Reproducing global potential energy surfaces with continuous-filter convolutional neural networks.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Neural networks offer accurate analytic potential energy surfaces (PES) at reduced computational cost.
• SchNet architecture utilizes continuous-filter convolutional neural networks for PES fitting.
• Previous studies validated SchNet for reproducing energies and forces from molecular dynamics simulations.

Purpose of the Study:

• To demonstrate SchNet's capability in reproducing global potential energy surfaces (PES) for chemical reactions.
• To evaluate SchNet's performance on systems involving bond breaking and formation.
• To assess the accuracy of SchNet models for energies and forces across various chemical systems.

Main Methods:

• Utilized the SchNet architecture, a type of continuous-filter convolutional neural network.
• Trained and tested SchNet models on potential energy surfaces of specific chemical reactions and systems.
• Evaluated model performance using root-mean-squared error (RMSE) for energies and mean absolute error (MAE) for forces.

Main Results:

• SchNet models accurately reproduced the potential energy surfaces for H + H2 and Cl + H2 reactions.
• Achieved low test set RMSE for energies: 0.52 meV (H + H2) and 2.01 meV (Cl + H2).
• Demonstrated good performance on OCHCO+ and H2CO/cis-HCOH/trans-HCOH systems with test set RMSEs of 2.92 meV and 13.55 meV, respectively.

Conclusions:

• SchNet architecture effectively models global potential energy surfaces, including regions with bond breaking/formation.
• The accuracy of SchNet approaches that of ab initio methods at a significantly lower computational cost.
• SchNet shows promise for accelerating molecular simulations and chemical reaction studies.