Microsoft Skala 1.1 Boosts DFT Accuracy and Expands Into Major Chemistry Platforms

Summary

Microsoft Research announced Skala 1.1 on August 20, 2026, advancing its deep-learning exchange-correlation functional with 2.5 times more training data than the first public release. It records a 2.8 kcal/mol weighted average error on GMTKN55, ranks first in 32 of its 55 chemistry categories, and outperforms leading global and range-separated hybrid functionals at semi-local meta-GGA cost. Improvements cover main-group thermochemistry, reaction kinetics, noncovalent interactions, molecular structures, electron densities, dipole moments and geometries. Microsoft expanded its Accurate Chemistry Collection, MSR-ACC, with high-accuracy wavefunction data including electron affinities and noncovalent clusters.

Skala 1.1 is available in the open-source CP2K package after integration with Professor Thomas D. Kühne’s Center for Advanced Systems Understanding, CASUS, using GauXC. CP2K and PySCF results agree within 0.1 kcal/mol mean absolute deviation on a representative GMTKN55 subset, apart from one difficult radical. Psi4 integration is underway, which would join CP2K and the PySCF-based Skala Community Edition as three open-source options; Microsoft is also working with FHI-aims, ORCA and VASP developers. The Community Edition supports ASE and optimized CPU and GPU execution. Skala matches r2SCAN cost on GPUs, while B3LYP and M06-2X become costlier above roughly 1,000 orbitals; its CPU overhead disappears beyond 20 to 30 atoms, or roughly 300 orbitals. A public benchmarking harness and living performance report will track implementations, hardware and optimizations across future releases.