Neural network based coupled diabatic potential energy surfaces for reactive scattering.

An approach for the construction of vibronically coupled potential energy surfaces describing reactive collisions is proposed. The scheme utilizes neural networks to obtain the elements of the diabatic potential energy matrix. The training of the neural network employs a diabatization by the Ansatz approach and is solely based on adiabatic electronic energies. Furthermore, no system-specific symmetry consideration is required. As the first example, the H2+Cl→H+HCl reaction, which shows a conical intersection in the entrance channel, is studied. The capability of the approach to accurately reproduce the adiabatic reference energies is investigated. The accuracy of the fit is found to crucially depend on the number of data points as well as the size of the neural network. 5000 data points and a neural network with two hidden layers and 40 neurons in each layer result in a fit with a root mean square error below 1 meV for the relevant geometries. The coupled diabatic potential energies are found to vary smoothly with the coordinates, but the conical intersection is erroneously represented as a very weakly avoided crossing. This shortcoming can be avoided if symmetry constraints for the coupling potential are incorporated into the neural network design.