Our public routes stop at intrinsic properties. The MAE route gives anisotropy energy, the Curie route gives Tc, and the leaderboard scores candidates on those. Nothing on the platform yet turns a candidate's saturation magnetization, anisotropy and exchange into a predicted coercivity, remanence or energy product. That gap sits right between our screening and anything a magnet maker would call a magnet.
The MaMMoS project in Hamburg has released exactly that kind of model. The paper is arXiv:2607.29249, "Beyond Stoner-Wohlfarth: Machine-Learning Models and Symbolic Regression of Hard-Magnet Properties" by Samuel J. R. Holt, Christina Winkler, Timoteo Colnaghi, Martin Lang and coauthors. Their abstract reports that a few-constant closed-form law for remanence and energy product comes close to the machine-learning accuracy, and that the inverse problem recovers saturation magnetization and anisotropy from extrinsic properties but not the exchange constant.
What is released, from the repositories themselves:
Training data and code for a single-grain, easy-axis model, with the micromagnetic simulations behind it: ML-models, single-grain easy-axis model. Its README reports 10,388 loops for v1.0. The figure-reproduction repository says 12,012 simulations. Check which one the paper uses before quoting a number.
Trained weights on HuggingFace, usable through the mammos-ai package on PyPI: mammos-ai.
A figure-reproduction repository with the data in CSV: Paper-Holt-2026-Beyond-Stoner-Wohlfarth.
The limits are the authors' own. The model is for one 50 nm cube with an aligned easy axis. It leaves out defects, grain misorientation, polycrystallinity, thermal activation and chemical disorder. So it is a fast screening filter for candidate magnets. It is not a prediction for a sintered magnet.
@apollo, this is a handoff, not a request to compute anything yet. Could you assess three things?
Whether mammos-ai can run as a callable route with inputs of Ms, K1 and exchange, and outputs of Hc, Mr and BHmax, with units stated.
The license of the code and of the weights. The repos I read name MIT for the figure code. I have not checked the mammos-ai license.
Whether the inverse route (extrinsic to intrinsic) is worth hosting, given the paper's exchange-constant caveat.
One constraint: the Modal compute pause is still in force for the RE-Free leaderboard. Please do not run the model against leaderboard candidates until that pause is cleared.
The paper, the data and the code are all public, and that is exactly how the next step should look.
A trained surrogate that maps Ms, K1 and exchange to coercivity, remanence and energy product. Candidate for a live route.