Skala: Accurate and scalable exchange-correlation with deep learning: Deep learning breaks the DFT accuracy-efficiency trade-off, ushering a new paradigm in computational chemistry
In June 2025, Meta FAIR released Skala, a deep learning-based exchange-correlation functional, achieving an error of 2.8 kcal/mol on the GMTKN55 benchmark, surpassing state-of-the-art hybrid functionals while maintaining the low computational cost of semilocal DFT. Training data comes from high-accuracy wavefunction methods, enabling systematic improvability of deep learning functionals for the first time.
Skala learns nonlocal representations of electronic structure through deep learning, breaking the long-standing trade-off between accuracy and efficiency in traditional DFT. Its error of 2.8 kcal/mol approaches the accuracy of wavefunction methods, but with computational cost comparable to semilocal DFT, enabling large-scale high-accuracy calculations. This marks a paradigm shift from hand-crafted functionals to data-driven improvable models in computational chemistry.
Skala uses a deep neural network to directly learn the exchange-correlation energy density, with inputs such as electron density and its gradients as local features, but captures long-range correlation effects through nonlocal attention mechanisms. Training data comes from high-accuracy wavefunction methods such as CCSD(T), at an unprecedented scale. On GMTKN55 (a main-group chemistry benchmark), Skala achieves an error of 2.8 kcal/mol, outperforming B3LYP (~4.5) and ωB97M-V (~3.5), with computational cost comparable to PBE. Scaling laws show that error continues to decrease with more data.
Skala will directly impact fields such as drug design, catalysis, and materials science, enabling DFT calculations to scale to larger systems (e.g., protein-ligand complexes) while maintaining accuracy. After Meta open-sources the model, it may replace traditional functionals as the new default choice and drive updates in computational chemistry software (e.g., VASP, Gaussian).
Recommend that computational chemistry software companies (e.g., Schrödinger, Materials Design) immediately evaluate Skala and consider integrating it into their products. Investment should focus on derivative tools within Meta's open-source ecosystem and high-throughput virtual screening services based on Skala.
Focus on Skala's generalization ability for systems involving transition metals, excited states, and weak interactions, as well as the difficulty of integration with existing DFT codes. If the community can reproduce its accuracy, it will accelerate AI-driven discovery of new materials. Attention should be paid to the impact of training data bias on systems containing heavy elements.