Physics-Informed AI Charts a New Path Toward Rare-Earth-Free Permanent Magnets

A team at a U.S. Department of Energy national laboratory has laid out a systematic, AI-guided roadmap for discovering permanent magnet materials that do not depend on rare-earth elements — a long-standing goal in materials science that has resisted brute-force experimentation for more than two decades.

Key takeaways

• Researchers at Ames National Laboratory have combined physics-based modeling with a reasoning-oriented AI system to screen candidate magnet materials before they are ever synthesized in a lab.
• The approach builds on an existing physics-trained model, originally developed to evaluate ductility in advanced alloys for fusion reactors, now being adapted toward magnetic materials design.
• The work is part of the Department of Energy’s Genesis Mission, aimed at strengthening domestic supply chains for critical minerals.
• Independent commentators caution that a true rare-earth-free replacement for today’s strongest magnets is still not on the table — this is a methodology breakthrough, not a finished product.

Why Rare-Earth Magnets Are a Strategic Bottleneck

Permanent magnets sit quietly at the center of modern life. They spin the motors in electric vehicles, stabilize hard drives, guide precision instruments in medical imaging machines, and anchor a long list of defense and energy technologies. The strongest commercial magnets in use today rely on neodymium, samarium, and other rare-earth elements, prized for the way they resist demagnetization even under intense heat and mechanical stress.

The trouble is supply. Refining rare-earth elements is concentrated overwhelmingly in a small number of countries, and building a resilient, domestic alternative supply chain has proven slow and expensive. For more than twenty years, materials scientists have searched for magnet chemistries that skip rare earths entirely while still holding onto comparable magnetic strength. Progress has historically been incremental, largely because the search space of possible atomic combinations is enormous, and testing each candidate physically in a lab is slow and costly.

Teaching an AI Model the Physics, Not Just the Data

This is where the new work out of Ames National Laboratory, in Iowa, stands apart from typical machine-learning approaches to materials discovery. Most AI materials models are trained purely on existing experimental datasets, which means they are fundamentally limited to interpolating between things that have already been measured. If a promising material sits far outside that training distribution, a conventional model has little basis to recognize it.

The Ames team instead built their AI workflow around a physics-informed foundation — a system originally developed under the name DuctGPT, which was designed to predict ductility in refractory, multi-element alloys intended for extreme environments such as fusion reactor components and aerospace turbines. Rather than only pattern-matching against historical alloy data, the model was trained to reason about how atomic structure and electronic behavior give rise to physical properties, allowing it to extrapolate into genuinely unexplored chemical territory.

Applied to the magnet problem, the extended workflow evaluates how combinations of elements influence magnetization strength, energy storage capacity, resistance to demagnetization, and thermal stability — the four properties that, together, determine whether a candidate material could ever function as a viable permanent magnet. Crucially, the researchers also folded supply-chain and manufacturing practicality directly into the model’s evaluation criteria, so that a chemically promising candidate that would be nearly impossible to source or produce at scale gets flagged early, rather than after years of downstream investment.

Screening Before the Lab, Not Instead of It

The practical value of this approach is speed. Traditional materials discovery follows a slow loop: propose a candidate, synthesize it, test it, measure the gap between prediction and reality, and repeat. By front-loading physics-grounded computational screening, the Ames workflow allows researchers to rule out large swaths of unpromising chemistries computationally, reserving expensive laboratory synthesis for the candidates most likely to succeed.

According to the researchers involved, Ames Laboratory’s decades of accumulated experimental data on magnetic materials gave the project a foundation that few other institutions could match — a long institutional memory of which element combinations behave in which ways, now encoded into a system that can reason about combinations no one has actually synthesized yet.

The team is also developing more interactive tooling on top of the core model, aimed at letting researchers query the system directly, adjust design constraints on the fly, and iterate on candidate shortlists in something closer to real time, rather than waiting on batch computational runs.

A Note of Caution: Roadmap, Not Replacement

It is worth being precise about what this milestone actually represents, because early press coverage of the Ames announcement occasionally overstated the result. This is not the unveiling of a finished rare-earth-free magnet ready to compete with today’s neodymium-based products. It is a methodology — a validated, physics-grounded computational pipeline that makes the search for such a material dramatically more efficient than brute-force experimentation.

Independent materials analysts have pointed out that commercially viable rare-earth-free substitutes remain in an early stage of technological maturity, and that established supply chains for rare-earth magnet manufacturing are not going away in the near term. The Ames roadmap should be read as a serious tightening of the search process — not as an announcement that the search is over.

Part of a Bigger Federal Push

The project sits inside the Department of Energy’s Genesis Mission, an initiative intended to bring together national laboratories, industry partners, and academic researchers around the shared goal of applying AI to hard problems in energy, discovery science, and national security — with critical mineral security named explicitly as a priority area. Reducing reliance on imported rare earths carries implications well beyond defense applications, touching renewable energy hardware, electric vehicle production, and consumer electronics manufacturing, all of which depend on stable, affordable magnet supply chains.

If the underlying physics-informed approach continues to generalize, the Ames team’s broader ambition is to turn this into a repeatable template — a way of building AI systems for materials discovery generally, not just for magnets. Given that the same underlying model was originally developed for an entirely different problem, in fusion-relevant alloys, that generalization ambition does not look far-fetched.

What Comes Next

The immediate next step for the team is expanding the pool of computationally screened candidates and beginning targeted laboratory synthesis of the most promising chemistries identified so far. Given the scale of the search space, researchers expect this to be a multi-year effort rather than a rapid turnaround, but the compressed screening timeline made possible by physics-informed AI could meaningfully shorten what has historically been a decades-long search.

For an industry watching costs, supply security, and geopolitical exposure all rise at once, even a modest acceleration in the search for alternative magnet chemistries carries outsized strategic weight. Whether or not a commercial rare-earth-free magnet emerges from this particular pipeline, the underlying lesson — that grounding AI models in physics, rather than only historical data, unlocks genuine extrapolation into unexplored materials — is likely to shape materials science efforts well beyond this one project.

How This Differs From Earlier AI Materials Screening Efforts

Materials scientists have used machine learning to accelerate discovery for well over a decade, but most of that earlier work fell into one of two camps. The first camp trained models purely on databases of previously measured materials, which works reasonably well for finding incremental improvements on known chemistries but struggles badly whenever the ideal answer lies far outside the training data. The second camp used brute-force computational simulation, running expensive first-principles physics calculations across huge numbers of candidate structures, which is thorough but often too slow and too costly to scale to the full space of plausible element combinations.

The Ames approach is explicitly designed to sit between these two camps, borrowing the extrapolation ability that comes from grounding predictions in real physics, while borrowing the speed and scalability that comes from a trained AI system rather than running full simulations on every single candidate. That hybrid positioning is precisely why researchers describe it as a roadmap rather than a single discovery: it is meant to be a repeatable pipeline that can be pointed at other materials problems entirely, not a one-time result specific to magnets.

Frequently Asked Questions

Has a working rare-earth-free magnet actually been built yet?
No. The Ames team has published a validated computational methodology for screening candidate materials far more efficiently than before. Laboratory synthesis and testing of the most promising candidates identified through this pipeline is the next phase of work, and researchers caution it will likely take years before a commercially competitive material emerges, if one does at all.

Why does this matter if no finished product exists yet?
Because the search space for viable magnet chemistries is enormous, and the previous bottleneck was largely how quickly researchers could rule out unpromising candidates before committing to expensive lab synthesis. A validated computational screening pipeline compresses that bottleneck substantially, even before any single winning material has been identified.

Is this specific to magnets, or could it apply elsewhere?
The core AI system was originally built for an unrelated materials problem — predicting ductility in alloys destined for fusion reactors — and has now been extended to magnetic materials. That track record suggests the underlying physics-informed approach could plausibly be adapted to other classes of materials discovery problems as well, which is part of why the Department of Energy is backing the work as a broader methodology investment rather than a narrow, single-purpose project.