A wave of artificial intelligence-based applications is accelerating fusion energy research across several fronts: IBM has unveiled a foundational model for analyzing plasma behavior, a U.S. national laboratory has demonstrated an AI approach to materials development, and a British fusion company has deployed its own AI programming assistant for its scientists.

IBM's TokaMind Decodes Tokamak Plasma 

IBM, the UK Atomic Energy Authority (UKAEA), and the STFC Hartree Centre have open-sourced TokaMind, the world's first foundational AI model for fusion plasma. Trained on data from the UKAEA's Mega Ampere Spherical Tokamak (MAST), the model integrates approximately 40 different sensor signals with varying frequencies into a unified representation of plasma behavior within the tokamak.

At 9 million parameters, the model is small by generative AI standards, yet it outperformed traditional machine learning models—trained separately for each task—on nearly all 14 plasma analysis tests in the new TokaMark evaluation suite. The most significant advantage appeared in long-term forecasting. "It seems this really helped the model build a transferable representation of plasma dynamics," noted Tobia Boschi, an IBM researcher involved in training the model.

UKAEA scientists plan to expand TokaMind using data from the upgraded MAST-U facility and potentially other tokamaks; the findings are expected to aid in designing STEP, the UK's prototype fusion power plant planned for the 2040s. The UKAEA’s upcoming AI supercomputer, Sunrise, will support the next phase of this collaboration.

DuctGPT Targets the Fusion Materials Challenge 

Separately, on April 24, researchers from the U.S. Department of Energy’s Ames National Laboratory introduced DuctGPT—a generative transformer for predicting tensile ductility in refractory multi-principal element alloys. These materials are considered top candidates for the extreme conditions of a fusion reactor's first wall. The tool aims to solve a key problem: while refractory alloys have exceptional heat resistance, they are inherently brittle, which limits their practical use. Developed under the DOE ARPA-E CHADWICK program, DuctGPT is designed to accelerate the search for alloys that combine durability with mechanical resilience.

First Light Fusion Implements Specialized AI Coder

 In the UK, First Light Fusion announced on April 23 a contract with Locai Labs for a highly specialized large language model designed as an AI programming assistant for the company's scientific and engineering teams. The model was trained using Locai’s "Forget-Me-Not" framework based on First Light Fusion’s own scientific workflows. In February, the companies signed a memorandum of understanding to explore AI applications in inertial confinement fusion research, and this contract marks the first commercial outcome of that partnership.

"The computational challenges we face in fusion can only be solved by bringing together quantum computing, AI, and high-performance computing," said Alessandro Curioni, IBM Fellow, reflecting the broad momentum behind these efforts.