Learning ab initio phase-field models

The theoretical derivation from the exact generalized Langevin equation to the practical mesoscopic model is clearly laid out, and the controlled comparison on the Lennard-Jones mixture—where all models use identical data, loss, and mobility but differ only in the free-energy ansatz—fairly isolates the benefit of the richer rung-3 functional. The physical applications are compelling: the iron–boron model predicts that the FeB$_4$ melt is spinodally unstable at ambient pressure but stabilized at 10 GPa, and the hydrogen–helium simulation reaches 2.2 million atoms, far beyond atomistic reach.

The three governing approximations invoked to obtain a tractable model are strong and remain largely unverified. The Markovian assumption requires the memory kernel to decay much faster than the density-field evolution, yet no comparison of $\tau_{\mathrm{mem}}$ and $\tau_{\mathrm{field}}$ is provided for Fe–B or H–He, and the paper itself notes that this assumption may fail in supercooled liquids or glasses. Similarly, mobility is restricted to depend only on local densities because fully nonlocal mobility is difficult to identify from data; this discards possible long-range kinetic couplings that could matter in dense metallic melts or plasmas. Finally, replacing the exact fluctuating force by Gaussian white noise fixes the covariance via the fluctuation–dissipation relation but does not test whether higher-order statistics of the unresolved forces are indeed negligible.

A second limitation is the omission of momentum density, which means the mesoscopic model cannot capture hydrodynamic flow. In the helium-rain demonstration, helium accumulates downward by diffusion under an enhanced gravitational potential rather than by droplet sedimentation, so the dynamics, while scalable, are not fully fluid-mechanical. More generally, the learned thermodynamic boundaries carry no uncertainty estimates: the binodal and spinodal surfaces are reported as single curves although they depend on the choice of parametrization, neural-network training, and the underlying machine-learning interatomic potential, whose own errors propagate unchecked into the phase-field model.

The evidence supports the key claims for the Lennard-Jones benchmark, where the learned rung-3 functional outperforms Landau and Flory–Huggins baselines for both the phase diagram and coarsening exponent. For the real materials, the Fe–B spinodal classification aligns with prior molecular-dynamics results, and the H–He immiscibility boundary is broadly consistent with published binodals, though the comparison in Figure 5d shows non-negligible scatter across literature sources that the present work does not resolve. The authors fairly contrast their approach with neural classical density functional theory, emphasizing that they learn dynamics jointly with equilibrium thermodynamics from short trajectories without external potentials, making the method more data efficient.

The code and trained models are publicly hosted on GitHub, which supports independent testing, but the molecular-dynamics trajectories and training datasets are available only upon request owing to their size. Critical hyperparameters—including the loss weights $\lambda_M$, $\lambda_S$, $\lambda_{\mathrm{bulk}}$, and $\lambda_P$ and regularization schedules—are relegated to Supplemental Material Section S3, so reproducing the training exactly requires consulting that source. Inference uses a documented pseudospectral solver, and the implementation details are given in Supplemental Material Section S2 A. Because the entire pipeline depends on prior machine-learning interatomic potentials (MACE) trained on density-functional-theory data, any reproduction is ultimately bottlenecked by the accuracy and transferability of those potentials.