Preprint reports lower GPU costs for molecular energy calculations

Researchers report a neural network that calculates molecular electronic states at previously untrained geometries without further optimization, with an estimated 25.8× reduction in end-to-end GPU costs on a 161-point nitrogen molecule grid. The results appear in an arXiv preprint. The model learns from sparse reference geometries. An orbital alignment procedure keeps orbital identities and phases consistent as molecular geometry changes. Across three paired nitrogen training seeds, alignment reduced mean absolute energy error at untrained geometries from 34-37 mHa to 0.049-0.085 mHa. At approximately 1 mHa mean absolute error, evaluation without further optimization reduced per-geometry cost by 986× relative to independent optimization. The reported cost comparison applies to the nitrogen grid.