Constructing Collective Variables Using Invariant Learned Representations

Area of Science:

• Computational chemistry
• Materials science
• Chemical engineering

Background:

• Atomistic simulations model chemical reactions as rare events, necessitating bias along collective variables (CVs).
• Identifying suitable CVs a priori is often complex and challenging for accurate reaction pathway exploration.

Purpose of the Study:

• To develop a novel, automated method for discovering collective variables (CVs) in chemical reactions.
• To enable efficient and accurate atomistic simulations of rare chemical events, particularly in complex catalytic systems.

Main Methods:

• A new method for CV discovery using dimensionality reduction of atomic representation vectors.
• Employing linear-scaling and invariant representations, either fixed or learned via supervised training of machine learning potentials.
• Demonstration on four high-barrier reactions, including gas-phase and heterogeneous catalysis.

Main Results:

• The proposed method demonstrates high data efficiency and automated feature extraction.
• Learned representations capture both structural and energetic features, transferable across reactions.
• Successful application to complex reactions in heterogeneous catalysts, showcasing favorable scaling and invariance preservation.

Conclusions:

• This approach facilitates fast and largely automatic construction of suitable CVs for complex reactive scenarios.
• The method is expected to significantly advance the study of reactive/catalytic transformations, including those at solid-liquid interfaces.