Microsoft Releases Skala 1.1 for Chemistry Simulations

Microsoft Research has launched Skala 1.1, the latest version of its deep-learning exchange-correlation functional designed to accelerate density functional theory (DFT) calculations. Trained on 2.5 times more data than its predecessor from the Microsoft Research Accurate Chemistry Collection (MSR-ACC), the updated model offers significantly improved accuracy for molecular structure prediction, reaction kinetics, and main-group thermochemistry.

In testing, Skala 1.1 achieved a weighted average error of 2.8 kcal/mol on the GMTKN55 benchmark, which spans 55 categories of chemical problems. The model ranked first in 32 of these 55 categories, outperforming expensive global hybrid functionals like B3LYP and M06-2X while maintaining the computational efficiency of a semi-local meta-GGA functional like r2SCAN. On graphics processing units (GPUs), Skala 1.1 matches the cost of r2SCAN, while traditional hybrid functionals become more expensive for molecular systems with more than 1,000 orbitals. On central processing units (CPUs), Skala's initial overhead disappears for systems with more than 300 orbitals or roughly 20 to 30 atoms.

To make these capabilities useful to practitioners, Microsoft is expanding Skala's software ecosystem. The model is now natively available in the open-source CP2K package, integrated via GauXC in collaboration with the Center for Advanced Systems Understanding (CASUS). Validation tests show that the CP2K and PySCF implementations of Skala 1.1 agree within 0.1 kcal/mol mean absolute deviation (MAD). Microsoft is also actively integrating the model into Psi4, FHI-aims, ORCA, and VASP, building on its existing availability in PySCF and the Atomic Simulation Environment (ASE).

To help researchers track efficiency gains, Microsoft is introducing a living benchmark and a benchmarking harness. This allows developers to measure computational performance across different hardware platforms and software packages. For practitioners in drug discovery, materials science, and catalysis, these updates mean they can now run highly accurate, predictive quantum-chemistry simulations at a fraction of the traditional computational cost directly within their existing workflows.