AI Grounded in Physics Seeks Rare-Earth-Free Magnets

For more than two decades, the search for a magnet that could rival the strength of rare-earth materials without actually containing any rare earths has followed a familiar, slow rhythm: mix two elements, melt them together, measure what comes out, write down the result, and start again. It is a method that has produced incremental progress and the occasional surprise, but it advances one data point at a time. A new effort out of Ames National Laboratory argues that the rhythm itself is the bottleneck — and that an artificial intelligence grounded in physics, rather than trained only on the accumulated record of past experiments, can change the tempo entirely.

What Happened

In early June, Ames National Laboratory — a U.S. Department of Energy lab operated by Iowa State University — published an AI-driven roadmap for designing the next generation of rare-earth-free permanent magnets. The work, authored by Ames Lab scientist Prashant Singh and laid out in a paper in Advanced Functional Materials, was picked up across science and technology outlets through the middle of the month, drawing fresh attention to a quiet but consequential idea: that the most useful AI for materials discovery may be the kind that already understands the rules of the game before it starts playing.

The roadmap is part of the DOE's Genesis Mission, an initiative that pulls together the national labs, universities, and industry to point artificial intelligence at problems in energy, fundamental science, and national security — including the long-running challenge of securing a domestic supply of rare earths and other critical minerals. Rather than treating the magnet problem as a purely chemical guessing game, Singh's approach combines physics-based modeling, high-throughput simulation, and reasoning-capable AI tools to narrow the field of candidate materials before anything is ever synthesized in a furnace.

The distinction matters. A conventional machine-learning model trained on a catalog of known magnets is, in a sense, a very capable librarian: it can interpolate within what it has already seen, but it struggles to propose something genuinely outside that record. A physics-informed model is closer to a working scientist. Because it encodes how a material's atomic arrangement and electronic structure give rise to magnetic properties — magnetization strength, energy storage, resistance to demagnetization, behavior at high temperature — it can reason about combinations no one has tried yet, and search what Singh calls "arbitrary material space" rather than nudging at the edges of the familiar.

Why It Matters

Permanent magnets are one of those technologies that disappear into the background precisely because they are everywhere. They spin the rotors of electric-vehicle traction motors, hold the read-write heads in disk drives, drive the actuators in robots, and sit inside the generators of wind turbines. They are also deeply woven into defense systems — radar arrays, fighter aircraft, submarines, and uncrewed vehicles all depend on magnets that hold their field under heat, vibration, and time. The strongest of these magnets owe their performance to rare-earth elements such as neodymium and dysprosium.

The trouble is not that rare earths are geologically scarce — they are not especially rare — but that the supply chain for mining and, above all, refining them is concentrated and fragile. The economics are volatile, the processing is environmentally costly, and a great deal of it happens far from where the finished magnets are eventually used. For a country trying to build resilient domestic manufacturing in electronics, clean energy, transportation, and defense, a magnet that delivers comparable performance without rare earths is not a minor convenience. It is a strategic hedge.

This is where the AI angle becomes more than a research curiosity. The reason rare-earth-free magnets have remained stubbornly elusive is that the search space is enormous and the experiments are expensive. Every promising idea has historically demanded weeks of synthesis and characterization to confirm or reject. By embedding the physics that actually governs magnetic behavior into the computational models, Singh's framework aims to do the rejecting and the prioritizing up front — in simulation — so that the costly lab work is reserved for the candidates most likely to succeed. Crucially, the tools can also weigh practical constraints that pure science usually ignores: how much a material costs, how easy its ingredients are to source, and whether a "solution" simply trades one supply-chain vulnerability for another.

How the Field Is Reacting

The broader materials-science community has spent the past few years working through a genuine debate about how much to trust AI as a discovery engine, and the Ames work lands squarely inside it. The optimistic case has been building since large data-driven projects — Google DeepMind's GNoME and the autonomous "A-Lab" at Lawrence Berkeley among them — claimed to predict and synthesize large numbers of new inorganic compounds. The skeptical case followed close behind, pointing out that many AI-proposed materials turn out to be variations on known ones, or simply not stable enough to make.

Singh's argument is, in effect, a response to that skepticism. The fix for data-only models that overpromise, he suggests, is not more data but better grounding. "If you just use the data to train your models, you are going to get only the predictions within the range of information you have," he has said. "But once you understand the physics of what controls specific properties, then you and your agentic tools or AI frameworks can search arbitrary material space." It is a notably humble framing of what AI is for: not a replacement for scientific understanding, but an accelerant bolted onto it. Coverage of the announcement has reflected both the excitement and the caution, with some outlets celebrating a path to magnets free of rare-earth dependence and others reminding readers that a roadmap is not yet a magnet on a shelf.

Part of what gives the Ames claim weight is institutional rather than algorithmic. The lab points to roughly seven decades of work in critical materials and a body of proprietary magnetic-materials data that, it argues, no other institution can match. In a field where the quality of training data often determines the quality of the prediction, that accumulated, carefully measured record may be as much of a competitive advantage as any model architecture.

What Comes Next

The roadmap is meant to be built upon, and the most interesting near-term piece is how scientists will actually talk to these models. The team is developing AI-assisted interfaces that let a researcher pose a design question, refine the requirements, and explore candidate materials interactively, rather than running rigid batch simulations and waiting for output. That work draws on Singh's earlier tool, DuctGPT — an agentic AI originally built to help discover materials tough enough to survive inside fusion reactors, where components must endure extreme heat and radiation while remaining ductile enough to be formed into parts.

There is also a concrete precedent for the payoff. Ames researchers have already developed a rare-earth-free magnet based on manganese and bismuth (MnBi) — a material with the unusual property of becoming more resistant to demagnetization as it heats up, nearly doubling its room-temperature coercivity at 120°C, and one that recently earned an R&D 100 Award after being tested in an industrial pump motor. The AI roadmap is, in part, an attempt to find the next MnBi faster: to compress the years of trial and error that produced that magnet into a search that simulation can run in a fraction of the time.

If the approach scales, the implications reach well beyond defense procurement. More flexible magnet supply chains would ripple into electric vehicles, grid-scale wind power, industrial automation, and consumer electronics — anywhere a motor spins or a field needs to hold. Just as importantly, a physics-informed search engine for magnets is, in principle, a template for other classes of materials, from catalysts to superconductors, wherever the underlying physics is well enough understood to teach a model the rules.

Closing Thoughts

It is tempting to read every materials-AI headline as another step toward machines that simply invent the future for us. The Ames roadmap suggests something more interesting and more modest. The breakthrough here is not that an AI replaced the physicist; it is that the physics was put back into the AI. A model that knows why a magnet behaves the way it does can venture into unfamiliar territory with some confidence, while a model that has only memorized the past tends to stay close to home.

That framing carries a quiet lesson that extends past magnets. As AI tools spread into one scientific domain after another, the projects that endure may be the ones that treat decades of human expertise not as something to be automated away, but as the very thing that makes the automation trustworthy. The fastest way forward, in other words, might run straight through the slow, accumulated knowledge it appears to leave behind — and the most powerful tool in a materials lab may turn out to be an AI that has been taught, patiently, to understand the rules of the game.

한글 요약

미국 에너지부 산하 에임스 국립연구소(Ames National Laboratory)의 프라샨트 싱(Prashant Singh) 연구원이 6월 초 Advanced Functional Materials에 희토류를 쓰지 않는 차세대 영구자석 설계를 위한 'AI 기반 로드맵'을 발표했고, 6월 중순 여러 과학·기술 매체가 이를 비중 있게 다뤘습니다. 핵심은 기존 데이터만 학습한 AI가 아니라, 물질의 원자·전자 구조가 자성을 어떻게 결정하는지에 대한 '물리 지식'을 모델에 내장한 물리 기반(physics-informed) AI라는 점입니다. 이 방식은 실험실에서 재료를 만들기 전에 시뮬레이션 단계에서 유망 후보를 추려, 20년 넘게 이어진 시행착오 방식의 속도 한계를 넘어서려 합니다. 이 연구는 AI로 에너지·과학·안보 난제를 풀려는 DOE의 '제네시스 미션(Genesis Mission)'의 일부입니다.

영구자석은 전기차 모터, 풍력 발전기, 로봇, 그리고 레이더·전투기·잠수함 같은 국방 시스템까지 폭넓게 쓰이며, 최고 성능 자석은 네오디뮴 같은 희토류에 의존합니다. 문제는 희토류 자체의 희소성이 아니라 채굴·정제 공급망이 특정 지역에 집중돼 가격이 불안정하고 환경 비용이 크다는 점입니다. 싱 연구원의 프레임워크는 자성 성능뿐 아니라 재료 가격과 조달 가능성 같은 현실적 제약까지 함께 고려해, 또 다른 공급망 취약점을 만들지 않는 실용적 후보를 찾는 것을 목표로 합니다. 에임스는 70년에 걸친 임계 소재 연구와 독자적인 자성 재료 데이터를 강점으로 내세웁니다.

이 발표는 데이터 기반 AI 소재 탐색을 둘러싼 학계의 논쟁 한가운데 놓여 있습니다. 데이터만으로 학습한 모델은 이미 알려진 물질의 변형을 제안하는 데 그친다는 비판에 대해, 싱 연구원은 "물리를 이해하면 임의의 물질 공간을 탐색할 수 있다"고 답합니다. 다음 단계로 연구진은 과학자가 설계 질문을 던지고 후보를 대화형으로 탐색할 수 있는 AI 인터페이스를 개발 중이며, 이는 핵융합로 소재 탐색을 위해 만든 에이전트형 도구 'DuctGPT'에 뿌리를 둡니다. 이미 개발한 망간·비스무트(MnBi) 자석이 그 가능성을 보여줍니다. 결국 이 연구의 메시지는 AI가 물리학자를 대체한 것이 아니라, 물리학을 다시 AI 안에 넣었다는 데 있습니다.

참고: Ames National Laboratory · BGR · Advanced Functional Materials · DOE Genesis Mission