Microsoft Research's Skala 1.1 Beats Top Chemistry AI at Half the Cost

Microsoft Research has shipped Skala 1.1, an updated version of its machine-learned exchange-correlation functional for density functional theory (DFT). If you have never touched a quantum chemistry code, the short version is this: DFT is the workhorse method scientists use to predict how molecules and materials behave, and the exchange-correlation functional is the approximation at its core that trades off accuracy against compute cost. Skala replaces that hand-crafted approximation with a neural network trained on high-accuracy reference data.

The 1.1 release is less about a single new capability and more about proving Skala can be improved on a schedule, like any other ML model. Each release is designed to supersede the previous one, and as new data, model architectures, and training strategies become available, the model improves while maintaining the same practical computational cost.

What actually changed

Skala 1.1 was trained on 2.5x more data than the first public version, with improvements across main-group thermochemistry, reaction kinetics, and molecular structure prediction. On the standard GMTKN55 benchmark, a suite of 55 categories covering thermochemistry, reaction barriers, and noncovalent interactions, Skala 1.1 reaches a weighted average error of 2.8 kcal/mol. That surpasses today's leading global (range-separated) hybrid functionals while retaining the efficiency of a semi-local functional. Concretely, it ranks first in 32 of 55 GMTKN55 categories at the cost of a meta-GGA.

The gains came from expanding the Microsoft Research Accurate Chemistry Collection (MSR-ACC), the training dataset generated with expensive coupled-cluster and other wavefunction methods. For 1.1, they added categories including electron affinities and noncovalent clusters, increasing both size and diversity of the training data. Beyond raw energies, the model now also delivers accurate electron densities, dipole moments, and geometries.

Getting out of the PySCF sandbox

The bigger story is distribution. Until now, Skala only ran through the open-source community release built on GPU4PySCF. That is fine for evaluation, but production computational chemistry lives in a fragmented ecosystem of decades-old codes tuned to specific problems. Skala is now landing where scientists actually work:

• CP2K - available today, integrated in collaboration with the CASUS team via the GauXC library, targeting large-scale systems and molecular dynamics.
• Psi4 - integration underway, giving three widely used open-source options alongside PySCF and CP2K.
• FHI-aims, ORCA, VASP - work in progress with the respective developer teams to cover the major commercial and academic platforms in chemistry and materials science.

To validate the CP2K port, the team ran matched calculations on a subset of GMTKN55 against the PySCF reference. The two implementations agree to within 0.1 kcal/mol MAD across the entire subset, with a single outlier corresponding to a particularly challenging radical system. The implementation details are written up in a joint paper with CASUS.

A living performance benchmark

The third piece of the release is a public benchmarking harness and living performance report that will be updated as new releases and library optimizations land. This matters because ML functionals have historically been hard to compare fairly across codes and hardware. The harness lets package developers benchmark their own Skala implementations against a shared reference.

The current numbers are competitive. On GPU, Skala 1.1 has the same cost as r2SCAN, and the hybrid functionals become more expensive for systems with more than roughly 1000 orbitals. On CPU, Skala has overhead versus other functionals for smaller systems that disappears for systems with more than about 300 orbitals.

Why this matters for practitioners

If you run DFT for catalysis, battery materials, or drug discovery, the practical pitch is that you can get accuracy previously reserved for expensive hybrid functionals at meta-GGA cost, called from a code you already use. The continuous-improvement framing is also a departure from the traditional DFT world, where dozens of functionals coexist and picking one is part science, part folklore. If Skala genuinely supersedes itself each release, that decision tree collapses over time.

The full technical writeup is on arXiv, and the community edition is on GitHub for anyone who wants to try it against their own systems before waiting for their preferred code to ship native support.