Broadening access to Skala creates a faster path to predictive DFT
Skala 1.1 is now in CP2K, with more major electronic-structure codes next in line.
Microsoft Research says the updated deep-learning DFT functional was trained on 2.5x more data and improves accuracy on thermochemistry, reaction kinetics, and structure prediction. It reports a 2.8 kcal/mol weighted average error on GMTKN55 and says Skala 1.1 ranked first in 32 of the benchmark’s 55 categories. The CP2K implementation matched PySCF results within 0.1 kcal/mol MAD on a representative GMTKN55 subset. Microsoft is also publishing a living performance benchmark for future Skala releases and implementations. Microsoft Research AI's note
Microsoft Research says the updated deep-learning DFT functional was trained on 2.5x more data and improves accuracy on thermochemistry, reaction kinetics, and structure prediction. It reports a 2.8 kcal/mol weighted average error on GMTKN55 and says Skala 1.1 ranked first in 32 of the benchmark’s 55 categories. The CP2K implementation matched PySCF results within 0.1 kcal/mol MAD on a representative GMTKN55 subset. Microsoft is also publishing a living performance benchmark for future Skala releases and implementations. Microsoft Research AI's note
score 5