Berkeley Lab Researchers Use Machine Learning to Accelerate Fusion Energy Progress Through Real-Time Plasma Heating Predictions

Researchers from Berkeley Lab, Princeton Plasma Physics Laboratory, MIT, San Francisco State University, and CEA-IRFM collaborated to develop machine learning models that accelerate predictions of plasma heating in fusion energy devices. These models, trained on NERSC’s Cori and Perlmutter supercomputers, enable real-time adjustments during tokamak operations by predicting particle heating via radiofrequency waves six orders of magnitude faster than traditional simulations. The research, published in Nuclear Fusion, focuses on ion cyclotron range of frequencies (ICRF) heating and aims to advance fusion energy progress by overcoming computational bottlenecks, supported by the U.S. Department of Energy.

Researchers from multiple institutions collaborated to develop machine learning models aimed at advancing fusion energy research. This interdisciplinary effort involved teams from Berkeley Lab, Princeton Plasma Physics Laboratory, MIT, San Francisco State University, and CEA-IRFM, each contributing expertise to enhance predictive capabilities in plasma heating.

The collaboration focused on creating models that significantly outperformed traditional simulations by achieving speeds six orders of magnitude faster. These advancements were critical for enabling real-time adjustments during experiments, a capability essential for optimizing fusion reactor performance.

Supercomputers at the National Energy Research Scientific Computing Center (NERSC), specifically Cori and Perlmutter, played a pivotal role in this research. These high-performance computing resources facilitated the intensive simulations required to build the database and optimize machine learning algorithms.

The models were applied to specific fusion devices such as NSTXU and WEST, demonstrating their effectiveness in predicting plasma behavior under ion cyclotron range of frequencies (ICRF) heating. This application highlighted the potential for machine learning to revolutionize fusion energy research by providing rapid, accurate simulations that could accelerate progress toward sustainable fusion power.

The findings underscored the versatility of these computational tools, suggesting broader applications across various challenges in fusion energy. By overcoming longstanding computational limitations, this research offers a promising path forward for achieving practical fusion energy solutions.

Broader Implications Of Machine Learning In Advancing Fusion Energy Research

Researchers from Berkeley Lab, Princeton Plasma Physics Laboratory, MIT, San Francisco State University, and CEA-IRFM collaborated to develop machine learning models for predicting plasma heating behavior under ion cyclotron range of frequencies (ICRF). These models achieved processing speeds six orders of magnitude faster than traditional simulation methods while maintaining high accuracy, enabling real-time adjustments during experiments.

Supercomputers at the National Energy Research Scientific Computing Center (NERSC), including Cori and Perlmutter, were instrumental in this research. These resources supported intensive simulations required to build the database and optimize machine learning algorithms, particularly for surrogate model training and hyperparameter optimization.

The models were successfully applied to specific fusion devices such as the National Spherical Torus Experiment Upgrade (NSTXU) and WEST, demonstrating their effectiveness in predicting plasma behavior under ICRF heating. This application highlights the potential of machine learning to accelerate progress toward sustainable fusion power by providing rapid, accurate simulations.

The findings suggest broader applications across various challenges in fusion energy research, offering a path forward for overcoming longstanding computational limitations and potentially reducing the timeline to practical fusion energy solutions.