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

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Deep learning models are increasingly used in chemistry as faster alternatives to quantum mechanics (QM) for predicting molecular properties.
• Accurate prediction of molecular energy is crucial for various chemical applications.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for predicting the total molecular energy of chemical structures.
• To compare the performance of the proposed method against existing models.

Main Methods:

• Utilized a multilayer perceptron neural network (ReMLP-NET) trained on optimized structures and total energies from the Retrievium repository.
• Employed an atomic environment vector as the feature set and a Genetic Algorithm for selecting symmetry function hyperparameters.
• Trained and assessed the model using molecules from the GDB13 and DUD-E sets.

Main Results:

• The ReMLP-NET model achieved a Mean Absolute Error (MAE) of 1.29 kcal/mol and a root mean squared error (RMSE) of 1.81 kcal/mol.
• These results represent an improvement over the ReANI-2x method, which had MAE and RMSE values of 1.53 and 2.16 kcal/mol, respectively.

Conclusions:

• The developed ReMLP-NET model demonstrates superior performance in predicting molecular total energy compared to ReANI-2x.
• This deep learning approach offers a computationally efficient and accurate method for molecular energy prediction in chemistry.