Abstract
Density Functional Theory (DFT) is the most widely used electronic structuremethod for predicting the properties of molecules and materials. Although DFTis, in principle, an exact reformulation of the Schr\"odinger equation,practical applications rely on approximations to the unknownexchange-correlation (XC) functional. Most existing XC functionals areconstructed using a limited set of increasingly complex, hand-crafted featuresthat improve accuracy at the expense of computational efficiency. Yet, nocurrent approximation achieves the accuracy and generality for predictivemodeling of laboratory experiments at chemical accuracy -- typically defined aserrors below 1 kcal/mol. In this work, we present Skala, a modern deeplearning-based XC functional that bypasses expensive hand-designed features bylearning representations directly from data. Skala achieves chemical accuracyfor atomization energies of small molecules while retaining the computationalefficiency typical of semi-local DFT. This performance is enabled by trainingon an unprecedented volume of high-accuracy reference data generated usingcomputationally intensive wavefunction-based methods. Notably, Skalasystematically improves with additional training data covering diversechemistry. By incorporating a modest amount of additional high-accuracy datatailored to chemistry beyond atomization energies, Skala achieves accuracycompetitive with the best-performing hybrid functionals across general maingroup chemistry, at the cost of semi-local DFT. As the training datasetcontinues to expand, Skala is poised to further enhance the predictive power offirst-principles simulations.