Machine-learning force-field models for dynamical simulations of metallic magnets

Abstract: We review recent advances in ML force-field methods for Landau-Lifshitz-Gilbert (LLG) simulations of itinerant electron magnets, focusing on scalability and transferability. Built on the principle of locality, a deep neural network model is developed to efficiently and accurately predict the electron-mediated forces governing spin dynamics. Symmetry-aware descriptors constructed through a group-theoretical approach ensure rigorous incorporation of both lattice and spin-rotation symmetries. The framework is demonstrated using the prototypical s-d exchange model widely employed in spintronics. ML-enabled large-scale simulations reveal novel nonequilibrium phenomena, including anomalous coarsening of tetrahedral spin order on the triangular lattice and the freezing of phase separation dynamics in lightly hole-doped, strong-coupling square-lattice systems. These results establish ML force-field frameworks as scalable, accurate, and versatile tools for modeling nonequilibrium spin dynamics in itinerant magnets.

• The paper develops symmetry-preserving neural-network force fields for itinerant magnets that reduce LLG simulation costs from cubic to linear scaling while retaining quantum-informed exchange fields.
• The paper demonstrates nearly 1,000-fold speedups over exact diagonalization and about fivefold gains over GPU KPM, enabling large-scale simulations that reproduce benchmark spin dynamics and extrapolate beyond training times.
• The paper finds linear chiral-domain coarsening and late-time sublogarithmic phase-separation growth, linking these anomalies to corner-driven interfaces and correlation-induced hole self-trapping.

Overview

This review by Chern, Fan, Zhang, and Zhang presents a machine-learning (ML) force-field framework for large-scale Landau-Lifshitz-Gilbert (LLG) simulations of itinerant electron magnets (2602.18213). The central problem addressed is computational: each LLG time step requires evaluating exchange fields $\mathbf H_i = -\partial E/\partial \mathbf S_i$, where $E$ is obtained by integrating out itinerant electrons for a frozen spin configuration. Exact diagonalization scales as $\mathcal O(N^3)$, and even linear-scaling kernel polynomial methods (KPM) carry substantial overhead and cannot handle electron-electron interactions. The authors generalize the Behler-Parrinello (BP) architecture from ab initio molecular dynamics to spin dynamics, exploiting Kohn's locality ("nearsightedness") principle to achieve linear scaling while retaining quantum-mechanical accuracy.

The framework is demonstrated on the one-band s-d exchange model, with two applications that yield new nonequilibrium physics: anomalous linear coarsening of tetrahedral chiral domains on the triangular lattice, and correlation-induced freezing of phase separation in a lightly hole-doped double-exchange system.

Symmetry-preserving magnetic descriptors

The effective energy is decomposed as $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$, where $\mathcal C_i$ is the set of spins within a cutoff radius $r_c$ of site $i$. A neural network approximates the universal local function $\varepsilon$, and automatic differentiation supplies the exchange fields. The key methodological contribution is the descriptor construction. Spin-rotation invariance is guaranteed by expressing the local environment through bond variables $b_{jk} = \mathbf S_j \cdot \mathbf S_k$ and scalar chiralities $$\chi_{jmn} = \mathbf S_j \cdot \mathbf S_m \times \mathbf S_n$$. Lattice point-group symmetry is handled by decomposing these variables into irreducible representations (IRs) of the on-site group (e.g., $E$0 on the square lattice), then forming power-spectrum invariants $E$1. Because the power spectrum discards relative phases between IRs, the authors introduce reference IR coefficients from coarse-grained blocks to define phase variables $E$2, supplemented by a selected set of bispectrum coefficients. This group-theoretical construction ensures descriptors are invariant under both lattice symmetries and global SO(3)/SU(2) spin rotations while remaining differentiable with respect to spin orientations — a requirement specific to magnetic systems that distinguishes this work from standard atomic BP descriptors built on the Euclidean group $E$3.

Computational performance

The efficiency gains are substantial and quantified directly. A 10,000-step LLG simulation of a $E$4 lattice takes roughly 20 CPU hours with exact diagonalization but only about 5 minutes with the ML force field — a speedup of nearly three orders of magnitude, attributable to the change from $E$5 to $E$6 scaling. Even against the GPU-implemented KPM, which shares linear scaling, the ML model retains an approximately fivefold advantage (100 minutes versus 20 minutes for a 500-step run of a $E$7 system). These numbers establish ML force fields not merely as an approximation scheme but as the practical route to large-scale adiabatic spin dynamics.

Chiral domain coarsening: violation of Allen-Cahn growth

At filling fraction $E$8 on the triangular lattice, the s-d model realizes noncoplanar tetrahedral order describable as a triple-$E$9 state with ordering vectors at the midpoints of the hexagonal Brillouin-zone edge. This order carries a discrete $\mathcal O(N^3)$0 chirality whose spontaneous breaking yields a quantized anomalous Hall conductivity $\mathcal O(N^3)$1. Prior Monte Carlo work showed the finite-temperature chiral transition is strongly first order despite its Ising-like order parameter, motivating the question of whether coarsening deviates from the Allen-Cahn law $\mathcal O(N^3)$2.

A network with eight hidden layers (1806 input features, up to 2048 neurons per layer) was trained on KPM-LLG data with a loss combining torque mean-square error and total energy error; the torque-only model achieves an MSE of $\mathcal O(N^3)$3 per spin without overfitting. Notably, the authors report that including the energy term makes training harder and recommend ramping up its weight only in later epochs. Thermal-quench simulations show excellent agreement between ML- and KPM-driven dynamics for the chirality structure factor, and — importantly — the model trained only on data up to $\mathcal O(N^3)$4 continues to predict the coarsening law reliably at much longer times, demonstrating extrapolation beyond the training window.

The principal physical result is that the chiral domain size grows linearly in time, $\mathcal O(N^3)$5, rather than as $\mathcal O(N^3)$6. The authors attribute this to the domain morphology: late-stage interfaces are nearly straight and aligned with principal lattice directions, so curvature-driven Allen-Cahn motion vanishes and growth must be governed by corner dynamics treated as point defects. This observation implies that existing coarsening theory for non-conserved Ising-like order parameters is inadequate here and calls for a coarsening framework based on corner motion.

Phase separation: breakdown of Lifshitz-Slyozov-Wagner scaling

In the strong-coupling (double-exchange) regime of the square-lattice model at slight hole doping, the ground state is a mixed phase of hole-rich ferromagnetic puddles in a half-filled antiferromagnetic background — the canonical setting for colossal magnetoresistance manganites. Since hole number is conserved, classical theory predicts LSW coarsening with $\mathcal O(N^3)$7. Whether this holds for correlation-driven electronic phase separation was previously untested.

A six-layer network trained on 3500 ED-LLG snapshots from a $\mathcal O(N^3)$8 lattice predicts exchange fields with Gaussian-distributed errors of variance $\mathcal O(N^3)$9. The authors make an interesting interpretive claim: because the prediction errors are normally distributed, they may act as an effective Langevin temperature, which would explain why noise-free ML-LLG runs reproduce correlations matching ED-LLG simulations at $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$0. This equivalence is suggestive but asserted rather than rigorously established.

Large-scale quenches on a $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$1 lattice at 1.5% doping reveal that FM cluster growth follows $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$2 only at early times; at later stages growth slows dramatically, fitting a sublogarithmic form $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$3 with $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$4. The proposed mechanism is self-trapping: as antiferromagnetic correlations develop, each doped hole becomes surrounded by parallel spins via double exchange, suppressing hole evaporation-condensation between clusters and halting LSW coarsening. This result implies that nanoscale phase textures in CMR-like systems can be dynamically arrested by correlation effects alone, independent of disorder pinning — a distinction relevant to interpreting experimental inhomogeneity.

Limitations and open questions

Several caveats bear directly on the results. The locality assumption underlying the site-energy decomposition has a fixed cutoff radius $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$5; interactions such as RKKY oscillations at weak coupling are long-ranged, so transferability across coupling regimes is not demonstrated within a single trained model. Training is expensive (~5 days for the eight-layer triangular-lattice model), and the reported accuracy is specific to the training parameters ($$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$6, chemical potential, temperature range); no systematic study of extrapolation in parameter space is provided. The identification of ML prediction noise with an effective thermal bath remains a hypothesis consistent with, but not proven by, the benchmark data. Physically, the linear coarsening law lacks a quantitative theory — the corner-motion picture is qualitative — and the sublogarithmic fit exponent $$E = \sum_i \epsilon_i = \sum_i \varepsilon(\mathcal C_i)$$7 is empirical, leaving open whether the freezing is asymptotic or a crossover. Finally, the framework treats spins classically within the adiabatic approximation, excluding nonadiabatic spin-transfer torques and quantum spin fluctuations.

Conclusion

This paper establishes a rigorous, symmetry-aware extension of BP-type ML force fields to itinerant magnetism, delivering linear-scaling LLG simulations with near-exact accuracy and enabling system sizes and timescales inaccessible to diagonalization-based methods. Its two case studies — linear chiral coarsening violating Allen-Cahn scaling, and correlation-induced arrest of electronic phase separation — demonstrate that the approach is not merely an accelerant but a discovery tool for nonequilibrium spin dynamics. The natural extensions flagged by the authors, namely equivariant neural networks and graph neural networks with localized message passing, pose concrete open questions about how far symmetry-constrained architectures can push accuracy and transferability for correlated spin-electron dynamics.