DM-PhiSNet: Equivariant Model Accelerates SCF Workflows by 49–81%
A new equivariant neural network, DM-PhiSNet, predicts one-electron reduced density matrices (1-RDMs) from molecular geometries to accelerate self-consistent field (SCF) calculations. The model uses a two-stage training schedule with physically motivated objectives, followed by a lightweight analytic refinement block that enforces electron-number conservation and idempotency. Tested on six closed-shell systems (H₂O, CH₄, NH₃, HF, ethanol, NO₃⁻), the refined 1-RDMs reduce SCF iteration steps by 49–81% compared to standard initializations. The approach also enables accurate one-shot property predictions without full SCF convergence.
Key facts
- DM-PhiSNet is a PhiSNet-based equivariant model for predicting 1-RDMs.
- Training uses a two-stage schedule with physically motivated objectives.
- An analytic refinement block enforces electron-number conservation and generalized idempotency.
- Tested on six closed-shell systems: H₂O, CH₄, NH₃, HF, ethanol, NO₃⁻.
- Refined 1-RDMs reduce SCF iteration steps by 49–81%.
- Enables accurate one-shot property predictions without full SCF convergence.
- Published on arXiv with ID 2604.27256.
- The model operates in an atomic-orbital (AO) basis.
Entities
Institutions
- arXiv