Improving Bond Dissociations of Reactive Machine Learning Potentials through Physics-Constrained Data Augmentation.

Area of Science:

• Computational chemistry
• Quantum mechanics
• Chemical kinetics

Background:

• Predicting bond dissociation energies (BDEs) is challenging for reactive systems due to multireference character.
• Single-reference methods and current machine learning interatomic potentials (MLIPs) struggle with accuracy for partially broken bonds.
• Generating accurate training data for dissociation pathways is computationally expensive.

Purpose of the Study:

• To improve the accuracy and reliability of reactive MLIPs for predicting BDEs and dissociation curves.
• To develop a cost-effective method for enhancing MLIP training data.
• To demonstrate the efficacy of the proposed approach on a relevant chemical system.

Main Methods:

• Physics-constrained data augmentation (PCDA) using the Morse potential to supplement training data.
• Augmenting existing MLIPs with inexpensive Morse potential data along dissociation pathways.
• Validating the improved MLIPs using a case study on methane combustion.

Main Results:

• The PCDA approach yields MLIPs with smooth bond dissociation curves.
• Achieved BDE predictions near coupled-cluster accuracy without expensive quantum calculations.
• The enhanced MLIP (ANI-1xnr improved with PCDA) shows superior performance for BDEs and dissociation curves compared to the original.
• The PCDA-trained MLIP maintains the reliability of the original model in reactive molecular dynamics simulations.

Conclusions:

• PCDA is an effective strategy for improving reactive MLIPs, particularly for BDE prediction and dissociation behavior.
• This method offers a significant advancement in computational chemistry by enabling accurate predictions without high computational cost.
• The approach successfully enhances existing MLIPs, demonstrating broad applicability in chemical reaction modeling.