[2506.14665] Accurate and scalable exchange-correlation with deep learning

[Submitted on 17 Jun 2025]

View the PDF file from the paper entitled “Development and Development” with deep learning, written by Julia Louise and 23 other books

PDF view

a summary:DFT theory is the most widely used way to predict the properties of molecules and materials. Although the DFT, in principle, a delicate formulation of the Schrödinger equation has been rewritten, practical applications depend on the unknown functionality of the exchange relationship (XC). Most of the current XC functions are created using a limited set of handmade complex features that improve accuracy at the expense of mathematical efficiency. However, no current and general approximation of the predictive modeling of laboratory experiments at chemical accuracy – which is usually defined as errors is less than 1 kilo calorie/mall. In this work, we offer Skala, a modern XC learning -based XC function that goes beyond manually designed features through learning directly from data. Skala achieves the chemical accuracy of slopes for small particles while maintaining the typical mathematical efficiency of the semi -local DFT. This performance is enabled by training on an unprecedented volume of high -resolution reference data created using intense functional methods. It is worth noting that Skala systematically improves with additional training data that covers a variety of chemistry. By integrating a modest amount of high -resolution additional data designed for chemistry that exceeds decay energies, Skala achieves accuracy competition with the best hybrid functions in the chemistry of the main general group, at the expense of the semi -local DFT. With the continued expansion of the training data set, Skala is preparing to increase the enhancement of predictive strength to simulate the first principles.