How machine learning is shaping fusion reactor control
Federico Felici and Jonas Buchli, former researchers at Google DeepMind, have founded Fusionality to develop control systems and simulation tools for fusion energy startups. Based in Zurich, the company emerged after repeated conversations with fusion companies revealed a critical gap in available technology for managing plasma stability and reactor operations. Their goal is to provide ready-made components that fusion firms can integrate directly into their control architectures, avoiding the need to build everything from scratch.
Breaking news
London-based CloudNC raises €17.2 million to scale AI-powered precision machining
AI Hardware Weekly: Chip Efficiency, Digital Twins, and Mobile Vision Models Lead September Discussions
AI Safety Researcher Quits, Warns Labs Racing Toward Dangerous Superintelligence
Meta launches Muse, a personal AI agent that books, buys and negotiates for youFusionality focuses on creating software environments that simulate fusion plasma behavior and enable real-time control adjustments. The founders drew on their expertise in machine learning and robotics from DeepMind to design systems capable of handling the complex, high-speed dynamics of fusion reactions. By offering modular tools, they aim to reduce development time and costs for startups working on tokamaks and other fusion devices. Felici noted that many fusion companies expressed frustration over the lack of commercial suppliers for essential control hardware and software, saying they would „ideally buy many of the components” if such options existed.
What challenges remain for scaling fusion power?
The startup’s approach leverages reinforcement learning techniques refined during their time at DeepMind to predict and stabilize plasma states within fusion reactors. These systems process sensor data in milliseconds to adjust magnetic fields and heating inputs, preventing disruptions that can damage equipment. Early simulations show their control algorithms can maintain stable plasma conditions longer than conventional methods, a key step toward net energy gain. Fusionality is currently testing its software with partner fusion companies in Europe and the United States, though specific names were not disclosed due to confidentiality agreements.
Despite progress in control systems, significant hurdles remain before fusion can contribute meaningfully to the grid. These include achieving sustained energy output, managing neutron radiation damage to reactor walls, and reducing the cost of superconducting magnets. Fusionality’s tools address only one layer of the fusion stack, but the founders believe reliable control is a prerequisite for scaling any fusion design. They plan to expand their simulation capabilities to include material behavior and thermal dynamics in future updates, aiming to offer a more comprehensive platform for reactor design and operation.
Frequently Asked Questions
How does Fusionality’s technology differ from existing fusion control systems? Fusionality provides modular, software-based control tools built using machine learning, allowing fusion startups to integrate advanced stabilization features without developing them internally. Unlike custom-built systems, their tools are designed for broader compatibility across different reactor types.
Can Fusionality’s tools help fusion reactors achieve net energy gain? While the startup’s control systems improve plasma stability and operational efficiency, achieving net energy gain depends on multiple factors including fuel density, confinement time, and heating efficiency. Their technology supports these conditions but does not guarantee them on its own.
