New physics-aligned methods boost MLIP generalization

Researchers have developed new methods, OrthoNormal-Loss (ON-Loss) and Grassmann Restricted Occupied-Orbital Training (GROOT), to improve the generalization capabilities of machine-learned interatomic potentials (MLIPs). These techniques align learning objectives with Kohn-Sham density functional theory (KS-DFT) by enforcing physical constraints. In experiments, these methods significantly reduced mean absolute errors for energy and forces, outperforming previous state-of-the-art density GSMs and demonstrating effectiveness in reactive chemistry tasks. AI

IMPACT Improves generalization of MLIPs for drug and material development.

RANK_REASON Research paper detailing new methods for machine-learned interatomic potentials. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Grassmann Restricted Occupied-Orbital Training
• OrthoNormal-Loss
• QM40
• QM9
• QMugs

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →