Machine learning based energy confinement time extrapolation via multi-tokamak database

초록

Predicting energy confinement time in future fusion devices like ITER is a significant challenge for traditional methods. This study introduces an advanced machine learning framework to address this. Our approach utilizes deep ensembles for robust uncertainty quantification and a novel feature alignment technique. This technique addresses the problem of covariate shift, where feature distributions differ significantly across devices, thereby mitigating the severe challenge of extrapolation. By creating a more unified representation of plasma parameters across diverse devices, the model's ability to generalize and make reliable predictions for new, unobserved conditions is substantially enhanced. The developed model shows improved predictive capability compared to existing approaches and offers a powerful tool for exploring and optimizing future reactor designs and operational scenarios, including the crucial impact of factors like wall materials. This work highlights the potential of sophisticated data-driven methods to advance fusion energy research. © 2025 The Authors

키워드