Speaker
Description
We present a framework for molecular vibrational coordinate optimisation based on generative machine-learning methods, in particular normalising flows.
The computational cost of vibrational configuration interaction (VCI) calculations is largely determined by the size of the correlation space. We show that this space can be substantially reduced by optimising curvilinear coordinates variationally within vibrational self-consistent field (VSCF) or VCI theory. Starting from conventional Z-matrix valence coordinates, a normalising-flow transformation is trained to minimise the ground state vibrational energy or the sum of energies of selected excited states.
The resulting coordinates provide a more compact representation of vibration correlation and accelerate the convergence of subsequent VCI and perturbation-theory calculations. For H2CO, CH3F, and trans-HCOOH, the optimised coordinates reduce the correlation spaces required for converged excited-state energies by factors of 3-7. The reduction reaches several tens for strongly coupled states in dense spectral regions.
The framework is implemented in Python/JAX, with its core routines parallelised for both CPUs and GPUs. Although coordinate optimisation does not remove the exponential scaling of VCI, it substantially extends the range of accurate state-specific calculations that are computationally feasible.