MIT Researchers Develop AI-Powered Model to Predict and Stabilize Plasma Rampdown in Fusion Tokamaks

Scientists at MIT have developed a new prediction model that could significantly improve the safety and reliability of fusion power plants by better managing the shutdown of plasma in tokamaks. Tokamaks, donut-shaped devices that use magnetic fields to confine superheated plasma, aim to replicate the fusion process that powers the sun. One of the major challenges in operating these machines is safely turning off the plasma current—often circulating at speeds of up to 100 kilometers per second and at temperatures exceeding 100 million degrees Celsius—without triggering damaging instabilities. When plasma becomes unstable, operators must ramp down the current to prevent damage. However, this process itself can sometimes cause disruptions that scar the tokamak’s interior, requiring costly repairs. To address this, MIT researchers led by graduate student Allen Wang created a hybrid model combining machine learning with physics-based simulations of plasma behavior. This approach allows the model to predict how plasma will respond during rampdown with high accuracy using relatively small amounts of experimental data—critical given the high cost and limited availability of tokamak experiments. The team trained and tested the model using data from the TCV tokamak, a small experimental device operated by the Swiss Plasma Center at EPFL. The model was able to learn how plasma evolves under different rampdown conditions using just a few hundred low-performance pulses and a handful of high-performance ones. This efficiency is a major advantage over purely data-driven machine learning models, which would require vast datasets to achieve similar results. The researchers also developed an algorithm that translates the model’s predictions into actionable control trajectories—real-time instructions for adjusting magnetic fields or heating to maintain stability. When tested on TCV experiments, the system successfully guided plasma rampdowns without disruptions, often faster and more smoothly than conventional methods. “We ramped the energy down to nothing,” Wang said. “We did it multiple times with statistical confidence that we made things better.” The work, published in Nature Communications, is part of a broader effort to make fusion energy a practical and reliable power source. The model’s success marks a key step toward enabling safe, repeatable operation of future large-scale fusion reactors. The research was supported by Commonwealth Fusion Systems (CFS), an MIT spinout building SPARC, a compact tokamak designed to achieve net-energy fusion. The team is now working with CFS to integrate the model into real-time control systems for next-generation fusion devices. The project also received backing from the EUROfusion Consortium, the Euratom Research and Training Program, and the Swiss State Secretariat for Education, Research, and Innovation. As Wang noted, while fusion remains a long-term challenge, this work represents meaningful progress toward making fusion energy routinely viable.