Reproduce “Physics-Aligned Electronic Ground-State Learning Improves Generalization” 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
Eike S. Eberhard, Xaver Kainz, Viktor Kotsev, Abdulrahman Aldossary, Stephan Günnemann
The authors train density and Hamiltonian predictors with losses that reflect the physical states their matrices represent, while GROOT also constrains density predictions. On QM9-to-QM40 size extrapolation, ROCKET reduces energy and force MAE by 99.8% and 95.9% against prior Hamiltonian baselines under a one-Fock-build budget. Evaluations cover closed-shell B3LYP/def2-SVP only.
Key insights
- Respecting the physical representation of an electronic state can improve generalization without changing the shared equivariant network backbone.
- Constraining density predictions to valid projectors can make energy errors less sensitive to small density errors and avoids unphysical electron occupations.
- A self-consistency residual can serve as a practical reliability signal, while self-consistency-based fine-tuning can adapt models without reference energy labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.