# An MIT AI built its own physics simulator and used it to redesign graphene > Source: > Published: 2026-09-30 03:26:06+00:00 *An MIT research group let an AI system design and build its own atomistic simulation tools, then ran it for days on graphene metamaterials with no one steering the wheel.* Feed a large language model five pictures, mostly biological patterns like leaf veins and nested fiber networks, and a short prompt, and it will usually describe what it sees. Markus Buehler's lab at MIT went further. According to Buehler's own account of the project, his team gave an AI system that prompt and those five reference images and asked it to infer a design language from them, then build the actual scientific instrument needed to test it. Within hours, the system had generated a working simulation toolkit: a geometry generator, a fracture solver, a validation module, a parameter scanner, and three browser-based virtual labs, all written by the AI itself, not assembled from an existing library. Then it kept going. Buehler describes an initial reasoning phase that ran for several days, covering hundreds of discovery simulations across a wide set of graphene architecture families. The AI formed hypotheses about how these atomically thin lattices would deform and fracture under pressure, tested them by running its own simulator, checked the results against the physics, and revised the hypotheses that failed. Nobody was choosing which structure to try next. That is the part that separates this from the AI-solves-a-math-proof stories that have circulated all year. This isn't language or symbolic reasoning. It's an AI system doing the unglamorous, iterative work of computational materials science: building its own physical model, atom by atom, using it to generate predictions, and confronting those predictions with the resulting evidence, on a loop, over days, without a human picking the next experiment. Buehler's account describes the system exploring dozens of graphene design families with relative densities between 0.77 and 0.83, essentially the same mass of material arranged differently at the atomic scale. Across those families, the density-normalized strength varied by a factor of 6.6, meaning some lattice geometries held together under stress dramatically better than others despite weighing the same. That's a real, checkable spread, not a marketing number, and it's the kind of result that would take a human researcher running one simulation at a time months to map out by hand. [Huawei rewrites chip physics to bypass US sanctions](https://startupfortune.com/huawei-rewrites-chip-physics-to-bypass-us-sanctions/) Huawei's new Tau Scaling Law could enable 1.4nm-equivalent chips by 2031, sidestepping US sanctions and accelerating China's AI computing power, analysts say. - [how to bypass US chip manufacturing sanctions](https://startupfortune.com/huawei-rewrites-chip-physics-to-bypass-us-sanctions/) - [Huawei Tau Scaling Law transistor density breakthrough](https://startupfortune.com/huawei-rewrites-chip-physics-to-bypass-us-sanctions/) Whether any specific design from this run has been fabricated and tested in a physical lab is a separate question, and it matters. Computational strength gains in a simulator are not the same as a verified sample under a load cell. Buehler's own earlier work at MIT already showed that porous 3D graphene structures can be roughly ten times stronger than steel at a fraction of the density, and that finding did eventually get built and confirmed. The current graphene metamaterial discovery work, by contrast, is described as reasoning and simulation, first. That distinction should matter to any materials engineer reading the claim, and it's worth being honest about it rather than treating a simulated result as a finished material. This isn't an isolated stunt, either. Researchers at Argonne National Laboratory, publishing in the journal Digital Discovery, built a separate multi-agent AI framework that automates atomistic simulations from start to finish, with one administrator agent assigning tasks to specialist agents. Aditya Koneru, an Argonne Scholar at the Argonne Leadership Computing Facility and one of the study's authors, said the goal is straightforward: by automating these exhaustive investigations, the team can potentially reduce the time needed to discover new materials from months or years to just days. **Also read:** [OpenAI and Anthropic Both Just Launched Cheaper Flagship AI Models](https://startupfortune.com/openai-and-anthropic-both-just-launched-cheaper-flagship-ai-models/) • [Anthropic warns a free Chinese AI model can already build working hacks](https://startupfortune.com/anthropic-warns-a-free-chinese-ai-model-can-already-build-working-hacks/) • [OpenAI launches Dots to rival Meta's Muse and it stumbles on stage](https://startupfortune.com/openai-launches-dots-to-rival-metas-muse-and-it-stumbles-on-stage/) *This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.* ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away.