Sixth IAEA Technical Meeting on Fusion Data Processing, Validation and Analysis

9 Sept 2025, 08:00 → 12 Sept 2025, 18:20 Europe/Vienna

Auditorium Hall HGX 102 (Guanghua Twin Tower) (Fudan University, Shanghai, China)

KEY DEADLINES

30 May 2025 Deadline for submission of abstracts through IAEA-INDICO for regular contributions. 

15 July 2025 Deadline for submission of Participation Form (Form A), and Grant Application Form (Form C) (if applicable) through the official channels. Notification of acceptance of abstracts.

---

The validation and analysis of experimental data obtained from diagnostics used to characterize fusion plasmas are crucial for a knowledge-based understanding of the physical processes governing the dynamics of these plasmas. This meeting aims at fostering, in particular, discussions about research and development results that set out or underline trends observed in the current major fusion confinement devices. Accurate data processing leads to a better understanding of the physics related to fusion research. It is essential for the careful estimate of the error bars of the raw measurements and processed data.

Objectives

The objective of the meeting is to provide a platform for discussing key topics related to fusion data processing, validation, and analysis, with a particular focus on addressing the extrapolation needs for next-step fusion devices such as ITER.

Target Audience

This event is suitable for experienced scientists and young scientists working in the domain of plasma diagnostics and synthetic diagnostics data analysis.

Discover the Meeting in Shanghai.pdf

EVT2405091 Form A Participation Form.rtf

EVT2405091 Form C Grant Application Form.rtf

EVT2405091 Information Sheet.docx

Tuesday 9 September

Registration

Opening

Conveners: Geert Verdoolaege (UGent), Mr Matteo Barbarino (International Atomic Energy Agency)

Physics-Based Machine Learning

Convener: Geert Verdoolaege (Ghent University)

1 Advancing Transparent Deep Learning for Modeling Turbulence in Fusion Plasmas

In response to the increasing intricacy of plasma behaviors within magnetically confined fusion systems, we propose a comprehensive approach leveraging interpretable artificial intelligence (AI) to enhance modeling, system reduction, and diagnostic analysis. The presented framework integrates a range of machine learning methods designed to uphold both physical insight and computational reliability.
A central feature of this research is the introduction of Layered Polynomial Neural Networks (LPNNs), a symbolic regression method tailored to reconstruct governing dynamical equations from trajectory data. Unlike traditional black-box models, LPNNs retain a clear polynomial structure, enabling transparent interpretation of nonlinear relationships. Their architecture builds complexity hierarchically, allowing them to outperform conventional networks in capturing chaotic dynamics observed in benchmark systems such as Lorenz, Rössler, and Lotka–Volterra models, as well as in reduced models relevant to L–H mode transitions in fusion plasmas.
This interpretability extends to model reduction. Using LPNN-based reduced-order modeling (termed AiPoG), we address canonical problems like the one-dimensional Burgers’ equation, achieving more stable and physically meaningful solutions than those obtained through conventional Proper Orthogonal Decomposition (POD) combined with Galerkin projection. The results underscore superior fidelity in resolving both advective and dissipative dynamics.
Simultaneously, we develop a hybrid Reduced-Order Model based on Neural Ordinary Differential Equations (NODEs) to simulate edge turbulence governed by the Hasegawa–Wakatani equations. This NODE-ROM model is built atop Galerkin-reduced modes and captures key dynamical features such as Lyapunov exponents, ensuring long-term trajectory fidelity and preserving system invariants—critical for real-time forecasting and potential feedback control in plasma operations.
Recognizing the challenges posed by chaotic regimes, especially regarding generalization, we introduce a novel phase-space-aware training strategy for NODEs. This method enhances extrapolation capability and stabilizes predictions over long time horizons by incorporating varied initial conditions during training. The approach proves effective in maintaining the structure of sensitive attractors like those in the Lorenz system.
Altogether, this work demonstrates how physically grounded AI tools can bridge simulation and theory offering new capabilities in modeling and understanding fusion plasma systems. These advances hold particular promise for improved control, monitoring, and design of next-generation fusion experiments.

Speaker: Mr David Garrido-Gonzalez (Aix Marseille University)

2 The Potential of Physics-Informed Neural Networks to Analyse Tokamak Diagnostic Measurements

Extracting net energy from fusion reactors will require a profound understanding of the underlying physics and the development of efficient control strategies. Plasma diagnostics are essential to these efforts, but obtaining accurate information from their measurements often involves solving quite delicate data analysis problems. Regrettably, many of the current approaches rely on simplifying assumptions, sometimes inaccurate or not completely verified, with consequent suboptimal outcomes. In order to overcome these challenges, the present study explores the potential of Physics-Informed Neural Networks (PINNs) to tackle various issues posed by the interpretation of diagnostic measurements in tokamaks. PINNs are a new branch of artificial intelligence that allows integrating data-driven methodology and physics equations in a very efficient way. The physics equations can be incomplete, leaving to the experimental data the task of providing the missing information. Moreover, they do not need a mesh and, with the approach of domain decomposition, can be easily applied to complex tasks. All these features have motivated the use of PINNs to address many challenging problems in various fields, ranging from fluid dynamics and physics to engineering and medicine. The present contribution describes the first attempts of developing PINNs to perform data analysis tasks in tokamaks. Various examples are provided ranging from equilibrium reconstruction to profile identification and tomography. The potential of the technology to perform integrated data analysis will also be briefly discussed. Overall, the undertaken study confirms the great potential of PINNs for data analysis in magnetic confinement thermonuclear fusion and highlights the benefits of using advanced machine learning techniques for the interpretation of several plasma diagnostic measurements.

Speaker: Dr Riccardo Rossi (Department of Industrial Engineering, University of Rome Tor Vergata)

3 When Explainable AI is not enough: Informed Machine Learning to Combine Fidelity and Interpretability

Recently the huge amounts of data collected in modern large experiments have motivated the deployment of machine learning (ML) in physics. Unfortunately, ML models are typically black boxes almost impossible to interpret. Consequently, in the last years eXplainable Artificial Intelligence, whose objective consists of improving the transparency of ML tools, has received a lot of attention. However in science, explainability assumes a flavour different from that in commercial applications and cannot be reduced to pure user convenience. Indeed the priority is also fidelity, which means developing models that reflect the actual mechanisms at play in the investigated phenomena. Genetic Programming supported Symbolic Regression (GPSR) has some obvious competitive advantages in trying to find an optimal trade-off between interpretability and realism. Unfortunately, the search spaces are typically too large; therefore the algorithms have to be steered to converge on meaningful solutions. The present work describes techniques to constrain GPSR and to combine it with deep leaning tools, so that the final models are expressed in terms of interpretable mathematical equations but at the same time reflect the physics at play in the phenomena under study. The strategies to combine fidelity and explainability include dimensional analysis, integration of prior information about symmetries and conservation laws and improvements of the fitness function. Great attention has been paid to devising practical solutions and they cover all the essential aspects of the data analysis process, from the treatment of the uncertainties to the quantification of the model complexity. All the main applications of ML, from regression to density estimation and classification, benefit from the proposed improvements. Theoretical considerations, systematic numerical tests, simulations with multiphysics codes and the results of actual experiments prove the potential of the devised upgrades.

Speaker: Andrea Murari (Consorzio RFX)

Coffee Break

Signal Processing and Anomaly Detection

Convener: Andrea Murari (Consorzio RFX)

4 A Comprehensive Strategy of Disruption Prediction to Avoid the Collapse of the Configuration in the Next Generation of Tokamak Devices

Disruptions are catastrophic forms of collapse that have affected all tokamak devices and are therefore one of the main potential showstoppers on the route to a commercial reactor. A new approach to proximity detection has been developed. It allows determining both the probability of and the time interval remaining before an incoming disruption. The methodology has been implemented with adaptive, from scratch, real time compatible routines. Moreover particular attention has been paid to identify machine independent indicators, to improve the transfer of the predictors to different devices. The new analysis methods have been deployed on thousands of JET discharges, covering the isotopic compositions from hydrogen to full tritium and including the last major D-T campaign. The nature of the main types of disruptions has been investigated to corroborate the potential of the devised solutions. The results indicate that physics based prediction and control tools can be developed, to deploy strategies of disruption avoidance and prevention, that meet the requirements of the next generation of tokamaks.

Keywords: Tokamaks, Disruptions, Radiation limit, Tomography, Adaptive predictors, Proximity control, Transfer Learning.

Speaker: Michela Gelfusa (University of Rome Tor Vergata)

5 Time series methods for fusion plasma disruption prediction

The approaching initial operation of major new tokomaks is rendering more pressing the need for effective disruption prediction techniques. The required tools should be not only accurate but also capable of operating with a minimum number of signals, because in the first campaigns of new devices typically only a very limited number of diagnostics is available. In addition, a very limited number of examples and extremely unbalanced sets of cases will make the problem even more difficult. These are very unfavorable conditions for the training of traditional machine learning classifiers. Therefore, approaches based on the identification of dynamical changes in the time series corresponding to routinely available diagnostic signals (locked mode, plasma current) has been developed recently. These methods are based on the identification of chaos onset, detection of concept drifts, changes in the complexity of the time series represented by ordinal patterns. The methods are capable of detecting the plasma drifting towards dangerous regions of the operational space in real time with high accuracy.

Speaker: Teddy CRACIUNESCU (National Institute for Lasers, Plasma and Radiation Physics INFLPR, Magurele, Romania)

6 Predictive Maintenance in Fusion Devices: Estimating the Remaining Useful Life of Plasma-Facing Component Units Using a Similarity-Based Approach

With a view to reliability, availability, maintainability, and inspectability (RAMI) for DEMO, achieving an availability of 30-60\% while minimizing unscheduled shutdowns is a critical prerequisite [1]. Plasma-facing components (PFCs), such as the divertor and first wall, are exposed to extreme thermal loads, shocks, and particle bombardment, conditions that can lead to component failure. In future fusion reactors, such failures may severely compromise plant availability, triggering costly remote maintenance interventions that typically last several weeks, resulting in substantial operational disruption directly impacting the final price of electricity. Predictive maintenance addresses these challenges by enabling proactive scheduling based on estimates of the remaining useful life (RUL) of components, derived from historical sensor data. However, monitoring PFCs in real time is constrained by limited diagnostic access, and full-scale modelling of reactor conditions is computationally prohibitive. In this study, we propose a similarity-based method for RUL estimation of beryllium tiles under steady-state heat loading from an electron beam [2]. Texture features including gray level co-occurrence matrix, local binary pattern, and fast Fourier transform are extracted from infrared images to capture the evolution of local pixel intensity patterns in the hottest surface regions and are fused into a health indicator. The RUL is then inferred by comparing the current health trajectory to those of previously observed tiles and interpolating from the most similar cases. This data-driven approach does not rely on a physical degradation model, making it particularly suited for conditions where failure mechanisms are complex or poorly understood. Although demonstrated under controlled conditions, the method is potentially extensible to more complex damage in real machine operations.

[1] Maisonnier, D. (2018). RAMI: The Main Challenge of Fusion Nuclear Technologies. Fusion Engineering and Design, 136, 1202–1208
[2] Hirai, T., Ezato, K., & Majerus, P. (2005). ITER Relevant High Heat Flux Testing on Plasma Facing Surfaces. Materials Transactions, 46(3), 412–424

Speaker: Geert Verdoolaege (Ghent University)

7 Determine the mode number and wave-number of plasma instability using harmonics in the spectrum

Harmonic phenomenon occurs frequently in the plasma diagnostic signal spectrum and is often considered useless or even harmful. Because the harmonic of plasma modes may overlap with other modes making it difficult for experimenters to distinguish between these modes. But in some case harmonics of mode may carry important information that helps further study plasma instability modes.
Harmonic phenomenon may come from mode coupling with itself. By analyzing the different harmonic frequencies in the spectrum, the toroidal mode number of the mode be determined. For example, Harmonics are seen in the spectrum of laser interferometer. There are mode and its harmonics in the spectrum at around 345~350ms. According to the magnetic probes the toroidal mode number is n=3. After analyzing the harmonics in the spectrum, the toroidal mode number n=3 or 4 is obtained. This is generally consistent with the measurements obtained from the magnetic probes.
The Bragg condition shows that the most efficient reflection of microwave occurs at kf=2k0, where kf is the wavenumber of the perturbation spectrum, k0 is the wavenumber of incident wave. The wavenumber of harmonic increases with the order of harmonics. When the wavenumber of harmonics meets the Bragg condition, it will show a stronger reflect phenomena. This phenomenon has been observed on HL-3’s microwave reflectometer signal. And after determining the wave number of the incident wave, the wave number range of the mode can be roughly determined.

Speaker: Liwen Hu (核工业西南物理研究院)

Lunch

Signal Processing and Anomaly Detection

Convener: Andrea Murari (Consorzio RFX)

8 On Data-driven Approaches in Predicting Turbulence Characteristics of Tokamak Plasmas

Abstract
Confinement mode transitions, particularly between L-mode, H-mode, and I-mode, are among the most critical phenomena in tokamak plasmas. These transitions are intrinsically linked to the formation and evolution of transport barriers, including internal transport barriers (ITBs), edge
transport barriers (ETBs), and double transport barriers (DTBs). The formation of these barriers is governed by the interplay of turbulence, shear flows, and equilibrium profiles. Accurate prediction of instability characteristics during such transitions is therefore essential for optimizing the plasma performance.
In this work, we propose a data-driven framework to analyze and predict turbulence type associated with the L-H transitions in tokamak experiments. We are trying to use diagnostic data from magnetic probes, ECE, reflectometry, and FIDA, and extract key turbulence features such as
frequency evolution, spectral broadening, and mode amplitude across different confinement modes. Given the limited availability and complexity of experimental data, we complement these analyses with reduced transport modes, i. e. , the Dynamical Critical Gradient (DCG) model. By systematically scanning the heating power, density, and pressure gradient thresholds, our method can reproduce the turbulence behaviors and transport barrier formation, creating a database that
links experimental signatures. Leveraging this database, we are trying to construct transformer based models capable of efficiently recognizing and predicting the turbulence transport, as well as identifying barrier-related thresholds, directly related with experimental observations.
Looking forward, we aim to extend this data-driven approach to HL-3 and EAST experiments, with particular emphasis on predicting power threshold of L-H transitions. This methodology has the potential to guide real-time confinement optimization and contribute to the development of advanced control strategies for next-generation fusion devices.

Speaker: Wenyang Li (Nankai University)

On Data-driven Approaches in Predicting Turbulence Characteristics of Tokamak Plasmas; Wenyang Li 6th IAEA TM on FDPVA Synopses.pdf

9 Real-Time Identification of sawtooth on HL-3 using a deep learning framework

Sawtooth instability is one of the most violent magnetohydrodynamic (MHD) instabilities that must be actively controlled in a tokamak fusion reactor.In present tokamak experiments, various auxiliary heating systems, such as neutral beam injection and ion or electron cyclotron resonance heating, are used. Electron cyclotron (EC) beams are particularly well-suited for controlling plasma instabilities due to their flexible power deposition, allowing for rapid adjustments across the plasma cross-section using mirror actuators. However, the current EC beam control system lacks real-time capabilities, making it difficult for sawtooth recognition algorithms to meet the required real-time processing speeds.
To address this challenge, a hybrid deep learning model integrating a long short-term memory (LSTM) network with a convolutional neural network (CNN) was developed. LSTM, a type of recurrent neural network specifically designed for long time-series data, effectively captures temporal dependencies in sawtooth oscillations. Meanwhile, CNN extracts spatial features from diagnostic signals, enhancing the model’s ability to detect intricate waveform patterns. This combination leverages both sequential and spatial feature extraction, improving recognition accuracy and robustness.The model was trained using over 10,000 sawtooth cycles derived from soft X-ray and electron cyclotron emission diagnostic data, reliably capturing characteristic sawtooth patterns in fusion plasmas. During validation, the algorithm achieved an accuracy of 92.5% in real-time experiments and 95.3% on post-processed data, demonstrating strong performance in both speed and accuracy. Additionally, cross-verification confirmed that the model’s statistical predictions align with the physical properties of HL-3 plasmas, making it a reliable tool for real-time sawtooth period control. This work contributes to the advancement of plasma diagnostics and control systems in fusion research.

Speaker: Hongjia OuYang (Swip)

Inverse Problems and Image Processing

Conveners: Andrea Murari (Consorzio RFX), Didier Mazon (CEA Cadarache)

10 Physics-Informed Neural Networks for Multi-Diagnostic Equilibrium Reconstruction in Tokamaks

Equilibrium reconstruction is a fundamental task in tokamaks, as it provides the distribution of the fields and currents inside the plasma. In recent years, the magnetic configurations and plasma scenarios have become increasingly complex. Their accurate identification is therefore particulalrly important to achieve the required performances. Accurate knowledge of the magnetic fields is also often a prerequisite for the interpretation of various diagnostic measurements. Unfortunately the equilibrium reconstruction is a severely ill-posed problem. It is therefore essential to constrain the algorithms with multiple diagnostics to achieve accurate results. The subject of the present work is a Physics-Informed Neural Network (PINN) algorithm for reconstructing plasma equilibrium using a multi-diagnostic approach, which includes magnetics, kinetic pressure, and interferometer-polarimeter data. Among these, the interferometer-polarimetric measurements are among the most valuable, as they provide line-integrated information about the internal magnetic fields. However, the polarisation evolution of an electromagnetic wave traversing a magnetised plasma exhibits non-linear effects, which render quite challenging the integration of this information into the reconstruction process. In the past this difficulty was circumvented by making recourse to various approximations. Unfortunately, these linearisations and approximations significantly limit the accuracy of the reconstructions in many plasma scenarios, particularly at high fields and currents. The developed PINN algorithm implements a comprehensive model (defined as the hot plasma model) that accounts for these nonlinearities and also thermal effects, both relativistic and non-relativistic. Parametric analyses conducted on synthetic cases demonstrate that the hot plasma model consistently yields results much more accurate than those obtained with the cold-plasma approximation or linearization of the polarimetric measurements. The PINN model has been also extensively tested with JET data of various high current campaigns, confirming the quality of the obtained reconstructions in all the investigated experimental conditions.

Speaker: Novella Rutigliano (Department of Industrial Engineering, University of Rome "Tor Vergata")

11 Improving resolution of plasma velocity-distribution-function measurements using multi-resolution data integration

To achieve high-performance plasmas, it is necessary to measure and understand perturbation components of plasma velocity distribution functions and avoid disruptive instabilities or continuous losses of plasmas by turbulent transport. One of the most reliable and prevailing methods for measuring velocity distribution functions is the charge exchange recombination spectroscopy (CXS). CXS determines the velocity distribution function and, therefore, the ion density, temperature, and velocity depending on the emission intensity, the Doppler broadening, and the shift of an ion emission line 1. However, as in other high-dimensional data measurement techniques, CXS has a trade-off between the resolution in real space, velocity space, and time and the signal-to-noise ratio (SNR). Currently, attempts are being made to improve both the SNR and resolution through improvements in measurement instruments and experimental ingenuities, but such hardware improvements are limited due to development costs and physical constraints. Here we attempt to improve both the SNR and resolution by using a novel data analysis method.
In high-dimensional data analyses, binning axes in dimensions that are not of interest can reduce noise, the effects of missing data and outliers, and data volume, while simultaneously revealing the structure of the original data in the dimensions of interest. Through formulating such binning operations in a Bayesian manner, we developed a multi-resolution data integration method that determines the binning size for each axis that most preserves the structure of the original signal and integrates the optimally binned data for each axis to restore the original signal. We applied this method to CXS data and achieved improvements of the resolution in real space, velocity space, and time, as well as SNR.

This work was supported by JST (Moonshot R&D Program) Japan Grant Number JPMJMS24A3.

Reference
1 Kobayashi, T., M. Yoshinuma, and K. Ida. "Full-image operation of fast charge exchange recombination spectroscopy with high-spatial and high-wavelength resolutions in large helical device." Review of Scientific Instruments 96.3 (2025).

Speaker: Toru Aonishi (The University of Tokyo)

12 First wall Heat flux estimations with THEODOR – an update: 3D solver, material inhomogeneities, resolved surface layers

In plasma experiments heat loads to material surfaces are of interested, which involves solving an inverse problem given temperature information. Infrared cameras can provide surface temperature information at high spatial and temporal resolution. To determine the heat flux density to the material the heat diffusion equation in the solid needs to be solved, respecting relevant boundary conditions and non-linearities of the transport.
At ASDEX Upgrade (AUG) the idea for the THEODOR code (THermal Energy Onto DivertOR) was established 1995 [1] to estimate surface heat loads on the divertor for investigations of plasma physics [2] and used in a Bayesian Framework [3] for high fidelity reconstructions. In recent years improvements have been made, such as a transition from an explicit to an implicit solver, inhomogeneous and optimised depth discretisation, an extension to 3D, and mixed material properties leading to better treatment of surface layers.
These developments are presented using the example of the new upper divertor in AUG for Alternative Divertor Concept (ADC) scenarios. The result is a non-toroidally symmetric heat load pattern on tiles with a trapezoidal plasma-facing surface and voids in the material due to mounting points. The finite difference solver provides an efficient scheme with simple setup and application. While finite difference schemes generally restrict the tile geometry to orthogonal grids, modifications to the local coefficients can be used to introduce spatially variable material parameter, or features like voids.
The currently predominantly used version is a Python implementation with uniform interfaces for the 1D, 2D, and 3D solver, core components compiled with numba for comparable speed to the C++ version, and parallelisation providing about 100 Hz evaluation speed for 3D geometries (200x100x50 cells), making it compatible with online analysis. It is also easy to extract in-tile temperatures for comparisons to thermocouples, for which of course also the borehole can be included in the geometry. The default non-uniform depth resolution allows for high resolution – and good response – on the surface, while saving computational resource towards the rear of the tile.

References
[1] A Herrmann et al 1995 Plasma Phys. Control. Fusion 37 17
[2] B Sieglin et al 2013 Plasma Phys. Control. Fusion 55
[3] D Nille et al 2017 Springer Proceedings in Mathematics & Statistics, vol 239

Speaker: Dirk Stieglitz (Max Planck Institute for Plasma Physics)

Stieglitz_THEODOR_IAEA_6th_TechnicalMeeting.pdf

Coffee Break

Pattern Recognition

Convener: Geert Verdoolaege (Ghent University)

13 Estimation of Plasma Parameters Based on Discharge Settings on WEST

Chenguang Wan ^{1,3, *}, Feda Almuhisen ² , Philippe Moreau ² , Remy Nouailletas ² , Zhisong Qu ¹ , Youngwoo Cho ¹ , Robin Varennes ¹ , Kyungtak Lim ¹ , Kunpeng Li ¹ , Zhengping Luo ³ , Qiping Yuan ³ , Xavier Garbet ^1,2

¹ School of Physical and Mathematical Sciences, Nanyang Technological University, Singapore
² CEA, IRFM, Saint Paul-lez-Durance, France
³ Institute of Plasma Physics, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei, China

^* Who is the speaker

Email: chenguang.wan@ntu.edu.sg

---

Tokamak discharge settings play a crucial role in determining the execution of tokamak experiments. Estimating discharge results prior to execution is highly beneficial for validating diagnostic signal consistency, assisting in experimental data analysis, verifying theoretical models, and advancing control technology research and development.

Traditional physics-driven modeling tools rely on empirical models or first-principle derivations, commonly referred to as Integrated Modeling. This approach consists of a suite of modular codes that address various physical processes within the tokamak, including core transport, equilibrium, stability, boundary physics, heating, fueling, and current drive. Notable modeling workflows include ETS, PTRANSP, TSC, CRONOS, JINTRAC, METIS, ASTRA, and TOPICS. However, due to the complex nonlinear interactions between plasma and the tokamak system, directly estimating discharge results based on the discharge setup remains a significant scientific challenge.

To address this, researchers have increasingly turned to data-driven approaches, exploring applications such as disruption prediction, tokamak operation using reinforcement learning, plasma tomography, radiated power estimation, instability identification, neutral beam effect estimation, confinement regime classification, scaling law determination, filament detection on MAST-U, electron temperature profile estimation via SXR with Thomson scattering, coil current prediction in W7-X based on heat load patterns, equilibrium reconstruction, and equilibrium solvers.

Furthermore, machine learning methods have been explored for control purposes, particularly in accelerating computational processes. Notably, there are several works focused on directly controlling tokamak parameters [1, 2]. However, to the best of the authors' knowledge, few studies [3, 4] have directly estimated tokamak discharge results based on discharge setup parameters.

In this study, we train a neural network model using a temporal sequence of discharge settings and diagnostic signals from 550 WEST discharges. The model successfully reproduces six plasma parameters:

• Normalized beta ($\beta_n$)

• Toroidal beta ($\beta_t$)

• Poloidal beta ($\beta_p$)

• Plasma stored energy ($W_{mhd}$)

• Safety factor at magnetic axis ($q_0$)

• Safety factor at 95% flux surface ($q_{95}$)

---

Acknowledgements

The computational work for this article was partially performed using resources of the [National Supercomputing Centre, Singapore (https://www.nscc.sg).

This research is supported by the National Research Foundation, Singapore.

---

References

1. Degrave, J., Felici, F., Buchli, J. et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature, 602(7897), 414 (2022).
2. Seo, J., Lee, Y., Park, J. et al. Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature, 626(8000), 746 (2024).
3. Wan, C., Yu, Z., Wang, F., Liu, X., Li, J. Experiment data-driven modeling of tokamak discharge in EAST. Nuclear Fusion, 61(6), 066015 (2021).
4. Wan, C., Yu, Z., Pau, A., Liu, X., Li, J. EAST discharge prediction without integrating simulation results. Nuclear Fusion, 62(12), 126060 (2022).

Speaker: Dr Chenguang Wan (Nanyang Technological University)

14 A Bayesian approach for the selection of predictor variables in the L-H power threshold scaling in metal-wall machines

Understanding the scaling of the L-H transition power threshold $P_\mathrm{LH}$ is crucial for the operation of future tokamaks such as ITER and SPARC[1~3]. However, multicollinearity of the predictor variables in the scaling makes it difficult to disentangle the effect on the power threshold of the individual plasma variables. In this study, we analyze the scaling of $P_\mathrm{LH}$ in a database with experimental data from three metal-wall tokamaks: JET-ILW (W divertor, Be wall), ASDEX Upgrade (tungsten wall) and Alcator C-Mod (molybdenum wall)[4]. Power-law regression is carried out using a hierarchical Bayesian model with a Dirichlet process prior and a variable selection prior [5]. This method simultaneously allows identifying redundant predictor variables, as well as grouping highly correlated predictors. The analysis provides strong statistical evidence that, in the database studied, the poloidal magnetic field $B_\mathrm{p}$ is a better predictor of the threshold power than the toroidal field $B_\mathrm{t}$ or plasma current $I_\mathrm{p}$. Furthermore, the method systematically evaluates the impact of ten different divertor configurations for the three tokamaks. For ITER, the new model predicts that under the standard H-mode line-averaged density of $n_e = 1.0 \times 10^{20}~\mathrm{m}^{-3}$, the threshold power for deuterium plasmas can vary between 100 and 200 MW, with the large uncertainty mainly influenced by the divertor geometry. This work provides a systematic, statistically sound analysis of LH threshold data in a multi-machine database, potentially offering additional insight into the physics of the L-H transition.

Speaker: Peizheng Zhang (Ghent university)

15 Study of tungsten ions emissivity profiles in ITER plasmas using CR model

Intrinsic impurities in a magnetically confined fusion (MCF) plasma dilute the reaction fuel in the hot plasma core and, through their radiation, are detrimental to global energy confinement. Thus, the radiative patterns of impurity elements in a fusion plasma must be understood across the full range of temperatures achieved in present day MCF experiments.

For this purpose, we present a comprehensive atomic processes investigation of tungsten ions (W$^{44+}$-W$^{47+}$,W$^{62+}$-W$^{65+}$) in ITER plasmas. Using the FAC code, we calculated excitation rate coefficients and radiative transition rates, then solved the collisional-radiative model (CR) to derive emission spectra across 100 eV–100 keV. Prominent spectral lines were identified, and their photon emissivity coefficients (PEC) were evaluated. Then by employing the Integrated Modelling & Analysis Suite (IMAS) database for the given simulation "shot", 7.5 MA scenario with impurity concentrations ( W: 10−4) and Te and ne profiles, we computed emissivity profiles for key spectral lines. This work provides critical atomic data for diagnosing impurity transport and radiation losses in ITER, benchmarking CR model and FAC calculation with ADAS database. Finally, by integrating the radial profile and the line of sight of the X-ray spectrometer, line integrated spectra with high spatial resolution were obtained.

The results highlight the role of charge-state distribution and line emission in fusion-grade plasmas, supporting spectroscopic diagnostics for ITER.

Speaker: Runjia Bao (ASIPP)

Inverse Problems and Image Processing

Conveners: Andrea Murari (Consorzio RFX), Didier Mazon (CEA Cadarache)

16 Latest Developments of the Maximum Likelihood Approach to Tomography for both Offline and Real Time Investigation of the Total Emission of Radiation

Quantification of the total emitted radiation is essential for the understanding and control of magnetic confinement plasmas. Its relevance is going to increase in the next generation of metallic devices that will have to operate at very high radiated fractions. The local emission from the bolometric integrated measurements is obtained with sophisticated tomographic algorithms. The layout of the diagnostics and the radiation patterns encountered in practice typically require solving very ill-posed inversion problems. The maximum likelihood tomography is one of the most advanced inversion methods and in this contribution the latest developments of the technique are presented. Firstly, the computational times are reduced of orders of magnitude by a matrix formulation of the problem, rendering the approach suitable for real-time feedback control. Secondly an adaptive procedure autonomously adjust the filtering to the radiation patterns, eliminating the need for human tuning of the hyperparameters and improving the capability of the technique to discover unexpected radiation patterns. Finally, the error estimation, a specific competitive advantage of the technique, is improved and validated with systematic Monte Carlo simulations. The performances of the new versions of the algorithms are compared with those of other methods reported in the literature with both synthetic and experimental data. The potential of the new improvements is substantiated by the analysis of the emitted radiation in phenomena such as MARFE, temperature hollowness and detachment in JET with a metallic wall.

Speaker: Ivan Wyss (1) University of Rome “Tor Vergata”,Department of Industrial Engineering, Via del Politecnico 1, 00133 Rome, Italy)

17 Gaussian process tomography for soft x-ray array and MHD mode analysis on HL-3 tokamak

A new soft X-ray (SXR) array system based on ceramic circuit boards has been installed on HL-3 tokamak to provide line-integrated measurements of the plasma emissivity in the energy range of 0.1–20 keV, which contains important information for the study of magnetohydrodynamic (MHD) activity. The lines-of-sight of this diagnostic cover a substantial portion of the poloidal cross-section, creating favorable conditions for two-dimensional tomographic reconstruction of the SXR emissivity. To investigate MHD behavior in HL-3, a Bayesian-based Gaussian Process Tomography (GPT) method has been developed to reconstruct local emissivity profiles from a limited number of noisy line-integrated measurements. To enhance the accuracy and reliability of the tomographic reconstruction, the non-negativity of emissivity is enforced in GPT using truncated Gaussian by Gibbs sampler, and knowledge about equilibrium magnetic geometry is incorporated into the construction through an anisotropic covariance matrix. The singular value decomposition (SVD) technique is employed to analyze the SXR emissivity and to extract the spatial and temporal features of MHD activity. The GPT method has been validated using both synthetic and experimental SXR data from HL-3 plasmas. For example, a growing m=1 mode localizing near the q=1 surface was observed as a precursor to a sawtooth oscillation, appearing before the crash and rapidly vanishing afterward. MHD activity under high-$\beta_N$ conditions has also been investigated. During a discharge with $\beta_N\sim$ 2.1, both 2/1 and 1/1 modes were simultaneously observed, which exhibited the same frequency but were located at different radial positions.

Speaker: Sen Xu (Southwestern Institute of Physics)

Conference dinner

Wednesday 10 September

Registration

Sensor Fusion and Integrated Data Analysis

Conveners: Rainer Fischer, Sehyun Kwak (Max-Planck-Institute for Plasma Physics)

18 Applications of Bayesian Data Analysis in KSTAR

Fusion devices, such as the Korea Superconducting Tokamak Advanced Research (KSTAR) facility, operate under conditions characterized by substantial noise and complexity, necessitating advanced non-invasive diagnostic methods. In recent studies at KSTAR, Bayesian inference techniques are applied to significantly improve the accuracy and reliability of plasma diagnostics. Specifically, Bayesian methods refine plasma edge density profiles by integrating hydrogen beam emission spectroscopy (BES) and two-color interferometry (TCI), resulting in improved resolution and enhanced understanding of edge plasma behavior.
Moreover, a nonnegative Gaussian process tomography framework based on Bayesian inference addresses radiative power losses during plasma disruptions. This approach successfully overcomes the challenges of reconstructing spatial profiles from limited diagnostic data, enabling precise characterization of disruption events and enhancing predictive capabilities.
Additionally, Kalman filter-based data fusion methods integrate magnetic coil and Hall sensor measurements, effectively leveraging the complementary strengths of each sensor type. Bayesian statistics systematically optimize Kalman filter hyperparameters. This combined approach significantly reduces measurement uncertainties and is anticipated to enhance the accuracy of magnetic equilibrium reconstructions, which is crucial for precise plasma control and stable tokamak operation.
Through these applications, Bayesian techniques provide robust solutions to challenges posed by high-dimensional, noisy, and incomplete experimental data, facilitating more accurate plasma characterization and advancing magnetic fusion research.

Speaker: Jaewook Kim (Korea Institute of Fusion Energy (KFE))

19 Leveraging neural networks for real-time plasma density inference using beam emission spectroscopy

Beam emission spectroscopy [1] (BES) is an active plasma diagnostic employed for plasma density measurements. In multiple BES applications such as synthetic diagnostics, density inference models, and plasma control frameworks computationally expensive emission inference calculations are utilised to determine the expected emission for a given density profile. The resource intensiveness of such calculations limits the applicability of BES for real-time measurements and control, while also restricting the speed of synthetic diagnostics.
In this work, we present two possible solutions to this problem in the form of a neural networks that can predict beam emission relevant profiles on a sub-millisecond timescale, both of which explore the use of extreme learning machine, multi-layer perceptron and convolution neural networks with scalable physics informed loss functions [2].
A forward inference model was developed focusing on emission inference along the beam based on assumed along the beam density profiles in an effort to be integrated and speed-up forward modelling based density inference frameworks, such as IDA [3]. The framework aims to enable real-time density measurement features, developed for the ASDEX-Upgrade Li-BES system [4].
A reverse inference model was developed focusing on plasma density inference along the beam based on beam emission measurements in order to enhance existing plasma density reconstruction methods used on the W7X stellerators Alkali-BES system [5].
Following preliminary results, inference uncertainty was assessed by use of multiple networks and found to be within acceptable margins, inference smoothness was deemed comparable to that of numerical methods and network performance was not affected by BES specific spatial resolutions.

References:
[1] D.M. Thomas et al. Fusion Sci. Technol., 53 487-527 (2008)
[2] M. Karacsonyi et al. 49th EPS, P1.029 (2023)
[3] R. Fischer et al. arXiv: 2411.09270 (2024)
[4] M. Willensdorfer et al. Plasma Phys. Control. Fusion, 56 025008 (2014)
[5] M. Vecsei et al. Rev. Sci. Instrum., 92 113501 (2021)

Speaker: Dr Örs Asztalos (HUN REN Centre for Energy Research)

20 Data fusion and uncertainty quantification of density data on EAST tokamak

In Tokamak, plasma density is a key parameter influencing confinement and transport. The rapid acquisition of reliable plasma density profiles is crucial for the real-time feedback control of density. However, measurement uncertainty leads to inconsistence between different density diagnostic systems, making it difficult to use and understand data. This report introduces data fusion and uncertainty quantification algorithms in machine learning to integrate experiment data from multiple density diagnostic systems in EAST Tokamak. The forward diagnostic models of the Polarimeter-Interferometer (POINT), the interferometer (HCN), and the microwave reflector (DPR) are constructed according to the diagnostic principles and layouts of corresponding density diagnostic systems. Bayesian inference is used to fuse diagnostic data from POINT, HCN, and DPR systems, providing uncertainty-quantified density distributions with higher confidence than individual diagnostic system. Convolutional Neural Network (CNN) is employed to integrate diagnostic data from POINT and HCN systems, achieving plasma density distributions and their uncertainty in milliseconds. The advantage of Bayesian-based method is its probabilistic characterization of physical quantities and measurement uncertainties, while providing an intuitive and flexible inference framework that enables data fusion across heterogeneous diagnostic systems through the product of likelihood functions. In contrast, deep learning-based method has the advantage of rapidly generating determination results, making it suitable for real-time data processing. This work is beneficial for providing a self-consistent and reliable density distribution for the research of confinement and transport and for the real-time feedback control of density. The algorithms and models established by this work, would provide reference for data fusion of multiple diagnostic systems and uncertainty quantification of key physical quantities in future fusion reactors.
References
[1] Fischer R, Fuchs C J, Kurzan B, et al. Integrated data analysis of profile diagnostics at ASDEX Upgrade[J]. Fusion science and technology, 2010, 58(2): 675-684.
[2] Lan T, Li G S, Liu H Q, Wang S X, Zhu X. Full poloidal section inversion using data of interferometer and reflectometer based on deep learning techniques. 2025, on submitting.
[3] Xie XP, Lan T, Liu HQ, Zhu X, Mao WZ, Lan T, Ding WX. Neural-network based electron density profile inversion for interferometer on EAST tokamak[J]. Plasma Physics and Controlled Fusion, 2025, 67(4): 045001.
[4] Lan T, Liu HQ, Ren QL, Zhu X, Mao WZ, Yuan Y, Wang YF. Electron density profile reconstruction with convolutional neural networks[J]. Plasma Physics and Controlled Fusion, 2022, 64(12): 124003.
[5] Lan T, Liu J*, Qin H, Xu L L. Time-domain global similarity method for automatic data cleaning for multi-channel measurement systems in magnetic confinement fusion devices[J]. Computer Physics Communications, 2019, 234:159-166.

Speaker: Ting Lan (Institute of Plasma Physics, Chinese Academy of Sciences Hefei)

10:25 Coffee break

Data Analysis for Feedback Control

Convener: Geert Verdoolaege (Ghent University)

21 Integrated Intelligent Control Framework for Plasma in HL-3 Tokamak

In magnetic confinement fusion, precise control of plasma dynamics and shape is essential for stable operation. We present two complementary developments toward real‑time, intelligent control on the HL‑3 tokamak. First, we build a high‑fidelity, fully data‑driven dynamics model to accelerate reinforcement learning (RL)–based trajectory control. By addressing compounding errors inherent to autoregressive simulation, our model achieves accurate long‑term predictions of plasma current and last closed flux surface. Coupled with the EFITNN surrogate for magnetic equilibrium reconstruction, the RL agent learns within minutes to issue magnetic coil commands at $1$ kHz, sustaining a $400$ ms control horizon with engineering‑level waveform tracking. The agent also demonstrates zero‑shot adaptation to new triangularity targets, confirming the robustness of the learned dynamics.

Second, we develop a non‑magnetic, vision based method for real‑time plasma shape detection. We adapt the Swin Transformer into a Poolformer Swin Transformer (PST) that interprets CCD camera images to infer six shape parameters under visual interference, without manual labeling. Through multi‑task learning and knowledge distillation, PST estimates the radial and vertical positions ($R$ and $Z$) with mean average errors below $1.1$ cm and $1.8$ cm, respectively, in under $2$  ms per frame—an $80$ percent speed gain over the smallest standard Swin model. Deployed via TensorRT, PST enables a $500 $ms stable PID feedback loop based on image‑computed horizontal displacement.

Together, these two streams lay the groundwork for a fully closed‑loop, vision‑informed RL control system. Although each module has been tested on its own, the next step is to link real‑time shape feedback with the RL‑trained coil actuator policy to enable continuous, model‑based control with minimal reliance on magnetic probes.

Speaker: Rongpeng Li (Zhejiang university)

22 Real-Time Detachment Forecaster: Decoding X-Point Radiation Oscillations in Impurity-Seeded Plasmas

Impurity seeding plays a pivotal role in achieving plasma detachment by reducing heat and particle fluxes to divertor targets, yet requires precise real-time control of seeding rates. Current diagnostic limitations and manual adjustments impede this process. For instance, the credibility of Langmuir probes becomes suspect under the heating of reactor level [#1]. Additionally, line-integrated measurements of the radiation spectrum can only yield rough and time-lagged two-dimensional radiation distributions [#2].

Consequently, a deep learning model has been developed for monitoring detachment in EAST, enabling instantaneous prediction of the electron temperature near divertor strike points. The model avoids reliance on Langmuir probes by utilizing photodiode radiation data and accommodating diverse operational conditions [#3].

Rigorous analysis has confirmed that the detachment state is primarily determined by the neutral beam injection (NBI) power, plasma current, line-averaged density, and impurity seeding rate [#4] (see Figure 1). Notably, it turns out that NBI synergizes with radio-frequency heating, broadening heat flux profiles and thereby facilitating plasma detachment. Moreover, the effect of impurity seeding remains consistent across different toroidal seeding locations [#3].

Crucially, the model demonstrates self-consistent predictions across nitrogen, neon, and argon seeding scenarios, despite being trained solely on nitrogen data. (see Figure 2). This consistency further validates the model’s applicability across diverse impurity seeding scenarios. The relative efficiencies among different impurity species are compared with the theoretical values in reference [#5], which rectifies the flaws of 1D models. This fresh perspective will advance the understanding of detachment control.

Figure 1. Cross-Correlations Between Input Variables and the Output

Figure 2. Validation of the Underlying Physics of the Model with Neon Seeding Scenarios

References
[#1] P.C.Stangeby, Plasma Phys. Control. Fusion 37, 1031 (1995).
[#2] W.Wen et al., Plasma Sci. Technol. 26, 095102 (2024).
[#3] Y.Yu et al., Plasma Phys. Control. Fusion 67, 025026 (2025).
[#4] S.S.Henderson et al., Nucl. Fusion 64, 066006 (2024).
[#5] A.Kallenbach et al., Plasma Phys. Control. Fusion 58, 045013 (2016).

Speaker: Yue Yu

IAEA-2025_YY.pptx

23 FusionMAE: large-scale pretrained model to optimize and simplify diagnostic and control of fusion plasma

The complex multiscale nonlinear dynamics of magnetically confined plasmas necessitate integrating massive diagnostic systems with control actuators in tokamak reactors. The complexity brought by such massive systems and their tangled interrelations has been a main obstacle in the way of fusion power plant. In this work, a large-scale model, fusion masked auto-encoder (FusionMAE) is pretrained to compresses information from 88 diagnostic signals into a concrete embedding, trying to provide a unified interface between diagnostic systems and control actuators. Two mechanisms are introduced to ensure a meaningful embedding. Firstly, the dimensionality of input array must be reduced to generate a smaller vector, while still allowing a decoder to restore all the raw data with minimal loss. Therefore, FusionMAE must capture the intrinsic low-dimensional manifold of plasma dynamics. Secondly, FusionMAE is trained reconstruct randomly masked input signals based on the intercorrelation across different channels. Surprisingly, FusionMAE accomplished these pretraining tasks and emerge additional intelligent functionalities. It emulates conventional data analysis pipelines, generates an all-purpose vector that enhance downstream control/simulation tasks, and enables diagnostic channel reduction while improving operational performance. This work pioneers large-scale AI model integration in fusion energy, demonstrating how pre-trained embeddings can simplify the system interface, reducing necessary diagnostic systems and optimize operation performance for future fusion reactors.

Speaker: Zongyu Yang

24 Validation and comparison of FPGA-based real-time Thomson Scattering data processing against offline analysis methods

A real-time signal processing unit based on FPGA has been developed for electron temperature (Te) evaluation in a Thomson scattering (TS) diagnostic system. The FPGA module is independent of the polychromator and is integrated with ADC and DAC components as a compact real-time processing unit. It receives five analog input channels directly from the spectral outputs of the polychromator and digitizes scattered signal pulses (~40 ns width) using a 16-bit, 1 GS/s ADC.
After background subtraction, each channel’s waveform is integrated, consuming 1.01 μs in simulation. The resulting integrals are then standardized across the five channels, a process taking 4.26 μs. These standardized values are matched against a precomputed 5001×5 spectral response database using a minimum-variance search algorithm. The database lookup takes 12.4605 μs in simulation, yielding the Te estimate. A 16-bit DAC converts the result to a 0–10 V analog voltage, linearly corresponding to 1–5000 eV. Preliminary validation has been performed for Te; the electron density (ne) algorithm is under development. The system demonstrates sub-15 μs processing latency and provides a viable foundation for future real-time plasma diagnostics and control applications.

Speaker: Wenyan Zhai

12:30 Lunch

Physics-Based Machine Learning

Convener: Andrea Murari (Consorzio RFX)

25 Prediction of NTM seed magnetic island trigger threshold in EAST based on supervised learning

The stability control of neoclassical tearing modes (NTMs) is critical for achieving high-performance steady-state operation in future magnetic confinement fusion devices. Active suppression of seed magnetic island formation represents a key early intervention strategy to minimize the cost of NTM control. This study addresses the critical threshold problem of NTM seed magnetic island triggering in the EAST tokamak, proposing a supervised learning-based temporal prediction framework to identify key
triggering parameters and quantify their abrupt transition characteristics. By integrating diagnostic signals (e.g., Mirnov probes, soft X-rays, electron cyclotron emission ECE) and inversion parameters (βp, q profile), a multimodal temporal database (time resolution ≤1 ms) containing magnetic island width evolution is constructed, focusing on capturing trigger event labels where the magnetic island width exceeds 2
cm. Using a hybrid deep network (HDL) and LightGBM algorithm with physics-informed feature engineering, the following objectives are achieved: 1) Establishing a correlation model between magnetic
island trigger thresholds and βp/ne, validating experimentally observed critical conditions; 2) Revealing the dominant roles of 1/1 internal kink mode coupling strength and error field harmonic components through SHAP value analysis and feature importance ranking for 2/1 NTMs; 3) Developing cross-device adaptation strategies to generalize the model to other tokamak data, verifying universal threshold patterns of normalized parameters (e.g., βN/q95). Experimental validation demonstrates high-precision prediction (AUC >0.91 with ≥20 ms warning window) on EAST historical data, showing consistency between key parameters (magnetic island growth rate, soft X-ray fluctuation amplitude) and theoretical/simulation results. This research provides a data-driven theoretical tool for analyzing NTM triggering mechanisms and active avoidance strategies in ITER and future fusion reactors.

Speaker: Feifei Long (University of Science and Technology of China)

26 Integrated modeling and experimental validation of H-mode divertor detachment and core confinement compatibility on HL-2A tokamak

Abstract: The divertor detachment and heat flux control under high-confinement H-mode conditions in tokamaks represent critical physical challenges in current magnetic confinement fusion research. Understanding the impact of detachment on H-mode boundary transport physics, particularly its compatibility with core confinement, is central to resolving divertor detachment physics. In this study, experimental results on divertor detachment and core confinement compatibility in H-mode plasmas from the HL-2A tokamak are presented. On the OMFIT (Objective MHD Framework for Integrated Tasks) integrated modeling platform, a novel neural network-based fast integrated modeling method for the divertor target region has been developed, by integrating a new edge neural network module (Kun-Lun Neural Networks, KLNN) to enhance divertor, scrape-off-layer and edge pedestal fast prediction capability. For the first time, this method is applied to conduct integrated simulations of divertor detachment and core confinement compatibility in HL-2A discharge #39007 in high-confinement mode. The simulation results are validated against experimental measurements, which are consistent well with each other. Further analysis reveals that in HL-2A H-mode detachment scenarios: turbulent transport in the core region () with high poloidal wave numbers () is dominated by ion temperature gradient (ITG) modes, while electron-driven turbulence prevails in the region . In the boundary region, electron turbulence dominates at low normalized poloidal wave numbers (), whereas ITG modes become predominant at higher wave numbers (), accompanied by minor electron turbulence contributions. The research results of this paper provide a certain foundation for integrated simulation and experimental verification in the study of core-edge coupling physics in tokamak devices and some insights for understanding of detachment-compatible H-mode scenarios in next-step fusion devices.

Keywords: tokamak, detachment, H mode, integrated simulation

Speaker: Mr Yukun Shu (Southwest Institute of Physics)

Information Retrieval and Visualisation

Convener: Geert Verdoolaege (UGent)

27 Innovative Applications of Visualization Technologies for Scientists and the Public in Magnetic Confinement Fusion

This report focuses on technological innovations in visualization within magnetic confinement fusion research, systematically elaborating on the groundbreaking applications of cinematic dynamic simulation, extended reality (XR) interaction, and intelligent 3D reconstruction in device modeling, theoretical demonstration, and scientific communication.

Based on high spatiotemporal resolution physical simulation data from the EAST tokamak, the research team constructed a movie-quality visualization of the formation process of boundary instabilities. This achievement (Symplectic Structure-Preserving Particle-in-Cell Whole-Volume Simulation of Tokamak Plasmas to 111.3 Trillion Particles and 25.7 Billion Grids) was successfully shortlisted as a finalist for the 2021 Gordon Bell Prize (only six global entries were selected) and presented at the virtual Supercomputing Conference (SC21) in St. Louis, Missouri, USA.

In terms of visualization-assisted fusion research, the team developed a 3D reconstruction algorithm based on parallel computing results, extracting the structure of weakly coherent modes (WCM) in I-mode and the filamentary structure of edge-localized modes (ELM) in H-mode. Through a series of mappings and dimensionality reduction techniques that preserved key information, the team successfully rendered these models in XR headsets, enabling scientists to intuitively compare the differences between the two instability structures and closely observe the simulation results.

In the field of engineering visualization, the team focused on the "Keda Torus Experiment" (KTX) device, establishing an intelligent conversion system from CAD drawings to lightweight 3D models. By employing an adaptive mesh simplification algorithm based on geometric feature recognition, the model's face count was reduced to 1% of the original data while maintaining solenoid winding topological accuracy (error < 0.05%). This optimization reduced single-frame rendering loads to within the acceptable latency threshold for Quest 3 headsets. Additionally, the team innovatively developed a cross-platform, multimodal interaction system supporting collaborative work between Vision Pro and Quest 3 devices, with multi-user synchronization latency controlled below 50ms. Using 3D Gaussian Splatting technology, a laboratory-grade digital twin system was constructed, accurately replicating the KTX laboratory at the University of Science and Technology of China (USTC) in virtual space. This provides a high-fidelity platform for future remote experimental rehearsals.

Speaker: Zixi Liu (University of Science and Technology of China)

28 In-pulse data analysis at ITER: evaluating workflow managers and 3D visualization tools

Understanding plasma behavior throughout the entire duration of a pulse is critical for achieving the objectives of ITER and future fusion devices. Plasma parameters such as temperature, density, current profiles, and impurity content evolve dynamically during a discharge, influencing key aspects of performance including energy confinement, stability, and fusion power production. Analyzing plasma performance during the pulse—rather than relying solely on post-pulse data—provides essential insights into transient phenomena, transport processes, and the interplay of heating and fueling. This in-pulse analysis is vital for guiding control strategies for sustained, high-performance fusion plasmas during an experimental session.

In this work, we present the performance evaluation of Apache Airflow as a workflow management tool for in-pulse analysis of plasma performance at ITER. Given the complexity and interdependence of data analysis tasks required to extract meaningful insights from plasma discharges—such as data extraction, pre-processing, profile reconstruction, and physics-based computations—there is a clear need for a robust and scalable orchestration framework. To this end, we have developed a mock workflow composed of artificial, interdependent actors that emulate a realistic plasma analysis pipeline. This synthetic workflow mimics key features of an ITER plasma performance analysis chain, including task dependencies, data flow, and execution timing. By executing and monitoring this mock workflow using Apache Airflow, we assess key performance metrics such as task scheduling efficiency, failure handling, scalability, and resource management. Our results provide insights into the potential of workflow orchestration tools like Airflow to support the demanding in-pulse analysis requirements of ITER, offering a pathway toward more automated, reproducible, and robust analysis pipelines in future fusion experiments.

We also demonstrate the usage of ParaView as a key visualization and data exploration tool, to support both static and interactive visualization of plasma parameters. By embedding ParaView scripts and tasks directly into the workflow, we can enable the generation of detailed visual outputs—such as 3D renderings of magnetic flux surfaces, current and temperature profiles, and diagnostic signal mappings—at various stages of the analysis. In addition, we can also leverage ParaView to ensure that complex, high-dimensional plasma data can be explored dynamically rather than only through static plots, empowering researchers to uncover subtle features such as asymmetries, localized instabilities, and temporal evolutions that might otherwise be overlooked. Users can manipulate views, apply filters, and adjust color maps interactively to investigate the spatial and temporal structure of plasma events, providing timely feedback on the quality and relevance of the analysis results. In addition, ParaView's ability to handle large, multi-dimensional datasets is critical in the context of ITER, where vast amounts of data must be processed, visualized, and interpreted quickly to inform operational decisions and advance scientific understanding.

By integrating ParaView as a modular, scriptable actor, we ensure that visualization tasks can be executed in parallel with other analysis tasks without introducing bottlenecks. This integration facilitates a workflow where data exploration is not an afterthought but a core component of the analysis pipeline, supporting transparent, reproducible, and actionable plasma performance insights for ITER and future fusion devices.

Speaker: Paulo Abreu (ITER Organization)

JINTRAC divertor.mp4

PAbreu - In-pulse data analysis.pdf

PAbreu - In-pulse data analysis.pptx

29 JDDB: A Flexible and Extensible Data Processing Framework to Accelerate AI4Fusion Research

The increasing integration of artificial intelligence (AI) into fusion research demands scalable, standardized, and traceable data infrastructures. To address this need, we introduce JDDB (J-TEXT Disruption Database), a flexible and extensible most of all, a light weight data processing framework designed to streamline the management and transformation of fusion data across tokamak experiments.
As the name suggests, JDDB started as a database for data driven disruption prediction research. As it evolves, it now has the potential to play a foundational role in supporting AI4Fusion research by enabling robust data handling practices, reproducible workflows, and cross-device compatibility.
The design philosophy of JDDB is to be simple and lightweight. As long as the data and metadata is kept complete and organized, the rest is not core concern of JDDB. The key feature of JDDB is the data is split into two repositories. One named FileRepo is a bunch of HDF5 files, simple and straight forward for data processing jobs to handle. The other is the MetaDB, which is a MongoDB database contain all the metadata used for searching and tracing the data. For the FileRepo storage format, the ITER Integrated Modelling & Analysis Suite (IMAS) data dictionary is used as it’s the most accepted standard for fusion scientific data. For metadata, the MongoDB is used because if offers flexible scheme so different metadata for different signals can be stored without modify the database scheme. For data processing, JDDB provides a parallel, modular workflow engine that ensures efficiency and reproducibility. A core principle of JDDB is data immutability: raw data remain unchanged, while all processed data are versioned with full provenance tracking. This guarantees traceability and transparency in complex analysis pipelines, which is critical for verification and reproducibility, collaboration, and regulatory compliance in fusion research.
By unifying data structure, metadata, and workflow management into a coherent framework, JDDB significantly lowers the barrier for developing and deploying AI-driven models in fusion applications. It enables researchers to focus on physics-informed modeling and data analysis, while ensuring that their pipelines are robust, scalable, and transparent. JDDB is currently in active use within the J-TEXT tokamak and is being extended to support cross-device applications and collaborative development. Its architecture is well-suited for integration into future AI foundation models, intelligent simulations, and control strategy optimization for magnetic confinement fusion devices.

Speaker: Wei Zheng (International Joint Research Laboratory of Magnetic Confinement Fusion and Plasma Physics, Huazhong University of Science and Technology)

30 Inital development of web interface and user access control of IMAS databases

It is well known that ITER utilizes IMAS for data storage and exchanging between experiments and simulations. The software infrastructure of the synthetic diagnostic platform (SDP) on EAST is based on IMAS. To enable the visualization of the data on SDP through web,a RESTful API has been developed. Benefiting from the mature web technologies, it is also possible to easily integrate the user access control (UAC) feature into the web interface, which could protect the resources owned by the developers and boost the development of synthetic diagnostics on EAST. The Flask web framework has been chosen for the development. The static and responsive frontend has been developed for user registion, data application and data authorization. The backend couples IMAS and SQL databases for retrieving and controlling the data flowing to the web. The administrator can nominate the developers to manage the data application for certain IMAS resources to the IDS level. The developers can review the data application submmited by the users. Once authorized, the users have the access to the limited IMAS resouces. To simplify the access of the IMAS data through web, the client API has been developed for local data analysis by the users. The web data visualization feature will be developed in the future to further simplify the visualization of the SDP data anywhere at anytime.

Speaker: Xiang Liu (ASIPP)

Poster session with tea

31 Investigation on Fast ions Lossess Induced by NTM in HL-3 High-Beta Plasmas

Fast ions generated by fusion reactions or auxiliary heating system play a quite essential role in magnetic confinement fusion reactor. The experiments on ASDEX-U and DIII-D have shown that low-frequency MHD activities like TM or NTM will induce the fast ion losses[1][2].  Theoretical and simulation analyses indicate two distinct mechanisms for low-frequency MHD-induced fast-ion losses (FIL). For circulating particles, the mechanism involves the overlap of drift islands in phase space, while for trapped particles, it results from resonant interactions with the perturbation field when the ratio between the bounce frequency and the toroidal precession frequency matches the mode frequency[3][4].

Fast ion loss induced by NTM in HL-3 high-beta plasma have been investigated. The main diagnostics for experiment is FILD -- Fast ion Loss Detector. The spectrogram of FILD shows that FIL increase concomitantly with a reduction in the rotation frequency of NTM(Fig 1). Another shot in HL-3 the FIL will onset during the NTM mode locking process and unlocking process. This thesis analysis the quantity relation between FIL and NTM. Detailed results will be presented at this conference.

Reference:
[1]M. García-Muñoz et al 2007 Nucl. Fusion 47 L10.
[2]Phys. Plasmas 15, 032501 (2008).
[3]Phys. Plasmas 15, 056115 (2008).
[4]National Science Review, Volume 9, Issue 11, November 2022, nwac019.

Speaker: Zhiyi Yin (Southwestern Institute of Physics)

32 One-Dimensional Modelling and Experimental Validation of Divertor Detachment on HL-3 and HL-2A

A new self-consistent 1D scrape-off layer model is recently developed in BOUT++ framework, named SD1D, which includes transport equations of various particle species (e.g. main plasma, neutrals and impurities) and couples collisional and radiative reactions by using open databases like ADAS and AMJUEL. In this work, SD1D is used to model divertor detachment experiments on two tokamak devices, HL-2A and HL-3. It is found that the variation of target electron temperature and the target ion current in simulations are consistent with experimental results on HL-2A and HL-3. The variation of divertor Dα radiation intensity in the modelling is qualitatively similar to the measured Dα signal, which is crucial for understanding the effects of plasma-neutral interactions on divertor detachment. The validation demonstrates that SD1D is capable of rapidly and effectively simulating divertor detachment experiments across various machines.
This work identifies a potential strategy for divertor impurity control on HL-3: increasing the upstream density enhances parallel transport, driving impurities (both intrinsic and extrinsic) toward the target plate. It may be helpful for control of the neon radiation front during detachment (closer to the target).

Speaker: Yulin Zhou

33 Overcoming Plasma-Induced Noise: Statistical Optimization of α-Particle Detection in EXL-50U p-B Reactions

Accurate measurement of charged fusion products in magnetically confined plasmas faces significant challenges due to plasma radiation exposure, electromagnetic interference, thermal loads, and background bombardment by electrons/ions, which generate substantial noise while yielding sparse signals. To address these issues, advanced solutions are required beyond conventional detector design and calibration, including signal discrimination techniques, machine learning-based noise suppression for enhanced signal-to-noise ratios, and statistical analysis to improve signal significance. This study focuses on the detection simulation of α-particles generated by proton-boron (p-B) fusion reactions in the ENN-operated ST-type device EXL-50U. By employing Monte Carlo simulations, we systematically evaluate the expected α-particle signals, reducible backgrounds (e.g., electromagnetic interference, plasma fluctuations, and energetic protons/electrons/photons), and irreducible backgrounds (e.g., high-energy proton pileup events and fortuitously energized helium impurities). The signal significance of measurable fusion products is quantified using statistical metrics, providing critical insights for optimizing α-particle diagnostics in p-B fusion experiments under high-background conditions.

Speaker: zhi li (ENN Science and Technology Development Co., Ltd.)

Pattern Recognition

Convener: Geert Verdoolaege (UGent)

34 Variability and risk of edge-localized modes at JET using machine learning

Edge-localized modes (ELMs) are quasi-periodic edge instabilities in tokamaks, accompanied by expulsion of heat and particles from the plasma. Large ELMs can heighten the risk of damage to the plasma-facing components (PFCs). However, under certain conditions, ELMs can exhibit strongly stochastic behavior, showing a mix of relatively small and larger bursts. This complicates the predictability of the effects of ELMs on plasma operation and plasma-wall interaction. It is therefore of key interest to understand the conditions governing the degree of ELM stochasticity.

In this work, we investigate probability distributions of ELM timing and ELM size at JET, as well as the dependence of the distributions on some of the main operational characteristics. This allows us to exploit information about ELM variability under stationary plasma conditions. More than 1000 JET discharges were analyzed, dating back to 2012 after the installation of the ITER-like wall. Advanced machine learning techniques were developed in order to robustly detect the ELMs under a broad variety of plasma conditions, based on a training set of 16000 manually marked ELMs. The inter-ELM time was determined for all ELM events in the database, as well as an estimate of the drop in the plasma stored energy as a measure of ELM size. Focusing on time windows with stationary plasma conditions, the inter-ELM time can be modeled adequately by a Weibull distribution, while the log-normal distribution is an appropriate model for the ELM size. We use regression analysis to quantify the dependence on machine operational parameters of various characteristics of the distributions, going beyond the analysis of averaged ELM properties by studying the variance, percentiles and tail heaviness. In particular, the tail heaviness of the ELM size distribution is an indicator of unaccounted risk, as the released energy may exceed wall tolerances defined on the basis of average ELM behavior. We highlight the influence of plasma current, triangularity and gas fueling rate on distributional properties. A more qualitative view is offered by a set of fuzzy logic rules, indicating areas of the operational space with the highest risk of tail ELMs. Another indicator of risk is the energy released by several consecutive ELMs. We show that the cumulative energy loss due to multiple ELMs is largely affected by heating power and plasma current.

Overall, this study provides both qualitative and quantitative insight into the occurrence of rare, but potentially impactful ELMs. The code for ELM detection, which can be easily adapted to other types of events, will be published as open source.

Speaker: Jerome Alhage (Ghent University)

jalhage_IAEA-TM6_v5.pdf

35 Data-Driven Tokamak Plasma Magnetic Response Modelling

Traditional first-principle-based tokamak plasma magnetic response models, while pos-
sessing clear physical significance, often exhibit significant deviations when compared
with actual experimental data, limiting our deep understanding of plasma dynamics pro-
cesses. This study develops an autoregressive neural network model based on improved
WaveNet architecture that directly learns plasma behavior patterns from HL-3 device
experimental data, achieving high-precision reproduction of real plasma responses.
The model takes experimental measurements of poloidal field coil currents and cen-
tral solenoid coil currents as inputs, predicting key parameters such as plasma current,
position, minor radius, elongation, and triangularity reconstructed by EFITNN[1]. Ad-
dressing the specific requirements of experimental data validation, we designed an in-
novative dual loss function mechanism that combines teacher forcing and autoregressive
training approaches, effectively mitigating exposure bias issues and enabling the model to
better handle long-term sequence autoregressive predictions. The loss function employs
a weighted combination of MSE, MAE, and MTE, providing differentiated optimization
gradients for prediction errors of different magnitudes, significantly improving the model’s
fitting capability for both regular variations and sudden events in experimental data.
To simulate measurement noise and uncertainties in real experimental environments,
we introduced random walk noise layers and Gaussian noise layers during model training.
This design enables the neural network to adapt to various perturbations present in actual
experimental data and effectively suppress cumulative errors in long-period predictions.
Validation based on experimental data from HL-3 device shots #2000 #6698 demon-
strates that the model can reproduce single-step experimental measurement results with
99.5% accuracy, while maintaining 97.5% accuracy in autoregressive mode. More impor-
tantly, the high consistency between model predictions and real experimental data indi-
cates that the neural network successfully captures the intrinsic physical laws of plasma
magnetic response, providing a new perspective for understanding tokamak plasma dy-
namic behavior.
Through comparative analysis with traditional physics models, we found that data-
driven methods have significant advantages in reproducing experimental observations,
particularly excelling in handling nonlinear coupling effects and long-term evolution pro-
cesses. This modeling approach based on real experimental data validation not only
provides effective tools for current fusion device physics research but also establishes a
foundation for plasma behavior prediction and analysis in future large-scale devices such
as ITER.

Speaker: jiyuan Li (Southwestern Institute of Physics)

Data Analysis for Feedback Control

Convener: Geert Verdoolaege (Ghent University)

36 Rapid analysis model and extrapolation method of neural network in spectral diagnostic

A robust and interpretable neural network (NN)-based model has been developed for analyzing charge exchange spectra on the HL-2A tokamak, achieving high accuracy in ion temperature (T_i) and toroidal rotation velocity (v_t) estimation. Trained and tested on around 122 thousand spectra, the model achieves a coefficient of determination (𝑅²) of 0.948 for T_i and 0.973 for v_t, with an inference time of less than 1 ms per spectrum. Its robustness against novel data eliminates the need for human intervention, enabling real-time plasma feedback control. Interpretability analysis using Integrated Gradients (IG) reveals that the model estimates T_i and v_t by detecting Gaussian spectral line features, aligning with fundamental physical principles. To extend the model’s applicability, a method is proposed to generate synthetic high-temperature spectral data based on low-temperature experimental data. A new NN-based model is trained on data with T_i<2keV, which performs poorly when extrapolated to T_i>2 keV. By adding 5% synthetic high-T_i data, the model’s extrapolation capability is extended to T_i < 4 keV, reducing the mean relative error in the 3–4 keV range from 35% to below 15%.
This work bridges the gap between NN-based models and traditional methods, establishing a reliable alternative for spectral analysis in fusion research. The use of synthetic data to enhance AI algorithms demonstrates significant potential for real-time ion temperature measurement and feedback control in future high-parameter fusion devices.

Speaker: Wenjing Tian

37 Imitation Learning-Based Feedforward Current Optimization for Tokamak Plasma Control

Accurate real-time control of plasma equilibrium is critical for stable tokamak operation. This study proposes a novel imitation learning framework to predict optimal feedforward currents for poloidal field (PF) coils based on plasma state observations. The model ingests high-dimensional state vectors including plasma boundary coordinates, plasma current centroid positions, and total plasma current. Through DTW-based discharge selection and behavioral cloning of expert trajectories, the neural network directly maps states to 12-dimensional PF coil currents.

Experimental validation on EAST tokamak data is currently underway, with preliminary simulations indicating the potential to achieve >92% trajectory tracking accuracy in reconstructing expert PF current sequences, computational latency below 2 ms per inference, and enhanced robustness to plasma perturbations through DTW-curated training data. This work pioneers the integration of imitation learning into magnetic confinement fusion control systems, establishing a scalable framework for adaptive feedforward compensation that could significantly improve discharge stability in next-step devices such as ITER.

Keywords:
Imitation learning, Plasma control, Feedforward prediction, Dynamic Time Warping (DTW)

Speaker: Jingjing Lu

Thursday 11 September

Registration

Next Fusion Device Concepts: Data Challenges and Design Optimization

Convener: Didier Mazon (CEA Cadarache)

38 An optimization method for the ITER Radial X-ray Camera line-of-sight design using on Bayesian uncertainty analysis

This paper presents a novel uncertainty optimization algorithm for the design of line-of-sight (LOS) systems used in tomographic inversion. By extending Gaussian process tomography from discrete pixel space to continuous function space through Bayesian inference, we introduce an uncertainty function and analyse its typical distributions. We develop an algorithm to minimize the uncertainty, which is then applied to optimize the LOS configuration of the internal camera in the ITER project. Uncertainty analysis and phantom testing are conducted to validate the effectiveness of the proposed algorithm. The results demonstrate improved accuracy and stability in tomographic reconstructions. This study contributes to the advancement of LOS design for tomographic inversion, offering a practical solution for optimizing diagnostic systems in complex experimental settings.

Speaker: Tianbo WANG (ITER Organization)

39 Bayesian Techniques for Design Optimization of Magnetic Diagnostics and Validation on the WEST tokamak

Next-generation fusion devices such as DEMO present significant challenges in diagnostic system design due to spatial and cost constraints. Previous work has demonstrated the successful application of Bayesian experimental design to DEMO, optimizing both the placement and orientation of magnetic coils to reduce sensor quantity while maintaining diagnostic accuracy.
In this study, we extend the application of Bayesian methods to the magnetic diagnostic system of WEST, focusing on optimizing the quantity of pick-up coils using mutual information as the criterion for sensor selection. The approach systematically identifies sensor configurations that maximize information gain while minimizing measurement uncertainty in key plasma parameters, including plasma current centroid, total current, and X-point position.
Preliminary results indicate that a reduction in the number of pick-up coils is feasible without compromising diagnostic accuracy, underscoring the effectiveness of Bayesian design in guiding optimal sensor configurations. This work provides a rigorous framework for sensor optimization under engineering and economic constraints, offering insights for diagnostic design in future fusion devices.

Speaker: Yangyang Zhang

40 Role of Human Inputs in Cross-tokamak Disruption Prediction in Magnetic Confinement Fusion

Disruption is a catastrophic event in tokamaks and represents a major challenge for future commercial fusion reactors. Although data-driven disruption prediction models trained on a single tokamak have successfully triggered disruption mitigation, obtaining large disruptive datasets for every new device is economically and operationally impractical. Cross-tokamak disruption prediction is therefore an essential transfer learning problem within machine learning. Differences in tokamak designs and operational conditions significantly affect data distributions between disruptive and non-disruptive phases. Additionally, diagnostic system variations across tokamaks directly influence whether the measured signals represent the same or similar plasma parameters. These factors present substantial challenges for achieving cross-tokamak disruption prediction using machine learning alone. This work highlights the important role of human inputs—including feature engineering, labeling strategies, and estimating data distributions for future tokamaks—in cross-tokamak disruption prediction research. Such human inputs enable improved predictive performance with fewer data from future tokamaks, while also providing interpretability to enhance model trustworthiness.

Speaker: Chengshuo Shen (Huazhong University of Science and Technology)

Coffee Break

Uncertainty Propagation, Verification and Validation

Convener: Simon Pinches (ITER Organization)

41 What is Detachment? Data Analysis Techniques to Understand and Control Detachment in Tokamaks

A degree of detachment is critical for the sustainable operation of a fusion energy plant within stringent safety limits [1]. Consequently, it is essential to develop a thorough and practical understanding of divertor detachment and control to move the technology readiness level of the tokamak from development to deployment. This topical review will cover the data analysis techniques and numerical methods applied in the field, with guidance on assessing the appropriateness of a statistical approach in the field of detachment physics with a focus on the main diagnostics using this field: Langmuir probes, atomic and molecular spectroscopy, filtered cameras, and bolometry [2]. A comparative analysis will be made with analysis techniques used in the field of divertor control, and a critique of the applicability of the reviewed material to reactor-scale devices will be made.

There are specific complexities to diagnosing divertors compared to the core due to the open magnetic topology, low levels of symmetry, and large variation in geometric configuration and spatial location within the device. It follows that the symmetry-based parameter mapping and tomographic inversions that are common in core physics have reduced application in the study of detachment. Experimental investigations are further complicated by the need to quantify the atomic and molecular processes that elucidate the power, particle, and momentum balance in the divertor and hence determine the state of detachment [3]. As a result, a rich field of data analysis and uncertainty quantification has developed to characterise detachment in a range of machines using both Bayesian inference and inverse problem solving [4]. These problems are solved with a diverse range of numerical solutions from Least Squares Regression and Monte Carlo sampling to Broyden–Fletcher–Goldfarb–Shanno (BFGS) and gradient-based conditional sampling to genetic algorithms.

Sensor fusion (or integrated data analysis) has been a growing trend to optimise the increasing level of accessible computational resources and to utilise the growing set of diagnostics and coverage on current devices [5]. This is somewhat contrary to the needs for a reactor which, for engineering and safety reasons, will have the absolute minimum required set of diagnostics [6]. Additionally, the actuator control times will be significantly larger than for current day devices. In preparation for reactor-scale devices, analysis schemes will need to be developed that accommodate limited diagnostics and diagnostic coverage, and for detachment control to operate within the narrow tolerances of a reactor [7].

References

1. K. Krieger et al 2025 Nucl. Fusion 65 043001 (link)
2. J L Terry and M L Reinke 2017 Plasma Phys. Control. Fusion 59 044004 (link)
3. K. Verhaegh et al 2023 Nucl. Fusion 63 126023 (link)
4. A Pavone et al 2023 Plasma Phys. Control. Fusion 65 053001 (link)
5. C Bowman et al 2020 Plasma Phys. Control. Fusion 62 045014 (link)
6. F.P. Orsitto et al 2016 Nucl. Fusion 56 026009 (link)
7. M Yoshida et al 2022 Plasma Phys. Control. Fusion 64 054004 (link)

Speaker: Dr Daljeet Singh Gahle

42 Development and validation of synthetic diagnostics and inference models

Preparing to interpret the data arising from ITER plasma operation requires the development and validation of models for each of the diagnostic systems installed.

Prioritising the development of synthetic diagnostics for the Start of Research Operations (SRO) has focused attention initially on models for the interferometers (TIP and DIP), polarimeter (PoPola, including full Stokes vector calculation, Faraday and Cotton-Mouton effects), the full magnetic systems (magnetic probes, flux loops, saddle coils and Rogowskis including white or coloured (1/f) noise on observations), Thomson scattering (core and edge), soft x-rays, XRCS (core, edge, and survey), hard X-rays and ECE. These models have been developed with the Minerva modelling framework.

Since all these models have been developed following the IMAS paradigm, they are applicable to any device whose configuration (Machine Description data) and data can be mapped to follow the IMAS Data Dictionary. This allows these models to be validated by comparing their predictions with experimental measurements. In Minerva, these synthetic diagnostic models can also easily be combined to construct inference models to infer key physics parameters such as the electron density and temperature profiles

Speaker: Simon Pinches (ITER Organization)

IAEA TM on FDPVA - Development and validation of synthetic diagnostics and interference models.pdf

43 Benchmarking Fusion Plasma Diagnostics: Analysis of Experimental W and Mo X-ray Data from EBIT Resolves Plasma Speed Contradiction

Tungsten will be a strong candidate of the material for ITER divertor. All Spectroscopies related to tungsten were then become highlight fields. Meanwhile Molybdenum is an impurity that are not be neglected in many Tokamak plasma, and thus be used for diagnostic widely.
The x-ray transitions from W45+,46+ ions and Mo32+ in the 5.19–5.26Å wavelength range that are relevant as a high-temperature tokamak diagnostic, in particular for JET in the ITER-like wall configuration, have been studied in this work with an electron beam ion trap which produces and confines highly charged ions for disentangle studies of plasma atomic processes.
Tungsten spectra were measured at the upgraded Shanghai Electron Beam Ion Trap operated with electron-beam energies from 3.16 to 4.55 keV. High-resolution measurements were performed by means of a flat Si 111 crystal spectrometer[1] quipped by a CCD camera. The experimental wavelengths were determined with an accuracy of 0.3–0.4 mÅ[2]. All measured wavelengths were compared with those measured from JET ITER-like wall plasmas and with other experiments and various theoretical predictions including COWAN, RELAC, multi-configurational Dirac-Fock (MCDF), and FAC calculations. To obtain a higher accuracy from theoretical predictions, the MCDF calculations were extended by taking into account correlation effects (configuration-interaction approach). And Mo32+ spectra in the same region were studied after that. With these very accurate wavelength data, contradiction between theoretical result and measured values was resolved related to impurity transportation speed in the fusion plasma.

Fig. 1 A general scheme of the experimental setup at the upgraded Shanghai EBIT.

Fig 2. Tungsten (W45+ and W46+) and molybdenum (Mo32+) x-ray lines observed in the spectrum measured at JET (shot #85909) at 𝑇𝑒≈3.9 keV and 𝑛𝑒≈3.2×10^19 m−3.

References
[1] Y. Yang, J. Xiao, D. Lu, Y. Shen, K. Yao, C. Chen, R. Hutton, and Y. Zou, A high precision flat crystal spectrometer compatible for ultra-high vacuum light Source, REVIEW OF SCIENTIFIC INSTRUMENTS 88 113108 (2017)
[2] J.Rzadkiewicz, Y. Yang, K. Koziol, M. G. O’Mullane, A. Patel, J. Xiao, K. Yao, Y. Shen, D. Lu, R. Hutton, Y. Zou, and JET Contributors, High-resolution tungsten spectroscopy relevant to the diagnostic of high-temperature tokamak plasmas, PHYSICAL REVIEW A 97 052501 (2018)

Speaker: YANG YANG (Fudan University)

44 Integrated simulation of the impacts of resonant magnetic perturbations on tungsten radiation on EAST

Tungsten (W), widely used in tokamak plasma-facing components due to its high melting point, poses a critical challenge for fusion performance due to potential core accumulation and associated radiative losses [1]. Recent experiments on EAST have observed a remarkable 70% reduction in core W radiation following the application of resonant magnetic perturbation (RMP) fields [2], yet the underlying mechanisms remain insufficiently understood. To investigate this phenomenon, an integrated simulation framework combining the MARS-F [3], EMC3-EIRENE [4], and OMFIT [5] codes has been developed to model W sputtering, edge-core transport, and radiation under RMP fields. The MARS-F code evaluates the vacuum and plasma response fields to determine the perturbed edge magnetic topology, which is then used in EMC3-EIRENE to simulate 3D edge plasma and impurity dynamics, including a newly implemented sputtering model based on varying plasma parameters. Core W transport and radiation are simulated using the STRAHL code within OMFIT, with boundary conditions self-consistently coupled from the edge simulation.
Simulation results reveal that while RMP fields increase both erosion source and W radiation when only divertor erosion and magnetic topology are considered, a significant reduction in W radiation is reproduced only when enhanced perpendicular transport of W impurities is taken into account. A parameter scan demonstrates that increasing the perpendicular diffusivity of W by an order of magnitude leads to a ~70% reduction in core radiation power, aligning with experimental observations. This study highlights the dominant role of impurity perpendicular transport in impurity control under RMP fields and underscores the necessity of integrated edge-core modelling for accurate interpretation of impurity behavior in future reactor-relevant scenarios.

[1] Loarte A, Koechl F, Leyland M J et al 2014 Nuclear Fusion 54 123014
[2] Vogel G, Zhang H, Shen Y et al 2018 IEEE Transactions on Plasma Science 46 1350-5
[3] Liu Y Q, Bondeson A, Fransson C M et al 2000 Physics of Plasmas 7 3681-90 3681
[4] Feng Y and Kisslinger J 2000 Contributions to Plasma Physics 40 271-5
[5] Meneghini O, Smith S P, Lao L L et al 2015 Nuclear Fusion 55 083008

Speaker: Zihao Gao

Lunch

Sensor Fusion and Integrated Data Analysis

Conveners: Rainer Fischer, Sehyun Kwak (Max-Planck-Institute for Plasma Physics)

45 Applications of Bayesian data analysis at W7-X stellarator

Wendelstein 7-X (W7-X) is a superconducting optimized stellarator built in Greifswald/Germany which started its first operation with limiter plasmas in 2015. Since 2022 it is being operated with fully water-cooled first wall components including high heat flux graphite divertors, allowing quasi steady-state plasma operation. Approx. 50 diagnostic systems are applied to get insights into the physics phenomena of the intrinsically 3D shaped stellarator plasma. Analysis of large amount of data provided by various types of diagnostics/sensors sampling the plasma at different positions poses a big challenge in many modern large-scale nuclear fusion experiments. This complexity can be handled by application of probabilistic data analysis methods based on the Bayes’ theorem in which all statistical and systematic uncertainties of the diagnostic setup itself, of the model parameters as well as interdependencies of global physics parameters can be incorporated providing reliable uncertainties of the inferred quantities as well as correlations between them. Several W7-X diagnostic models have been implemented in the Minerva scientific modelling framework. A few newer applications based on spectroscopic measurements will be presented, for example tomographic reconstruction of 2D impurity radiation profiles in the W7-X divertor plasma [1].

[1] M. Krychowiak et al., Gaussian Process Tomography of carbon radiation in the transition to detached plasmas in the Wendelstein 7-X stellarator, Proceedings of EPS conference 2021

Speaker: Maciej Krychowiak (Max-Planck-Institute for Plasma Physics)

46 Uncertainties of magnetic equilibrium reconstructions

Reliable magnetic equilibrium reconstruction is key for the interpretation of experimental data and for physics modelling. Uncertainties of equilibrium quantities are frequently not provided although essential for the validation and quantification of derived physical quantities. A Monte-Carlo approach applicable to a free-boundary equilibrium reconstruction is suitable to provide uncertainties on any scalar, profile or flux-surface magnetic quantities.
The Monte-Carlo approach is aplied to the IDE equilibrium code which couples a kinetic free-boundary Grad-Shafranov solver with a solver of the current diffusion equation [1,2]. The set of external and internal measurements available to constrain the equilibrium are thereby extended by a flux-surface averaged current density provided by the current diffusion. The covariance matrix evaluated from the response function coupling the equilibrium coefficents with the set of measurements is employed in a Monte-Carlo sampling approach to estimate any equilibrium quantity of interest. The set of scalar quantities comprise, e.g., the geometry coordinates of the magnetic axis, the saddle points, dedicated points on the separatrix, the strike-line positions, volume-averaged coordinates, shape coordinates such as upper and lower triangularity, plasma current, Wmhd, plasma self inductivity, various beta values, dedicated q-values, and the distance between the 1st and 2nd X-point. The set of profile quantities comprise, e.g., the current density profile, and the q- and magnetic shear profile. The list of physical quantities can easily be extended as with the applied method the uncertainty of any magnetic quantity can be estimated. The evaluation time is below 1 s for a single equilibrium time point with room for significant improvement applying parallelization techniques.
[1] R. Fischer et al. 2016 Fusion Sci. Technol. 69 526–36
[2] R. Fischer et al. 2019 Nucl. Fusion 59 056010

Speaker: Rainer Fischer

47 Bayesian inference of plasma and impurity parameters with visible spectroscopy at ITER

Effective impurity control is essential for sustaining high-performance plasma operation in magnetic confinement fusion devices such as ITER. The visible spectroscopy reference system (VSRS) is designed to measure visible radiation, including bremsstrahlung, spectral lines, and synchrotron radiation, for the inference of key plasma parameters, such as the effective ion charge $Z_\mathrm{eff}$, electron density $n_\mathrm{e}$, impurity content, and the presence of runaway electrons. This study presents the development, implementation, and validation of a Bayesian model for the VSRS within the Minerva scientific modelling framework. A real-time application based on spectrally integrated polychromator signals has been developed for fast $Z_\mathrm{eff}$ inference to enable integration into plasma control systems. More advanced applications make use of full survey spectrometer data, incorporating spectral line contributions through asymmetric predictive distributions and anomaly detection techniques. Validation is conducted using synthetic data generated from IMAS-stored scenarios, including equilibrium and profile data, and cross-verified with the camera and spectroscopy emission ray tracer CASPER simulations. Furthermore, the model is extended to extract bremsstrahlung-subtracted spectra using Bayesian techniques, thereby enabling the automated identification of weak impurity lines. Feasibility studies on the detection of synchrotron radiation from runaway electrons have also been undertaken, demonstrating the potential of the VSRS to support rapid disruption mitigation strategies at ITER.

Speaker: Sehyun Kwak (Max Planck Institute for Plasma Physics)

20250909 IAEA TM SehyunKwak.pdf

48 Bayesian inference and Gaussian Process for fusion diagnostic data analysis

International Thermal Nuclear Reactor (ITER) will be equipped with a large array of diagnostics and produce a huge amount of data characterized by redundant, complementarity and complex errors. A prominent challenge for ITER as well as other fusion devices, is how to make the best use of such a large amount of data to obtain as much useful information as possible while meeting the requirements of accuracy, computation speed and reliability assessment against data and model uncertainty for specific tasks such as physics study and real-time control. In this talk, we will give a brief review of the recent developments and applications of advanced data analysis techniques based on Bayesian inference [1] and Gaussian Process [2], including Non-stationary Gaussian Process Tomography (NSGPT) method for tomographic reconstruction [3] and Gaussian Process Regression (GPR) for plasma parameter profile inference [4], as well as others. By contrast, these statistical methods possess advantages of uncertainty quantification for reliability assessment and high flexibility in incorporating domain knowledge into the prior model for improved inference. We will introduce briefly the Bayesian approach to an integrated analysis of multiple sources of data from heterogeneous diagnostics, by which improved reliability and consistency of the results can be obtained via the synergistic effect. This study is expected to provide instructive reference not only for magnetically confined fusion research but also for other fields where advanced analysis techniques are essential.

References
[1] Udo Von Toussaint, "Bayesian Inference in Physics" Rev. Mod. Phys. 83 943 (2011).
[2] C.E. Rasmussen, "Gaussian Processes in Machine Learning" Springer-Verlag, Heidelberg (2004).
[3] Dong Li, J. Svenssen et al, "Bayesian soft X-ray tomography using non-stationary Gaussian Processes" Rev. Sci. Instrum. 84 083506 (2013).
[4] M.A. Chilenski, et al, "Improved profile fitting and quantification of uncertainty in experimental measurements of impurity transport coefficients using Gaussian process regression" Nucl. Fusion 55 023012 (2015).

Speaker: Prof. Dong Li

Coffee Break

Sensor Fusion and Integrated Data Analysis

Conveners: Rainer Fischer, Sehyun Kwak (Max-Planck-Institute for Plasma Physics)

49 Bayesian inference of 2D tungsten concentration profiles at WEST using soft X-ray and bolometer diagnostics

Tungsten (W) will be used in the ITER tokamak for its divertor and first wall. The strong plasma radiation from heavy impurities like tungsten may cause significant power loss and can pose a major risk to the energy confinement. Reliably estimating the tungsten concentration in a fusion device is therefore critical for transport studies and active impurity control. This is often achieved by analyzing the line-integrated plasma radiation recorded by various diagnostics, such as soft X-ray (SXR), bolometry and extreme ultraviolet (EUV). However, results from independent analyses of individual diagnostics often exhibit inconsistencies. A coherent combination of the data from multiple diagnostics can be achieved using the integrated data analysis (IDA) approach based on Bayesian probabilistic theory. This provides joint estimates of tungsten concentration profiles, as well as kinetic profiles (temperature and density), together with their uncertainties. The joint posterior probability distribution of the profiles must be explored by a Markov chain Monte Carlo (MCMC) sampler. The difficulty of sampling from the high-dimensional and strongly correlated joint posterior distribution has been resolved in this work by applying reparameterization. First results using synthetic data and experimental data from WEST indicate the strong potential of this method.

Speaker: Mr Hao Wu (Ghent University)

Thu_Sep_11_Hao Wu.pptx

Assembly for excursion

Excursion

Friday 12 September

Registration

Uncertainty Propagation, Verification and Validation

Convener: Simon Pinches (ITER Organization)

50 Bayesian modelling of MAST Upgrade CXRS diagnostic (CELESTE)

Stimulated emission by neutral beam (NB) injection is of key importance for diagnosing the state of the core plasma in tokamaks. This paper reports on the development of a new Bayesian model of a spectral diagnostic system with NB injection. Implemented within the Minerva framework and with data entry via the ITER Integrated Modelling & Analysis Suite (IMAS), the model provides a generic probabilistic treatment that can be applied to any appropriate Tokamak device subject to availability of suitable IMAS database instances. A feature of the model is that the NB injection includes the propagation of the beam from the final accelerator grid of the NB source. Downstream distributions, taking account also of the attenuating effects of edge baffles, are expressed as a summation of analytic distributions obtained by coordinate transformation (L. C. Appel, Computer Physics Communications, Vol 312, 2025). Continuing model developments incorporate NB reionisation due to cold mono-atomic and diatomic deuterium neutrals in the duct extending between the NB calorimeter and the core separatrix. A model of NB reionization is necessary without direct measurement of beam power at the plasma separatrix. Finally, the model incorporates a Collisional Radiative Model (S. Bannmann et al 2023 JINST 18 P10029) which gives an inference of both the spectral line emission and the beam attenuation through the plasma itself. Results will be presented demonstrating the use of Beam Emission Spectra for calibrating the sight-lines geometry of the CELESTE (N. J Conway et al. Rev of Scientific Instruments, Vol 77, 2006) diagnostic and determination of plasma minority temperatures, densities and toroidal rotation.

Speaker: Lynton Appel (UKAEA)

51 The uncertainty quantification of plasma equilibrium calculation on Experimental Advanced Superconducting Tokamak (EAST)

The precision of plasma equilibrium is a pivotal issue in the field of fusion plasma physics and the operation of tokamak plasmas. During the plasma equilibrium uncertainty quantification (UQ) campaign, input parameters for the equilibrium solver are assigned default value ranges derived from EAST experimental data. These input parameters are subsequently processed by an encoder, decoder, and analyzer within a sequence-to-sequence deep learning model. The outcomes of the UQ campaign provide insights into the calculated equilibrium output uncertainty and its sensitivity analysis. Drawing from the results of this sensitivity analysis, the rankings of influence for equilibrium input parameters on the calculated safety factor q profile, plasma shape, toroidal field, plasma pressure, and 0D outputs are presented. Meanwhile, the sensitivity of the magnetic probes on the equilibrium calculation for the EAST double null configuration is analyzed. By analyzing the magnitude of uncertainties in plasma equilibrium calculations, on the one hand, the sources of error can be identified for subsequent diagnostics and physical analysis, thereby improving the accuracy and robustness of the final results. On the other hand, the sensitivity analysis of parameter influence can provide valuable insights for optimizing the plasma discharges platform.

Speaker: Shuzhi Yuan (Institute of Plasma Physics, Hefei Institutes of Physical Science, Chinese Academy of Sciences)

Information Retrieval and Visualisation

Convener: Geert Verdoolaege (UGent)

52 Development of the IMAS Data Model Towards an Open Standard for Fusion Data

The cornerstone of the Integrated Modelling & Analysis Suite (IMAS) developed by ITER is its machine-agnostic data model, called the Data Dictionary (DD). It was recently made open access [1] to permit a broader adoption outside of the ITER Members and by private fusion ventures. Discussions and contributions are welcome and held publicly in GitHub, either for new extensions, clarifications or changes to the Interface Data Structures (IDS) that compose the DD.

While it aggregates decades of effort to standardize the description of fusion data for both simulations and experiments, it has been found to introduce a maintenance burden on adapted codes through the frequent introduction of non-backward compatible changes introduced in IDS that are still in an alpha development stage. The major release of DD version 4 in 2024 improved the situation with a 50% increase of the number of stable IDSs, focusing on diagnostic and sub-system IDSs, while in version 3 the focus was on IDSs describing simulations and processed data. To go further, ongoing efforts are exploring more disruptive approaches based on lessons learned from the historical design choices.

A first effort focuses on a better separation of the Data Model and the IDS interfaces designed in support of integrated modelling. It relies on the definition of Standard Names that can be used as metadata to describe physical quantities and their units without restricting the dimensionality or coordinates associated with the data. The Standard Names approach is adapted from the CF-Conventions [2] which is successfully applied to climate and geoscience data over 20+ years. This approach aims at reducing the complexity of the Data Model and improving backward compatibility and is intended to be complementary to the IDS definitions.

A second effort focuses on improving the storage format for IMAS data with the intention to lower maintenance needs by simplifying the software stack and improve performance, particularly for remote data access. The main objective is to define a self-describing data format that keeps metadata together with data and to use open-source tools and libraries for access and manipulation. Early comparisons are made between netCDF, Zarr and the legacy solutions based on MDSplus, HDF5 and UDA for remote access.

[1] https://github.com/iterorganization/IMAS-Data-Dictionary
[2] https://cfconventions.org/

Speaker: Olivier Hoenen (ITER Organization)

IMAS-Data-Model-Open-Standard.pptx

53 Global Data Sharing: A Comparative Study of Pelican and CVMFS for remote access to MASTU and DIII-D data for UFO detection workloads

Data aggregation across multiple fusion devices has enormous value for improving machine learning models and for validating simulation tools. One challenge in forming and using such datasets can simply be the latency caused by the distance between experimental sites and the computational facilities where data is used. Other challenges arise from the different data access interfaces exposed by each data provider, and the formats and representations of that data.

Pelican Platform[1] and CVMFS[2] are both examples of data distribution services which provide consolidated access interfaces and use caches and data mirrors to reduce latency when accessing multiple, globally distributed data sources. Pelican is a data federation platform which aims to unify access to different kinds of storage backends (S3, Posix, HTTP) through adaptor services, it uses XRootD for data transfer and includes features for user authorisation and authentication. CVMFS is a CERN-developed data distribution technology widely used in High Energy Physics, initially focused on sharing software, it uses the HTTP protocol for data transfer with a convenient virtual filesystem interface and aggressive use of local caching to improve performance for certain data access patterns.

In this presentation we will cover our experiences to date comparing these two potential infrastructure frameworks and how we made use of them to share fusion data between data sources and HPC facilities at both UKAEA and DIII-D. This data sharing involves the use of SciTokens[3] (a federated authentication and authorisation infrastructure) that has been used at DIII-D to enable remote connection to their dataset via the US DOE-funded Fusion Data Platform [5], an initiative led by General Atomics.

We use a UFO detection tool, ENEJETIC[4], as a demonstration HPC workload for processing remote image data from both the MAST and DII-D tokamaks. UFOs are a class of impurity within the plasma that can lead to damaging disruption events and are often formed from debris coming off the wall of the vessel. They are difficult to monitor during operation without human intervention, but quick responses are required to avoid disruption. Originally trained from JET image data, ENEJETIC (Enhanced Neural Engine for JET Image Classification) uses a ​​convolutional Neural Network to automate the process of the detection and logging of these events.

[1] Pelican Platform, https://pelicanplatform.org/
[2] CernVM File System, https://cvmfs.readthedocs.io/en/stable/
[3] SciTokens, https://scitokens.org/
[4] Phys. Plasmas 32, 042508 (2025) https://doi.org/10.1063/5.0261120
[5] Fusion Data Platform, https://ga-fdp.github.io/

Speaker: Stephen Dixon (UKAEA)

54 The IAEA Fusion Data Lake Project - Accelerating AI and Big Data Applications through Open Science and FAIR Data

The application of AI technologies is an ever-growing area of research within the development of fusion energy. The technology plays a key role in the production of surrogate models and digital twins that enable accelerated development by supplementing the use of computationally expensive 2D and 3D simulation codes, defining non-parametric models for insufficiently characterised phenomena, and computationally cheap alternatives for real-time applications [1]. The IAEA is supporting the fusion communities' efforts in this area with the AI for Fusion Coordinated Research Project (CRP), a five-year initiative launched in 2022, which involves 24 institutions across 11 countries [2]. A key goal of the CRP is to support the development of modern, scalable, and accessible data infrastructure that is required to produce diverse and rich data sets, which are required to develop generalised AI models that are machine agnostic and can safely extrapolate into the parameter space of future fusion power plants.
The IAEA is playing an active role in contributing to the data infrastructure with the Fusion Data Lake project. A modern data platform to enable the development of AI workflows in line with FAIR (Findable, Accessible, Interoperable, and Reusable) data principles. The platform comprises three major components:
1. An international data catalogue;
2. A centralised medium-term storage; and
3. A data federation of the various fusion data platforms around the world.
A proof of concept (PoC) has been built, which demonstrates the data cataloguing and federation capacity by integrating with the UKAEA’s MAST Data Catalog [3]. Currently, the second phase of the PoC will involve further generalisation of the data pipeline codebase and the ingestion of shot catalogues from two additional experimental fusion devices.
This presentation will present a high-level overview of the Fusion Data Lake project, including:
● Technical architecture and design, collaborations and contributions, and the PoC solution;
● Data and metadata model development and ontological concepts; and
● The approach to data governance and terms of service.
This presentation will illustrate the approach, results, and direction of the work, highlighting the high potential value to the fusion community of increasing the visibility and accessibility of the numerous international experimental data sets.
References
1. P. Brans, "AI ignites innovation in fusion", ITER Organization, 2025, website https://www.iter.org/node/20687/ai-ignites-innovation-fusion (accessed 5th September 2025)
2. AI for Fusion, International Atomic Energy Agency, 2025, website https://nucleus.iaea.org/sites/ai4atoms/ai4fusion/SitePages/AI4F.aspx (accessed 5th September 2025)
3. S. Jackson et al. 2025, IEEE Trans. on Plasma Sci., doi: 10.1109/TPS.2025.3583419. https://ieeexplore.ieee.org/document/11128905

Speakers: Daljeet Gahle (IAEA), Daljeet Singh Gahle (International Atomic Energy Agency)

Coffee Break

Round-table discussion

Conference closing

Lunch