Supervised NN Model for Data Augmentation of 4D PES and Application in Quantum Dynamics

The study aims to utilize neural network (NN) models for augmenting molecular potential energy surfaces (PES) within spectroscopic accuracy (i.e. error < 1 cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup>). The current methods for solving the Schrödinger equation within such accuracy scale poorly with system size. The PES is an essential starting point for studying any quantum dynamical process occurring in a given molecular system. Since the molecular Hamiltonian quickly surpasses the current computing limits due to the large dimensionality of the equation, the scalability and ability to handle a large number of dimensions make NN models an ideal candidate for such applications. The current paper takes an example of an NCCN-H<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> collisional system which results in a 4D PES under rigid rotor approximation. Several supervised NN models are created and benchmarked by splitting the dataset into training and testing datasets. The final model can augment the PES within hours, delivering results that otherwise need years of computation.