Accelerating Electronic Structure Calculations with Deep Neural Networks

Density Functional Theory (DFT) is a widely used method for computing numerous properties in materials science and chemistry. Despite its usefulness, its inherent computational scaling with system size often limits its applicability and makes large-scale simulations infeasible. Our team develops a machine-learning approach to accelerate Kohn-Sham DFT calculations using deep neural networks within the Materials Learning Algorithms (MALA) Python framework. Our method employs bispectrum descriptors to encode roto-translationally equivariant representations of atomic neighborhoods as input to the neural networks. The networks then predict the local density of states (LDOS) from which various quantities of interest can be calculated. Our MALA models are trained on DFT data which serve as the ground-truth for the machine-learning models. Our MALA models significantly reduce the computational demands compared to conventional DFT methods. In this poster, we demonstrate the efficacy of our approach on a system of hydrogen molecules under various pressure and temperature conditions, showcasing the potential of our methods for molecular systems. Furthermore, we explore the use of SE(3)-equivariant graph neural networks (equiformer GNNs) to enhance the generalizability and extrapolation capabilities of our models while further reducing the computational cost. Our results indicate that the MALA framework provides a powerful and efficient tool for accelerating Kohn-Sham DFT calculations in molecular systems. The proposed approach has the potential to revolutionize the field of materials science by enabling researchers to perform large-scale simulations and explore complex molecular systems more efficiently. This work paves the way for future research in developing advanced machine-learning algorithms for accelerating electronic structure calculations and determining properties in materials science and chemistry both accurately and efficiently.