A New Method of DFT Functionals Parametrization via Deep Learning Technologies

The project presents a new approach to parameterization of DFT functionals (Density Functional Theory), which combines the advantages of machine learning and physically grounded functionals. Using neural networks to optimize hyperparameters while preserving the original the functional formula, the accuracy of calculations is increased without compromising the physical basis of the functional. Promising results have been obtained, indicating a significant increase in the accuracy of calculations compared with traditional methods. It is important to note that this improvement was achieved while maintaining the theoretical basis of the original functionality. This approach has great potential to advance the development of DFT functionals, which is important for various fields, including computational chemistry and materials science. The importance of developing this approach lies in the fact that it allows you to take advantage of empirical parameterization without neglecting the physical basis of the functional. This makes it possible to combine existing approaches to the construction of DFT-functionals, taking only the strengths from each of them. Thus, it turns out to combine in one DFT architecture both strict physical validity and training on experimental or ab initio data. The proposed method can be easily applied to many other density functionals. In particular, our laboratory is currently conducting research on the application of this approach to the PBE GGA functionality. Perhaps neural network parameterization in the future will become a common way to improve the quality of routine calculations and will be implemented in one form or another in quantum chemical programs