[GS_C_MS] Attention makes machine-learned electron densities accurate and transferable enough to bypass self-consistency
ABSTRACT
We introduce an attention-based model that predicts the electron density on a three-dimensional real-space grid. Attention operates on coarsened grids, so added capacity is inexpensive — and unlike convolution it keeps paying off, which is what makes the model more accurate than previously reported approaches. The advantage widens out of distribution, where it transfers without fine-tuning to larger systems and to elements absent from training. The density is accurate enough to bypass self-consistency: a single non-self-consistent Kohn–Sham step, in place of an entire self-consistent cycle, gives total energies within chemical accuracy and geometries that recover equilibrium bond lengths and vibrational frequencies — opening a route to optimization and molecular dynamics.
