We’ve built ModeCast to predict linear instabilities in tokamaks: which modes grow, where they develop and how fast. It’s the first ML model to bring these predictions together across instability families, delivering a complete stability assessment up to 1,000,000× faster than our reference simulations. Fusion could provide abundant, low-carbon energy, but plasma instability is one of the biggest challenges for tokamaks: it can degrade confinement and rule out promising reactor designs. Cutting stability checks from hours to milliseconds could let engineers explore millions of plasma configurations and pursue more ambitious designs. We’re especially excited by the physics in these predictions: ModeCast reproduces magnetic islands and locates resonant layers. With just four new simulations per instability branch, it adapted its growth-rate predictions to a plasma outside the original training distribution. On unseen cases, it recovered the reference response to changing resistivity, with particularly close agreement for the tearing mode. The paper covers linear instabilities in the tokamak core and has been accepted at the Sim2Science workshop at NeurIPS 2026 in Paris. Credit to the team behind this work: @alexhigg_ , Zara Ercan, Abetharan Antony, Ph.D, Guzman Sanchez Gonzalez, Rafael Grazzini Placucci, Armand Kassaï Koupaïï and Artur Toshev. The full blog post, interactive figures and paper are linked in the comments. ModeCast is one part of our work in fusion at Zenithon. We’ll have much more to share very soon. #FusionEnergy #AIforScience #PlasmaPhysics #NeurIPS2026

Introducing NOVA by Zenithon, the first world model for extreme physics. NOVA predicts how complex physical systems evolve and respond to control inputs, in seconds on a single GPU. Built on a single transformer architecture, it is pretrained on diverse physical systems spanning plasma physics, fluid dynamics and more, using data from multiple simulation codes. Our first showcase is a problem close to our hearts: fusion energy. We demonstrate NOVA across magnetic, magneto-inertial and inertial fusion, adapting to new operating regimes with a few dozen simulations, or none in our tokamak example. For engineers, that means faster design exploration and experiment planning, with more possibilities tested before committing to hardware. Explore the technology and results: lnkd.in/e5mfnnKR We’re just getting started.

Our CEO @alexhigg_ recently spoke with SemiWiki.com about how world models for extreme physics can change semiconductor manufacturing. Speaking to Daniel Nenni, ( Founder of SemiWiki) he covers why trial and error can’t keep pace with the demands of manufacturing leading-edge nodes, and how we’re changing that. Full interview: lnkd.in/eh95i_2V

Last week we ran a workshop at Reuters Fusion Energy in Boston on how physics world models can transform the design, development and commercialisation of fusion. Our biggest takeaway from the two days was how urgent fusion now feels. Competition from China, the climate crisis, the push for energy sovereignty and the growing power demands of AI are all pointing the same way. That urgency puts pressure on the design process. Today's simulation tools only let teams explore a small part of the design space before they have to commit to hardware, so they often settle for designs that are good enough. If the industry wants power plants sooner, the design cycle has to get much faster, and that's what we're working on. We also gave the room a closed-door preview of something big we've been building. More on that next week. Thanks to @nuclearenergy1 for having us, and to everyone who came along and asked such sharp questions.

We’ve raised $10m to build the first world models for extreme physics and accelerate the technologies that will define the 21st century. Our backers include @BackedVC , @lunar_vc , @seraphim_space , @MMC_Ventures and @SOSV , as well as other funds, founders and hyperscaler directors. Alongside AI, this century will be shaped by rockets, fusion reactors and fabs. Building the next generation means understanding and controlling increasingly complex physics. Today’s tools cannot keep pace: when a single simulation or experiment takes days, even the best engineering teams can explore only a fraction of the possibilities. At Zenithon, we’re building models to explore a million design points in the time it takes to simulate one, and to learn directly from real-world experiments. Engineers can search beyond familiar designs, discover new levels of performance and keep improving how their machines operate. The team, co-founded by @alexhigg_ and , brings together researchers behind foundational breakthroughs in machine learning for physics and a wider team from leading labs globally. We’re proud to be building this together. Our ambition is to keep human innovation on an exponential path, even as the technologies we build become more complex. Linkedin linkedin.com/company/zenitho…