Improving fusion plasma predictions with multi-fidelity data science models

Tokyo, Japan (SPX) Dec 16, 2024 - The quest for fusion energy, a potential solution to global energy challenges, has advanced through innovative approaches to plasma performance prediction. Magnetic confinement fusion reactors rely on intricate systems that confine high-temperature plasma within powerful magnetic fields. While this effort integrates technologies like superconducting magnets and advanced heating devices, understanding and predicting plasma behavior remains a significant scientific challenge.

Energy and particle transport in confined plasma are studied through theory, numerical simulations, and experimental measurements. While physics-based simulations align with experiments to some degree, discrepancies persist, limiting their reliability. Empirical models based on experimental data offer insights, yet their application to future fusion scenarios remains uncertain due to limited data applicability.

Addressing this challenge, researchers at the National Institute for Fusion Science have introduced a multi-fidelity data fusion approach. This method uses nonlinear auto-regressive Gaussian process regression (NARGP) to improve the predictive accuracy of turbulent transport models in plasma. By integrating low-accuracy but abundant data with a limited set of highly accurate data, the method compensates for data scarcity in unexplored plasma conditions.

"The idea of NARGP is to express the prediction of high-fidelity data as a function of input and low-fidelity data," researchers explained. They applied this approach in three key scenarios: combining low- and high-resolution simulation data, predicting turbulent diffusion coefficients using experimental plasma data, and integrating simplified theoretical models with turbulence simulations. These applications demonstrated significant improvements in prediction accuracy.

The research bridges the gap between theoretical simulations and empirical data models. "This study combines the predictability of theory and simulation based on physical models with the quantitative information obtained from experimental data," the team added. By merging these approaches, the method provides a reliable pathway for predicting conditions in future nuclear fusion burning plasmas.

The implications extend beyond fusion research. The multi-fidelity modeling approach offers potential for developing predictive models across various fields where high-precision data are limited. It enables performance optimization and design innovation in fusion reactors and other complex systems, ensuring faster, more accurate predictions with minimal high-fidelity data.

The findings, published in *Scientific Reports*, represent a critical step toward efficient fusion reactor design and advancing scientific methodologies in other domains.

Research Report:Multi-Fidelity Information Fusion for Turbulent Transport Modeling in Magnetic Fusion Plasma