Modeling H2O/Rutile-TiO2(110) Potential Energy Surfaces with Deep Networks

摘要

Successful and cost-effective water splitting could be one of the most interesting energy sources of the future. The modification of the transition metal oxide rutile-TiO2 as photocatalyst can lead to improved performance in the water splitting process. For that purpose, an accurate description of the interaction potential of a water molecule and the rutile-TiO2(110) surface in the ground and electronically excited state after photoexcitation is crucial. The electronic Schrödinger equation for the states involved is solved pointwise for different nuclear configurations within the Born-Oppenheimer approximation, and accurate fits to these energy points are required to obtain an analytic expression for the potential energy surface. This is too computationally expensive for fine-grained surface calculations of quantum chemical models. In this paper, we propose to use state-of-the-art deep learning techniques to provide accurate fits for this problem. Namely, we employ a fully connected variant of ResNet and DenseNet with heavy regularization (L2, RReLU, Dropout, and BatchNormalization). Previous literature applied neural network approaches before, but with unsatisfactory accuracy. In an experimental evaluation we show that the root mean squared error (RMSE) can be 6.8 times lower for the exited state and 12.7 times lower for the ground state compared to former approaches.