Enhancing Vibronic-Coupling Hamiltonian Parameterization with Machine Learning: The PyVCHAM Tool.

Nonadiabatic effects play a key role in the photophysics and photochemistry of molecular systems, yet their efficient inclusion in quantum molecular dynamics simulations remains challenging, due to the need to construct accurate representations of the molecular Hamiltonian within the manifold of relevant electronic states. Here, we present the PyVCHAM library, which builds on the established multimode vibronic-coupling framework, enhanced by modern machine learning techniques for efficient parameter optimization. The code interfaces with electronic structure packages to generate potential energy surfaces, enabling the parametrization of diabatic Hamiltonians for quantum dynamics calculations. Leveraging the optimization of specialized loss functions, the use of automatic differentiation to compute their analytical gradient in parameter space, and the availability of a wide range of optimization algorithms, the core engine substantially improves in terms of accuracy, flexibility, and efficiency compared to existing implementations. The PyVCHAM library introduces a standardized format for storing vibronic-coupling Hamiltonians based on the JSON data format. It also introduces the ability to combine an arbitrary number of existing vibronic Hamiltonians into interacting supersystems or aggregates, where the constituents couple through dipole-dipole interactions. Several illustrative examples highlight the program's ability to treat complex, high-dimensional molecular systems that were previously difficult to access.