ReMLP-NET: A Neural Network Interaction Potential for Molecular Energy Prediction

Deep learning methods have seen increased applications and are now regularly used in chemistry, biology, and related areas. In chemical applications, they are used as an alternative to quantum mechanics (QM) methods to predict molecular energy with a more reasonable computational time. Here, we utilize a machine learning algorithm to predict the molecular total energy of structures containing elements H, C, N, O, F, S, and Cl. Optimized structures and total energies from the Retrievium repository, which includes molecules from the GDB13 and DUD-E sets, were used to train and assess a single Retrievium multilayer perceptron neural network (ReMLP-NET) using an atomic environment vector as the feature set. Symmetry function hyperparameters were selected using a Genetic Algorithm. The proposed method showed improved performance with MAE and root mean squared error values of 1.29 and 1.81 kcal/mol, respectively, compared to 1.53 and 2.16 kcal/mol for ReANI-2x.