Predicting the quasiparticle energies and bandgaps of bulk materials with machine learning techniques

This master thesis explores the use of machine learning (ML) techniques to predict bandstructures, specifically focusing on the GW approximation. The accurate prediction of bandstructures is critical for understanding electronic properties in materials science. While traditional methods like density functional theory (DFT) have limitations in capturing electronic correlation effects, the GW approximation offers more precise descriptions at a higher computational cost. ML presents an opportunity to bridge the gap between accuracy and computational efficiency: in this study, ML models such as RandomForest, ExtraTrees, XGboost and NeuralNetwork regressors were trained and evaluated on a dataset of 148 3D non-magnetic materials, achieving promising results in predicting G0W0 band gaps and energy corrections: the NeuralNetwork achieved the best accuracy in gap predictions on unseen data with a mean absolute error of 0.15 eV, while for the band energies prediction, the ExtraTrees, resulted in the best accuracy with a mean absolute error of 0.10 eV. The developed models offer accurate predictions at a fraction of the computational cost of traditional methods, opening avenues for accelerated materials discovery. Future extensions could include incorporating magnetic materials and other structural types to enhance model performance and versatility.