Speakers
Description
Self-consistent field (SCF) calculations remain the computational bottleneck in quantum chemistry workflows. The initial density matrix guess significantly impacts convergence speed, with the Superposition of Atomic Densities (SAD) being the de facto standard. We present a machine learning approach that achieves comparable performance to SAD using a remarkably compact model predicting only diagonal blocks of density matrices.
Our transformer-based architecture employs 3.2 million parameters to predict atomic-block diagonal elements of density matrices for arbitrary molecular systems. The model combines rotary position-encoded attention with physics-informed multipole interactions and is trained on ~4.9 million quantum chemistry calculations spanning diverse molecular geometries (1-200 atoms) and electronic structure methods (HF, B3LYP, ωB97X-D). The architecture is molecule-independent and handles systems up to typical quantum chemistry scales (~300 atoms).
We benchmark performance on a ~900 molecule subset of the GMTKN55 database, filtered to contain only elements H-Ar (periods 1-3). On average, SAD achieves 0.22 fewer SCF iterations than our model, demonstrating that the ML approach reaches near-parity with the established standard despite predicting only diagonal blocks. Notably, our model exhibits qualitatively different physical behavior: while SAD systematically underestimates electronic energy, our predictions overshoot the converged density, suggesting a more physically motivated initial electronic structure.
These results demonstrate that a remarkably small neural network can match decades-optimized classical methods for SCF initialization, opening pathways for more sophisticated architectures to exceed SAD performance while maintaining computational efficiency for routine quantum chemistry applications.