This thesis proposes a new solver by applying active learning to configu- ration interaction-based solvers in dynamical mean-field theory. Exact diag- onalization and configuration interaction methodologies face fundamental dif- ficulties due to expon...
This thesis proposes a new solver by applying active learning to configu- ration interaction-based solvers in dynamical mean-field theory. Exact diag- onalization and configuration interaction methodologies face fundamental dif- ficulties due to exponentially growing Hilbert space. This study presents a method that reduces the Hamiltonian matrix size by selectively choosing im- portant Slater determinants using machine learning. The proposed solver was validated on the one-dimensional Hubbard model by analyzing its convergence and scaling behavior. Calculations for Sr2RuO4 confirmed applicability to real materials, with self-energy and physical quantities validated against exact diag- onalization. The new solver is expected to be efficiently utilized in multi-orbital cluster dynamical mean-field theory.