Microsoft’s Skala 1.1 Broadens Predictive Chemistry
Microsoft’s updated neural chemistry functional improves accuracy and moves into software already used for molecular simulation.
A neural upgrade for a scientific workhorse
Microsoft Research has released Skala 1.1, an updated deep-learning exchange-correlation functional intended to improve the accuracy of density functional theory without imposing the high cost associated with conventional hybrid methods. DFT is widely used to model electronic structure in chemistry, materials science, catalysis, energy research and drug discovery.
The new version was trained on 2.5 times as much reference data as its predecessor, with added examples covering electron affinities and noncovalent molecular clusters. Microsoft reports a weighted average error of 2.8 kilocalories per mole across GMTKN55, a suite spanning 55 categories of thermochemistry, reaction barriers and noncovalent interactions. Skala 1.1 recorded the lowest error in 32 categories and also predicts electron densities, dipole moments and molecular geometries.
The company says those results exceed leading global hybrid functionals while retaining computational demands closer to a semi-local meta-GGA method. Its measurements put GPU cost roughly level with r2SCAN, while CPU overhead diminishes for larger molecular systems. These are research benchmarks rather than proof of accuracy across every industrial workload.
Deployment, not only benchmark gains
The accompanying access expansion is central to the announcement. Skala is now integrated into the open-source CP2K simulation package, with the CP2K and PySCF implementations agreeing closely in Microsoft’s validation. Work is also underway to add it to Psi4, FHI-aims, ORCA and VASP—tools embedded in established computational-chemistry and materials workflows.
Microsoft additionally released a benchmarking harness and a continuously updated performance report so package developers can compare implementations across hardware. The model and inference code remain available under an MIT license.
Why it matters: AI-for-science systems often remain disconnected from the software scientists actually use. Skala 1.1 combines a measurable accuracy improvement with integration into mature simulation environments. If independent users reproduce the reported balance of speed and precision, regularly updated learned functionals could replace today’s accumulation of static approximations and shorten computational screening cycles for molecules and materials.