Reproduce “Equivariant Flow Matching for Electron Density Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenxing Liang, Chengdong Wang, Yuchao Lin, Xiaofeng Qian, Shuiwang Ji
OrbFlow predicts molecular electron densities by transforming structured noise into coefficients for compact Gaussian-type basis functions, using a two-step flow with rollout supervision. On QM9, it achieves 0.152% NMAE, 13.6% below ELECTRA. When used to initialize SCF, it reduces PBE0 test iterations by 68.13% and wall-clock time by 60.98% versus atomic-density initialization; finer untrained integration does not improve accuracy.
Key insights
- Training through the same short rollout used at inference helps address accumulated trajectory error; accuracy is best at the trained two-step horizon and declines when more steps are used.
- The compact GTO representation yields a measurable fitting floor: on QM9, the reported projection-and-reconstruction floor is 0.061% NMAE, below the model’s 0.152% prediction error.
- Predicted densities can be useful beyond density reconstruction: they improve SCF starting points and provide multipole estimates without requiring SCF, though evidence is limited to the tested molecules and functionals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.