Predicting τₑ Scaling Using Interpretable Neural Networks & Bayesian Methods

This poster explores utilizing neural networks (NNs) and Bayesian methods to construct an energy confinement time scaling law with integrated epistemic and aleatoric uncertainty intervals leveraging the Global H-Mode Confinement Database v5.2.3^[1]. Accurate and interpretable scaling laws are critical for predicting the energy confinement time of next-generation tokamaks such as ITER and SPARC. Traditional methods using regression models assume no interdependencies of various parameters across the whole parameter space. However, both experimentally and theoretically there are linear and non-linear interactions between the various parameters. These interdependencies can be captured by neural networks and should lead to an uncertainty reduction. In addition, the Expected Gradients^[2] method, related to Shapley (SHAP) values, provides information on the weight of each parameter in the trained NN. This provides an opportunity to develop new interpretable scaling laws, purely driven by data, relaxing prior assumptions embedded in physics models. Additionally, the NN is combined with Bayesian uncertainty quantification for reliable uncertainty bounds, which is crucial to estimate the performance of future tokamaks.

[1] G. Verdoolaege et al 2021 Nucl. Fusion 61 076006

[2] M. Sundararajan et al 2017 ICML 70