This paper proposes a method for generating density functional theory (DFT) initial guesses by predicting electron densities in a compact auxiliary basis representation using an E(3)-equivariant neural network to accelerate the convergence of DFT calculations. Trained on small molecules with 20 atoms or fewer, the model achieves an average self-consistent field (SCF) step reduction of 33.3% for systems with up to 60 atoms, significantly outperforming existing Hamiltonian-centered and density matrix (DM)-centered models. Furthermore, it exhibits robust transport behavior across orbital basis sets and exchange-correlation (XC) functionals, suggesting it as a first candidate for a universally transportable DFT acceleration method. The researchers have made the SCFbench dataset and related code publicly available to support further research.
