Deep reinforcement learning of transition states

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Physics

Background:

• Understanding chemical reaction mechanisms is crucial for predicting reactivity and designing new chemical processes.
• Traditional methods for elucidating reaction pathways can be computationally expensive and may involve subjective interpretation.

Purpose of the Study:

• To develop an automated machine learning approach for unraveling chemical reaction mechanisms.
• To minimize subjective biases in the discovery of reaction pathways and transition states.

Main Methods:

• The proposed method, RL‡, integrates reinforcement learning (RL) with molecular dynamics (MD) simulations.
• Chemical reaction mechanism discovery is framed as a game, optimizing value and policy functions approximated by deep neural networks.
• RL‡ is trained tabula rasa, requiring minimal prior assumptions.

Main Results:

• RL‡ successfully identifies transition states and reaction pathways by formulating the problem as a game.
• The value function provides direct interpretability of the reaction mechanism.
• The policy function enables efficient sampling of the transition path ensemble for dynamics and kinetics analysis.

Conclusions:

• RL‡ offers an automated and unbiased approach to uncovering chemical reaction mechanisms.
• The method facilitates direct interpretation of reaction pathways and analysis of reaction dynamics.
• This integration of RL and MD holds significant potential for advancing computational chemistry research.