Denoise Pretraining on Nonequilibrium Molecules for Accurate and Transferable Neural Potentials

• DOI: 10.48550/arxiv.2303.02216
• Publication Date: 2023-01-01
• Authors: Yuyang Wang; Changwen Xu; Zijie Li; Amir Barati Farimani
• Keywords: Chemical Physics (physics.chem-ph); FOS: Computer and information sciences; Computer Science - Machine Learning; 0303 health sciences; 03 medical and health sciences; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Physics - Chemical Physics; FOS: Physical sciences; Machine Learning (cs.LG)

Recent advances in equivariant graph neural networks (GNNs) have made deep learning amenable to developing fast surrogate models expensive ab initio quantum mechanics (QM) approaches for molecular potential predictions. However, building accurate and transferable using GNNs remains challenging, as the data is greatly limited by computational costs level of theory QM methods, especially large complex systems. In this work, we propose denoise pretraining on nonequilibrium conformations achieve more GNN Specifically, atomic coordinates sampled are perturbed random noises pretrained which recovers original coordinates. Rigorous experiments multiple benchmarks reveal that significantly improves accuracy potentials. Furthermore, show proposed approach model-agnostic, it performance different invariant GNNs. Notably, our small molecules demonstrate remarkable transferability, improving when fine-tuned diverse systems, including elements, charged molecules, biomolecules, larger These results highlight leveraging build generalizable potentials