Skala 1.1 brings AI-powered density functional theory closer to real-world chemistry labs

Microsoft Research has released Skala 1.1, a machine-learning-based density functional theory (DFT) model that advances computational chemistry with greater accuracy and broader accessibility. Trained on 2.5 times more data than its predecessor, the tool now delivers measurably better predictions for molecular properties critical to drug discovery, materials science, and catalysis.

Better predictions across the board#

Skala 1.1 achieves a weighted average error of 2.8 kcal/mol on GMTKN55, a standard benchmark suite that tests performance across 55 categories of chemical problems—from thermochemistry to reaction barriers and noncovalent interactions. This accuracy level exceeds current hybrid functionals while maintaining the computational efficiency of simpler semi-local approaches. The model also accurately predicts electron densities, dipole moments, and molecular geometries.

The improvement stems from expanding the Microsoft Research Accurate Chemistry Collection (MSR-ACC), a dataset of high-precision quantum-chemistry reference data. For version 1.1, researchers added new categories, including electron affinities and noncovalent clusters, increasing both size and diversity. This data-driven strategy allows Skala to improve systematically with each iteration, moving closer to a genuinely predictive, scalable DFT framework.

Now integrated into standard tools#

To have scientific impact, advanced models must be available where researchers actually work. Skala is now accessible in CP2K, an open-source package used widely for large-scale molecular dynamics and electronic-structure work. Integration is underway with Psi4, FHI-aims, ORCA, and VASP—major platforms in computational chemistry and materials science. Earlier, Microsoft released Skala through a community edition built on PySCF, which works with the Atomic Simulation Environment (ASE).

The CP2K integration, developed with Prof. Thomas D. Kühne's team at the Center for Advanced Systems Understanding (CASUS), included comprehensive testing to ensure numerical consistency. Results confirm that the CP2K version reproduces the accuracy of the original implementation to within 0.04 kcal/mol.

Performance you can track#

Computational chemistry is not just about accuracy—speed matters. Skala achieves performance comparable to semi-local meta-GGAs on CPUs and GPUs, with minimal overhead for molecules larger than 20–30 atoms. Rather than offer a static benchmark, Microsoft is publishing a that updates as new optimizations become available and versions improve. The report covers various computational tasks and hardware platforms, helping developers validate and refine their own implementations.