Machine Learning Solutions
AI and ML are essential to match the complexity and speed of fusion. Our models are grounded in physics, trained on our own simulators, and validated on real devices.
Focus Areas
Our machine learning engineers work closely with physicists and fusion software developers to build solutions that are grounded in proven science and either device-agnostic or easily transferable across devices. We also focus on methods relevant to new fusion devices that start without a historical dataset of discharges to train models on.
Plasma control with reinforcement learning
We develop and deploy RL-based controllers for plasma shape and position, trained inside the NSFsim simulation environment and validated directly on DIII-D — eliminating the need for equilibrium reconstruction and enabling high-speed control execution.
Plasma parameters reconstruction
Combining model-based and machine-learning methods for real-time plasma parameters reconstruction and uncertainty quantification — from neural network reconstruction of plasma boundaries to L-H mode classification using reactor-relevant ECE diagnostics data.
Surrogate models and physics acceleration
Physics-informed neural networks trained on integrated modeling ensembles to replace expensive physics codes with fast, accurate surrogates — enabling real-time inference for plasma control, rapid scenario screening, and uncertainty quantification at scales impractical with first-principles simulations.
Discharge scenario builder
Combining physics-based simulation with machine learning to automate discharge scenario development — learning from large ensembles of NSFsim runs to optimize plasma current ramp-up, heating schemes, and confinement targets while respecting hardware and stability constraints.
Relevant Reading
Our ML work spans reinforcement learning for plasma control, confinement state classification, and neural network reconstruction — all developed in close collaboration with physicists and validated on real devices such as DIII-D.
Controlling plasma temperature and safety factor with gyrotrons using reinforcement learning
Collaboration with the DIII-D National Fusion Facility — training a reinforcement learning agent to control plasma temperature and safety factor using the ECRH gyrotron system.
Paper ↗Plasma confinement state classification via FPP relevant microwave diagnostics
Classifying plasma confinement states using fusion power plant-relevant microwave diagnostics, enabling real-time monitoring without direct profile measurements.
Paper ↗Neural network reconstruction of the DIII-D tokamak plasma boundary using a reduced set of diagnostics
Collaboration with the DIII-D National Fusion Facility — reconstructing the tokamak plasma boundary in real time using a neural network trained on a reduced set of magnetic diagnostics.
Paper ↗Reconstruction-free magnetic control of DIII-D plasma with deep reinforcement learning
Developed with the DIII-D National Fusion Facility — a reconstruction-free deep reinforcement learning controller for magnetic plasma control, deployed and tested on the live tokamak.
Paper ↗Robustness by design: Interface contracts for AI control in high-stakes physical systems
Formal interface contracts for AI control in high-stakes physical systems, establishing verifiable safety and robustness guarantees for plasma control deployments.
Blog ↗Developing an ML-based Surrogate Model for Plasma Boundary Prediction
Development of a transformer-based surrogate model for real-time plasma boundary prediction at DIII-D — achieving 95% accuracy within 1.3 cm mean error and enabling fast inference for control and scenario screening without running the full equilibrium solver.
See the publications page to learn more about our research and work.
Have an ML Problem to Solve?
Tell us about your challenge and we'll outline how machine learning can help — no obligation.
Get in touch