Construction, Evaluation, and Optimization of Permutationally-Invariant Polynomial Expansion of Molecular Potential Energy Surfaces

Potential energy surfaces (PESs) provide a fast route for a variety of semiclassical and quantum dynamics as well as kinetics simulations for molecular systems. However, construction of PESs often requires extensive human efforts to tackle the complicated PES landscape or a large amount of computing power due to the complicated fit function forms. In this work, we demonstrate a computationally cost-effective strategy to perform parameterization and evaluation of PESs of phenol and nitrobenzene using permutationally-invariant polynomials (PIPs). The large number of PIP terms for these systems are tamed using a GPU-accelerated pruning procedure. In conjunction with this strategy, a GPU/CUDA implementation demonstrates up to 300x speedup (comparing 1 GPU with a single CPU core) in energy and grandient evaluation, enabling sampling of semiclassical trajectories to a statistical convergence. In principle, this protocol is aplicable for general molecular systems of interest.