Speaker
Description
Accurately predicting the electronic properties of materials under realistic conditions remains a central challenge in materials science, particularly for systems that deviate from ideal crystalline order, such as those with disorder, defects, or thermal fluctuations. Many-body methods based on the GW approximation provide the accuracy needed to capture these effects, but their high computational cost has limited their use to small, high-symmetry systems. This presents a major obstacle to modelling complex materials, including amorphous solids and systems at finite temperatures where atomic positions fluctuate away from equilibrium. Here, I will discuss our progress in developing a machine-learning framework that learns simplified, interpretable models capable of reproducing GW-calculated electronic energies using only inputs from standard density functional theory. Trained on a single high-symmetry reference structure, these models retain GW-level accuracy while generalizing to systems with broken symmetry and increased structural complexity. This enables predictive electronic-structure calculations across length scales at a fraction of the cost, making many-body accuracy accessible for structurally complex material systems.
