Georg Trautmann

Georg Trautmann is a PhD student working on machine learning models in electronic structure theory. He is developing machine learning models for the prediction of Hamiltonians matrices, with the goal of enabling accurate electronic structure calculations for systems beyond the reach of conventional ab initio methods.

PhD Research

Started: October 2026

Georg’s doctoral research focuses on developing machine learning surrogates of electronic Hamiltonians that are trained directly on physical observables, such as eigenvalues, band structures, and densities of states, rather than on full matrix representations. This observable-based training approach aims to free Hamiltonian models from their dependence on a fixed orbital basis and electronic structure code, and to eventually enable the incorporation of experimental data as training targets.

Research Approach

His work combines equivariant deep learning with concepts from quantum chemistry, including:

• Equivariant graph neural network architectures for the prediction of electronic Hamiltonian matrices, building on the MACE-H framework
• Composite loss functions that incorporate both matrix element and observable targets, with numerically stable training through eigenvalue-dependent quantities
• Extension to correlated Hamiltonians beyond mean-field, learned from DMRG data via effective one-body representations

Background

Georg holds bachelor’s and master’s degrees in physics from the University of Heidelberg, where he specialized in theoretical and computational physics. He completed his master’s thesis at the Kirchhoff Institute of Physics on superfluid dynamics, combining a mean-field theoretic approach with a holographic description based on the AdS/CFT correspondence, providing him a strong foundation in theoretical physics and scientific computing.

During his studies he completed two Erasmus+ semesters at the University of Birmingham. Prior to his PhD, he worked as a research assistant at the Max Planck Institute for Meteorology in Hamburg, where he applied machine learning methods to the discovery of pattern formation in cloud time series data.

Research Impact

Accurate electronic structure calculations for large, complex material systems, such as moiré bilayers, organic-inorganic heterostructures, and strongly correlated molecules, remain out of reach for conventional ab initio methods. By developing Hamiltonian surrogates that learn from observables rather than full Hamiltonian matrices, Georg’s work aims to bridge the gap between first-principles theory and experimental spectroscopy, opening the door to electronic structure predictions at scales and accuracy levels currently inaccessible.