Researchers around the world are starting to realize the game-changing potential for machine learning (ML) in fusion. Inherently, fusion energy research involves tremendously large and complicated datasets, requiring enormous manual effort to analyze. By adapting developments in ML technology from outside of fusion to the fusion landscape, we can make rapid progress in several outstanding fusion applications, including:

• ML-Based, Real-time Control Schemes: ML algorithms, especially deep learning and surrogate models, can rapidly analyze sensor data from the reactor and adjust magnetic fields and other control mechanisms in real-time. This allows for faster reaction to instabilities, minimizing disruptions or plasma loss.
• Predicting and Preventing Disruption: ML models can be trained on vast amounts of historical data from fusion experiments to predict when a disruption is likely to occur. By identifying patterns that precede disruptions, we can institute protective actions to steer the plasma away from disruption limits.
• Optimizing Reactor Design: Designing a fusion reactor involves a vast design space with countless variables, such as magnetic field configurations, material properties, and cooling systems. ML can help optimize the design parameters more efficiently than traditional methods by providing mechanisms to assess a wide range of configurations in short time scales without a loss of fidelity.
• Accelerating Scientific Discovery: Theoretical models of fusion reactions are incredibly complex and require high-performance computing resources. By integrating ML with physics-based models to develop surrogates, we are working to improve the accuracy of simulations while significantly reducing computation time, allowing for faster experimentation and optimization.
• Data-Driven Insights from Experiments: Fusion experiments generate terabytes of data per second, making it impossible for human researchers to manually analyze it all. ML can efficiently extract patterns, correlations, and anomalies from this enormous volume of data, providing new insights into plasma behavior that might have been missed using traditional analysis methods.

As an example of a successful application of ML techniques to fusion data analysis, consider the following situation. The DIII-D tokamak has numerous diagnostic suites, some with higher temporal resolution and some with higher spatial resolution. However, there are many important physical phenomena, such as ELMs that require both rapid and highly spatially-resolved measurements to accurately study. Typically, we have not been able to capture the effects of ELMs and other rapid events in enough detail to develop predictive models, but with novel ML techniques we can enhance the resolution of important diagnostics, as illustrated below:

The main diag-to-diag methodology.

In the above figure, we display a typical workflow for Multimodal Super-Resolution Diagnostics. Diagnostic A is essential to capture fast transient events near the edge of plasma, but due to its low temporal resolution and accuracy it fails to track the evolution of such events. The diag-to-diag methodology solves this problem by generating a synthetic super resolution reconstruction of Diagnostic A by learning the correlation between Diagnostic A data and other diagnostic measurements with higher resolutions and better accuracy. This allows us to reproduce fast transient events with a surrogate of Diagnostic A, leading to new scientific understandings of the edge plasma!

In the reactor design space, I am working with complex, data-driven and ML-informed engines like FUSE, an open-source framework for the integrated design of Fusion Power Plants, and Next Step Fusion’s Digital Twin for Tokamaks, which integrates several ML-based surrogate models for fusion performance into a time-dependent pulse optimizer. These tools allow us to more effectively traverse and predict the landscape of future fusion power plants, leading to new insights into the optimization of these complex machines. Additional studies on this topic will be coming out shortly, so stay tuned!

Selected publications on this subject:

Muli-variate Optimization of the Negative Triganularity Fusion Reactor Using Surrogate Models
Nelson, A.O., and Slendebroek, T., in prep (2025).

Multimodal Super-Resolution: Discovering hidden physics and its application to fusion plasmas
Jalalvand, A., Curie, M., Kim, S., Seo, J., Hu, Q., Curie, M., Steiner, P., Nelson, A. O., Na, Y.-S & Kolemen, E., Science, in review (2024).

Alfvén eigenmode classification based on ECE diagnostics at DIII-D using deep recurrent neural networks
Jalalvand, A., Kappatou, A. A., Garcia, A. V., Nelson, A. O., Abbate, J., Austin, M., Verdoolaege, G., Brunton, S. L., Heidbrink, W. W. & Kolemen, E., Nuclear Fusion 62, 026007 (2022).

Neural Dynamical Systems: Balancing Structure and Flexibility in Physical Prediction
Mehta, V., Char, I., Neiswanger, W., Chung, Y., Nelson, A. O., Boyer, M. D., Kolemen, E. & Schneider, J., 2021 60th IEEE Conference on Decision and Control (CDC), 3735–3742 (2021).