Enhancing Molecular Energy Predictions with Physically Constrained Modifications to the Neural Network Potential

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning-based force fields (MLFFs) are crucial for molecular simulations.
• Prioritizing only predictive accuracy of MLFFs is insufficient for real-world applications.
• Assessing the chemical rationality and practical utility of MLFFs is essential.

Purpose of the Study:

• Introduce SWANI, an optimized neural network potential based on the ANI framework.
• Enhance MLFFs by incorporating physical constraints for improved chemical realism.
• Evaluate SWANI's performance against existing models like ANI and graph neural networks (GNNs).

Main Methods:

• Developed SWANI by optimizing the ANI framework with additional physical constraints.
• Compared SWANI's predictive accuracy and potential energy profiles against the original ANI model.
• Benchmarked SWANI against a leading GNN-based model for molecular simulations.

Main Results:

• SWANI yields more chemically rational potential energy profiles than standard ANI.
• SWANI demonstrates superior predictive accuracy compared to the ANI model.
• SWANI outperforms the GNN-based model, especially for molecules larger than the training data.

Conclusions:

• SWANI offers improved chemical realism and predictive power for molecular simulations.
• The enhanced model exhibits excellent generalization capabilities for larger and unseen molecular systems.
• SWANI represents a significant advancement in developing reliable and accurate MLFFs for computational chemistry.