Microsoft Research shipped Skala-1.1, a major update to its deep-learning approach to density functional theory, the computational engine behind much of chemistry and materials science.
Trained on 2.5 times more data than the first public version, the model reaches a weighted average error of 2.8 kcal/mol on the GMTKN55 benchmark, a suite spanning thermochemistry, reaction barriers, and noncovalent interactions. Microsoft says that beats today’s leading hybrid functionals while keeping the speed of a semi-local functional, and that the model also delivers accurate electron densities, dipole moments, and molecular geometries.
Each Skala release is designed to supersede the previous one rather than accumulate, Microsoft says, with accuracy improving at the same computational cost. The update adds new training categories including electron affinities and noncovalent clusters.
Accessibility is expanding alongside accuracy. Skala is now available in CP2K and is being integrated into Psi4, FHI-aims, ORCA, and VASP, putting the functional inside the codes that research groups already run for simulations of drug discovery, catalysis, and energy materials.
Microsoft also introduced a living benchmark that tracks the computational performance of successive Skala releases across software packages and hardware, so the community can measure progress toward faster, more predictive simulation.
