Speaker
Description
Fusion is fundamentally cutting-edge, and to achieve economic fusion energy the field must advance the understanding of engineering, materials, and plasma phenomena through experiments and test facilities. When designing a facility or experiment, traditional approaches of diagnostics, control, or experimental setup can often rely on manual or intuitive decision making. This can often be very time consuming and expensive and may pose significant risks to the aims of the experiment. In this work we have developed a novel methodology for designing fusion experiments using Bayesian experimental design.
Bayesian experimental design is a framework to assess the ability of a system to be measured and controlled, based upon Bayesian theory and probabilistic AI methods such as Gaussian Processes. The framework inherently accounts for uncertainties in diagnostics, controllers, and models. It can be used by diagnosticians or decision makers to assess the uncertainty or risk of a design, or make automated designs, such as the viewing angle or placement of a camera. Crucially, this is done through the lens of information gain - a measure of how well a sensing system can differentiate between different simulated states within uncertainty.
Practically, this novel approach represents a fundamental shift in how sensors and experiments are designed, to be more flexible, integrated, and holistic. Flexible, in that the single framework can answer many diverse questions about a design, including the uncertainty of key quantities of interest, or how sensing performance reduces under failure. Integrated, in that it determines the added information gained by combining data from sensors which may conventionally be kept separate. And holistic, in that it assesses a sensor set’s ability not to measure one value, but to distinguish between different possible states of the fusion experiment as an entire system.
The benefits of this Bayesian design framework make it seamlessly compatible with digital twins—virtual counterparts of physical fusion devices. In fact, the basic structure of a digital twin is a set of experimental data that is compared to simulations of the real asset, to identify which simulated state is most likely and track this digital version of the asset. Digital twins also propose to link information from disparate diagnostics together. The proposed framework designs exactly for this, analysing the ability for a complex integrated diagnostic system to distinguish between simulated states. Though digital twins are not the only application of this framework, they represent an application that is in need of modern design tools.
To validate the proposed method, the UKAEA and digiLab have applied the Bayesian design software to challenges across fusion. For example, this framework was used to assess how uncertainties in equilibrium reconstruction of MAST are impacted by various levels of sensor failure. Additionally, this framework was applied to automatically recommend optimal integrated designs of thermocouples on materials test facilities. Though this initial demonstration work shows the capabilities of Bayesian design, this tool can be applied to a wealth of other areas in the fusion.
| Country or International Organisation | United Kingdom |
|---|---|
| Affiliation | digiLab |
| Speaker's email address | cyd@digilab.co.uk |
