Reproduce “DeePAW: A universal machine learning model for orbital-free ab initio calculations” on OpenResearch

About this paper

Tianhao Su, Shunbo Hu, Yue Wu, Runhai Oyang, Xitao Wang, Musen Li, Jeffrey Reimers, Tong-Yi Zhang

DeePAW, a machine learning model, offers universal orbital-free *ab initio* calculations, predicting electron densities with a Normalized Mean Absolute Percentage Error (NMAPE) of 0.351% and formation energies. It achieves broad applicability across 88 elements and diverse crystal structures, significantly accelerating *ab initio* computations.

Key insights

• The E(3)-equivariant double MPNN architecture, integrated with atomic and electron density networks, effectively learns robust representations of chemical environments and spatial electron distributions.
• A PAW-inspired partitioning of electron density into smooth and residual components, handled by distinct MLPs and KANs, is critical for managing chemical diversity and maintaining high prediction fidelity.
• The model demonstrates strong zero-shot generalization capabilities across various crystal types and functionalities, indicating it has learned fundamental local chemical interactions rather than merely memorizing training data.

Summary by alphaXiv.