Tests of Artificial Neural Network-Based Diabatization Approaches on Simple 1D Models

Recently, a novel diabatization scheme has been proposed [Shu, Y.; Truhlar, D. G. J. Chem. Theory Comput. 2020, 16, 6456-6464] using artificial neural networks. Most importantly, the method almost exclusively requires the knowledge of adiabatic energies, which are routinely obtained from ab initio calculations. However, many questions related to the favorable performance of the method remain unanswered. In the present paper, some of these questions are considered for selected one-dimensional models with one configurational variable. In particular, various activation functions are tested, including nonlinear ones in the output layer, the effect of the regularization term in the loss function is analyzed, and computationally cheap extensions of training sets are proposed. Significant improvements of the performance of the original method have been achieved.