Including Physics-Informed Atomization Constraints in Neural Networks for Reactive Chemistry.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning interatomic potentials (MLIPs) offer high accuracy and efficiency for atomistic simulations.
• Current neural network (NN) based MLIPs struggle with accurately predicting energies of isolated or nearly isolated atoms.
• This limitation impacts the simulation of reactive processes involving such species.

Purpose of the Study:

• To develop a mathematical technique to enhance NN MLIPs for accurate prediction of isolated atom energies.
• To ensure consistent prediction of atomization energies (AE) across different system configurations.
• To improve the overall performance and reliability of MLIPs in chemical simulations.

Main Methods:

• Introduced a mathematical technique to modify existing atom-centered NN architectures.
• Developed AE-constrained versions of established MLIP models: HIP-NN-AE and ANI-AE.
• Evaluated model performance on AE prediction, bond dissociation energies, and extensibility tests.

Main Results:

• AE-constrained models demonstrated significantly improved AE prediction accuracy.
• The new technique ensures consistency in energy predictions, particularly for systems with isolated atoms.
• Performance improvements were observed in other tasks without compromising existing capabilities.

Conclusions:

• The proposed AE constraint method provides a robust solution for handling isolated atoms in MLIPs.
• This approach enhances the predictive power and reliability of neural network potentials.
• The technique offers a generalizable method to improve various NN MLIP architectures.