Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.20181865 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866901613002620928 |
|---|---|
| author | Jia, Bi |
| author_facet | Jia, Bi |
| contents | <p>This code implement the method that focuses on: </p> <p>Equivariant neural-network interatomic potentials, such as Machine-Learned Atomic Cluster Expansion (MACE), construct expressive tensorial representations of local atomic geometry. However, all predictions ultimately pass through an invariant scalar pathway that collapses these features into energies and their derivatives. We identify this scalar interface as a largely overlooked yet tractable design axis governing the physical fidelity of short-range equivariant potentials. Focusing on the MACE framework, we demonstrate that uniform neighborhood aggregation and spectrally limited nonlinearities in the scalar pathway produce four quantifiable physical artifacts: anomalously softened short-range repulsion, red-shifted optical phonons, energy drift under microcanonical integration, and nearest-neighbor over-approach in liquid radial distribution functions. To address these systematically, we introduce two lightweight, symmetry-preserving modules acting exclusively on $\ell{=}0$ channels: Physics-Aware Neighborhood (PAN) pooling, which replaces uniform aggregation with coordination-sensitive amplitude modulation, and Physics-Guided Spectral (PGS) mixers, which augment pair-level representations with a real-valued Fourier–Bessel basis and introduce a tapered Exponential-of-Semicircle expansion at readout. Benchmarked across metallic Ag, covalent Si, a short-range ionic Li--F subset, and the MD17/rMD17 molecular datasets, PAN+PGS reduces energy, force, and stress errors by 22-28\% relative to baseline MACE at under 5\% additional computational cost, while measurably attenuating all four physical artifacts. Results are reproducible across five training seeds, robust under a stricter grouped split for the Li--F subset, and transfer directionally to Allegro and NequIP backbones. These findings identify scalar-pathway fidelity--alongside tensor<br>expressivity--as a principled and accessible dimension for improving the<br>physical accuracy of short-range equivariant interatomic potentials.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_20181865 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Scalar-pathway fidelity improves physical accuracy in short-range equivariant interatomic potentials Jia, Bi <p>This code implement the method that focuses on: </p> <p>Equivariant neural-network interatomic potentials, such as Machine-Learned Atomic Cluster Expansion (MACE), construct expressive tensorial representations of local atomic geometry. However, all predictions ultimately pass through an invariant scalar pathway that collapses these features into energies and their derivatives. We identify this scalar interface as a largely overlooked yet tractable design axis governing the physical fidelity of short-range equivariant potentials. Focusing on the MACE framework, we demonstrate that uniform neighborhood aggregation and spectrally limited nonlinearities in the scalar pathway produce four quantifiable physical artifacts: anomalously softened short-range repulsion, red-shifted optical phonons, energy drift under microcanonical integration, and nearest-neighbor over-approach in liquid radial distribution functions. To address these systematically, we introduce two lightweight, symmetry-preserving modules acting exclusively on $\ell{=}0$ channels: Physics-Aware Neighborhood (PAN) pooling, which replaces uniform aggregation with coordination-sensitive amplitude modulation, and Physics-Guided Spectral (PGS) mixers, which augment pair-level representations with a real-valued Fourier–Bessel basis and introduce a tapered Exponential-of-Semicircle expansion at readout. Benchmarked across metallic Ag, covalent Si, a short-range ionic Li--F subset, and the MD17/rMD17 molecular datasets, PAN+PGS reduces energy, force, and stress errors by 22-28\% relative to baseline MACE at under 5\% additional computational cost, while measurably attenuating all four physical artifacts. Results are reproducible across five training seeds, robust under a stricter grouped split for the Li--F subset, and transfer directionally to Allegro and NequIP backbones. These findings identify scalar-pathway fidelity--alongside tensor<br>expressivity--as a principled and accessible dimension for improving the<br>physical accuracy of short-range equivariant interatomic potentials.</p> |
| title | Scalar-pathway fidelity improves physical accuracy in short-range equivariant interatomic potentials |
| url | https://doi.org/10.5281/zenodo.20181865 |