Microsoft Research has recently released a deep learning exchange-correlation (XC) functional called Skala, aimed at providing an efficient computational scheme for Kohn–Sham density functional theory (DFT).
Skala achieves computational efficiency comparable to current meta-GGA functionals by learning non-local effects, while reaching the accuracy level of hybrid functionals. In the atomic energy evaluation of the W4-17 molecular system, the mean absolute error (MAE) reached 1.06 kcal/mol, and even 0.85 kcal/mol on the single-reference subset; in the GMTKN55 benchmark test, Skala's weighted mean absolute deviation (WTMAD-2) was 3.89 kcal/mol. These results show that Skala can compete with top-tier hybrid functionals in terms of accuracy.
The design goal of Skala is to achieve strict main-group thermochemical calculations, rather than providing a universal functional applicable to all fields immediately. The model does not attempt to learn dispersion effects, and the initial version still uses fixed D3 (BJ) dispersion corrections. This tool is particularly suitable for main-group molecular chemistry fields requiring semi-local cost and hybrid-level accuracy, such as high-throughput reaction energy (ΔE), reaction barrier estimation, conformational/free radical stability ranking, and geometry and dipole prediction.
The architecture and training process of Skala are divided into two stages: first, pre-training on B3LYP density to extract XC labels of high-level wavefunction energy; second, fine-tuning within the SCF loop using Skala's own density without relying on backpropagation for SCF. The model training of Skala is based on a large-scale, high-quality dataset of atomic energies, including approximately 80,000 high-precision total atomic energies (MSR-ACC/TAE).
To ensure efficiency, Skala maintains a computational complexity of O(N³) and is optimized for GPU execution. The open-source code and toolkit of this model have been released on Azure AI Foundry Laboratory and GitHub, allowing users to run it directly on PySCF/ASE and GauXC platforms, facilitating efficient batch SCF calculations.
Project: https://github.com/microsoft/skala?tab=readme-ov-file
Key Points:
🌟 Skala has an MAE of 1.06 kcal/mol on W4-17 and 0.85 kcal/mol on the single-reference subset, demonstrating high precision.
🛠️ The model achieves computational efficiency similar to current meta-GGA functionals by learning non-local effects, focusing on main-group chemistry.
🚀 Skala is available on Azure AI Foundry Laboratory and GitHub, supporting efficient molecular computations and experimental sharing.
