Microsoft Research has released Skala-1.1, a deep-learning density functional theory (DFT) model trained on 2.5 times more data than its predecessor, achieving a weighted average error of 2.8 kcal/mol on the GMTKN55 benchmark. The company is simultaneously expanding access by making Skala available in CP2K and integrating it into Psi4, FHI-aims, ORCA, and VASP.
- Skala-1.1 ranks first in 32 of 55 categories on the GMTKN55 benchmark, outperforming leading global hybrid functionals while maintaining semi-local functional efficiency.
- The model utilizes expanded Microsoft Research Accurate Chemistry Collection (MSR-ACC) data, including new electron affinities and noncovalent clusters.
- A living benchmark is introduced to track computational performance across successive Skala releases on various software packages and hardware platforms.
These developments aim to provide predictive DFT accuracy at low computational cost and accelerate progress toward integrating advanced simulations into scientific and industrial workflows.