Microsoft Research Launches Skala: A Deep Learning Exchange-Correlation Functional for Semi-Local Costs with Mixed Precision

Microsoft Research introduces Skala, a deep learning exchange-correlation functional, significantly improving the computational efficiency of Kohn–Sham density functional theory. The model simulates non-local effects, achieving the accuracy of hybrid functionals while maintaining the speed comparable to meta-GGA. Tests show that the average absolute error in atomization energy prediction for W4-17 molecular systems is only 1.06 kcal/mol, and the error for a single-reference subset is reduced to 0.85 kcal/mol, with excellent performance on the GMTKN55 benchmark.