Learning the committor without collective variables.

Area of Science:

• Computational chemistry
• Machine learning in molecular dynamics
• Artificial intelligence for chemical processes

Background:

• Predicting molecular transitions is crucial for understanding chemical reactions and material properties.
• Traditional methods often rely on hand-crafted collective variables, limiting their applicability and interpretability.
• Developing automated methods for analyzing complex molecular dynamics is an ongoing challenge.

Purpose of the Study:

• To introduce a novel graph neural network architecture for predicting the committor function directly from atomic coordinates.
• To enable atom-level interpretability in molecular transition analysis without prior assumptions.
• To accurately estimate rate constants and identify key atomic contributions in molecular processes.

Main Methods:

• Development of a graph neural network architecture utilizing geometric vector perceptrons.
• Direct prediction of the committor function from raw atomic coordinates.
• Application and validation across diverse molecular systems.

Main Results:

• The graph neural network accurately infers the committor function across various molecular systems.
• The method provides atom-level interpretability, highlighting critical atoms in transition mechanisms.
• Precise estimation of rate constants for underlying molecular processes was achieved.

Conclusions:

• The proposed approach facilitates collective-variable-free learning for molecular dynamics.
• Automated identification of physically meaningful reaction coordinates is enabled.
• This method enhances the understanding and modeling of complex molecular transitions.