Skala

Skala is a deep-learning exchange–correlation (XC) functional for density functional theory that delivers near-chemical accuracy (~1 kcal/mol) on atomization energies while preserving the cost profile of semi-local DFT. Instead of relying on hand-designed descriptor hierarchies, Skala learns electron-density representations directly via deep learning, trained on a dataset of around 150,000 high-accuracy molecular structures labeled with coupled-cluster references. On the W4-17 and GMTKN55 benchmarks it remains competitive with hybrid functionals at roughly 10% of their compute cost.

Skala addresses a 60-year accuracy bottleneck in computational chemistry: practitioners have long had to choose between cheap, less-accurate semi-local DFT and expensive hybrids or post-Hartree–Fock methods. By delivering hybrid-class accuracy at semi-local cost, Skala enables predictive — rather than interpretive — molecular simulation for drug discovery, battery electrolytes, catalysts, and carbon-capture materials. The dataset and code are released to the community, with early industry partnerships (including Merck and Flagship Pioneering) signaling commercial relevance, marking a clear example of Microsoft AI for Science’s “fifth paradigm” in computational chemistry.