Machine learning of self-energies for accelerated DMFT calculations

We explore the use of machine learning for correlated d-electron systems within the DFT+DMFT framework. Specifically, we accelerate the DMFT calculations by training equivariant graph neural networks on self-energy and fermi level data, providing a better starting point for the impurity solver. First, high-throughput calculations on Fe, FeO and NiO using DMFT are conducted to build a suitable training set. Using the data generated from these calculations, we train equivariant graph neural networks in order to accelerate further calculations. We show that using this method, one can expect a 2-4 times speed-up in convergence time for DMFT calculations. We also utilize this method to generate more training points for Fe, construct a neural network potential, and use this potential to predict melting points at high pressure.