Using machine learning to go beyond potential energy surface benchmarking for chemical reactivity

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Chemical Reaction Dynamics

Background:

• Developing accurate potential energy surfaces (PES) for reactive chemistry, like hydrogen combustion, is computationally demanding.
• Traditional machine learning (ML) approaches for PES often rely on chemical intuition, leading to incomplete or unphysical configurations.
• Finite temperature and pressure conditions add complexity to modeling reactive systems.

Purpose of the Study:

• To develop an efficient and accurate ML model for predicting energies and forces in hydrogen combustion.
• To overcome limitations of data acquisition strategies in ML for reactive chemistry.
• To reduce the computational cost of molecular dynamics simulations for combustion processes.

Main Methods:

• Training an equivariant machine learning (ML) model for energy and force prediction.
• Implementing a 'negative design' data acquisition strategy using metadynamics within an active learning workflow.
• Utilizing a query-by-committee approach to efficiently select data for ab initio calculations.
• Developing a hybrid ML-physics model for molecular dynamics simulations.

Main Results:

• The 'negative design' strategy successfully generated ML models that avoid unphysical energy configurations.
• The ML potential energy surfaces converged more rapidly, improving efficiency.
• The hybrid ML-physics model achieved a two-orders-of-magnitude reduction in computational cost.
• Enabled prediction of free-energy changes in transition-state mechanisms for hydrogen combustion.

Conclusions:

• A 'negative design' data acquisition strategy is effective for creating robust ML potentials for reactive chemistry.
• Hybrid ML-physics models significantly enhance the efficiency of molecular dynamics simulations.
• This approach accelerates the study of complex chemical reactions like hydrogen combustion.