General spin models from noncollinear spin density functional theory and spin-cluster expansion

Abstract: We present a data-efficient framework for constructing general classical spin Hamiltonians from the spin-cluster expansion (SCE) combined with fully self-consistent noncollinear spin density functional theory (DFT). The key idea is to fit an SCE model to magnetic torques rather than to total energies. Because torques are site-resolved vectors, each configuration supplies many independent constraints, which makes the regression well conditioned and sharply reduces the number of DFT calculations needed, especially in large supercells. Applied to the B20-type chiral magnets ${\rm Mn}{1-x}{\rm Fe}{x}{\rm Ge}$ and ${\rm Fe}{1-y}{\rm Co}{y}{\rm Ge}$, the resulting models nonperturbatively extract the full pairwise exchange tensor (isotropic exchange, anisotropic symmetric exchange, and the Dzyaloshinskii--Moriya interaction) and predict helical spin period via a micromagnetic mapping. The composition trends and the divergence of the period near the chirality sign change are reproduced in line with experiments. Because the SCE framework is systematic, it also enables systematic analysis of interaction order; training on increasingly disordered spin configurations shows that the lowest-order model loses torque accuracy, whereas including higher-order interactions restores predictive power. These advances enable near-DFT-accurate spin models for finite-temperature magnetism and complex textures at modest data cost, while providing a systematic, extensible, and nonperturbative route to quantitative first-principles parameterization and predictive materials design. An open-source implementation is available as the Julia package, \textit{Magesty.jl}.