Microsoft Research releases Skala-1.1 with 2.5x more data and broad integration
Microsoft Research launched Skala-1.1, a deep-learning exchange-correlation functional, trained on 2.5 times more data than the original model, reducing weighted average error on the GMTKN55 benchmark to 2.8 kcal/mol.
Source: Microsoft Research · August 20, 2026 at 5:31 PM · AI-assisted report
Single-sourceKUALA LUMPUR, 21 AUGUST 2026 —
Listen to this article
DomainFork Audio · read aloud
Microsoft Research launched Skala-1.1, a deep-learning exchange-correlation functional, trained on 2.5 times more data than the original model, reducing weighted average error on the GMTKN55 benchmark to 2.8 kcal/mol.
Market Impact
The upgrade cuts errors in main-group thermochemistry, reaction kinetics and molecular structure prediction, while maintaining semi-local functional efficiency, according to the research team.
To widen access, Skala is now available in CP2K and being integrated into Psi4, FHI-aims, ORCA and VASP, the developers said.
A living performance benchmark and transparent reference harness will track successive Skala releases across hardware and software platforms, they added.
Skala-1.1’s accuracy surpasses today’s leading global hybrid functionals on GMTKN55, the team said, attributing the gain to a major expansion of the Microsoft Research Accurate Chemistry Collection, which added electron affinities and noncovalent clusters.
Microsoft first released Skala through an open-source community edition built on GPU4 PySCF and integrated with ASE, enabling CPU and GPU use with minimal setup, the researchers said.
The CP2K integration, developed with Prof. Thomas D. Kühne’s team at CASUS, targets large-scale simulations and long-timescale molecular dynamics while preserving computational efficiency, they said.
Psi4 integration is under way, and collaborations with FHI-aims, ORCA and VASP aim to extend Skala across the major computational chemistry platforms, the team added.
Performance remains comparable to semi-local meta-GGAs on CPUs and GPUs, with overhead disappearing for molecules larger than 20–30 atoms, the researchers said, and ongoing optimisations will continue to improve speed and accuracy.
Related: Microsoft