Skip to main content
Back to timeline
arXivSource publication:

WannDiffuse hands polarization to a diffusion model: trained on small cells, it computes infrared spectra for 512-molecule water and 512-atom amorphous silicon

Synopsis

Electronic polarization in insulators and its derivatives have required frame-by-frame first-principles calculations, which large systems and long trajectories cannot afford. WannDiffuse combines an E(3)-equivariant graph neural network with a diffusion process that generates Wannier charge centres as an unordered set and predicts their real-space spreads. On 100 unseen crystals the root-mean-square errors are 0.09 Šfor centres and 0.02 Ų for spreads; trained on 32-molecule water, it handles 512-molecule water and reproduces the infrared spectrum.

Source-provided article image: A Universal Diffusion Model for the Quantum Position Operator
arXiv

Interpretation

Polarization and its derivatives with respect to atomic displacement and strain can now come directly from a generative model given an atomic configuration, without a self-consistent first-principles calculation at every frame. Earlier routes ran density-functional perturbation theory or performed Wannierization, both solved frame by frame; existing machine-learning approaches predicted only atom-centred Wannier centroids or relied on assigning Wannier centres to specific atoms in advance. Across 100 unseen insulators, the root-mean-square errors are 0.09 Šfor Wannier centres and 0.02 Ų for spreads; every chemical species in these materials appears in training, but no crystal structure overlaps the training set.

The model treats Wannier charge centres as an extra atomic species in the graph and uses a diffusion process to iteratively denoise them from random starting positions toward the true centres. Diffusion generation handles the centres as an unordered set, sidestepping the ordering or assignment that regression requires, so it can accommodate interstitial centres between atoms. Separate models are trained for liquid water, amorphous silicon and BaTiO3 on 800, 120 and 50 structures respectively; the universal model is trained on 2,807 crystalline insulators spanning 67 chemical species.

The predicted centres support finite-difference derivatives, yielding Born effective charges, spontaneous polarization and piezoelectric tensors. These response quantities otherwise require density-functional perturbation theory or frame-by-frame Wannierization. For tetragonal BaTiO3, Born effective charges deviate by roughly a few percent for all atomic species relative to established first-principles benchmarks, and the anomalously large Ti and O charges are recovered.

Perspective

The demonstrated scope is insulating systems: liquid water, amorphous silicon, cubic and tetragonal BaTiO3, 100 unseen crystals and monolayer MoS2; the universal model requires a fixed number of Wannier centres per chemical species, and training data come from valence Wannier functions. The authors propose training WannDiffuse alongside a machine-learning interatomic potential in the same active-learning workflow, with the extra cost largely confined to the Wannierization step.

The amorphous-silicon infrared spectrum shows a small systematic redshift relative to experiment, which the authors attribute to the underlying atomic trajectory rather than the Wannier-centre prediction; the small peak below 1 Å in the Si–WCC radial distribution matches Wannier90 on the same trajectory but deviates from earlier literature, attributed to different trajectories and DFT settings. The universal model's accuracy depends on related chemical species and structural motifs appearing in training.