Exploring chemistry and catalysis by biasing skewed distributions via deep learning

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Reaction Dynamics

Background:

• Automated discovery of chemical and catalytic reactions is challenging, especially in complex systems.
• Conventional methods struggle to identify optimal search directions for reaction pathways.

Purpose of the Study:

• To introduce Loxodynamics, a novel machine-learning-driven approach for reaction exploration.
• To address limitations of conventional techniques in discovering chemical reactions efficiently.

Main Methods:

• Utilizing biased molecular dynamics guided by the skewness of local probability distributions.
• Employing Skewencoder, an Autoencoder with a skewness-based loss function, to extract reaction coordinates.
• Iterative sample-and-search cycles to adaptively map free energy surfaces and capture finite-temperature effects.

Main Results:

• Demonstrated efficient identification of low-energy barrier directions and metastable states.
• Successfully validated across model potentials, gas-phase reactions (S_N 2, Diels-Alder), and catalytic alcohol dehydration.
• Showcased acceleration of reaction discovery without requiring elevated temperatures or a priori collective variable knowledge.

Conclusions:

• Loxodynamics offers a systematic framework for automated reaction discovery in computational chemistry.
• The method effectively overcomes key limitations of traditional approaches for complex reactive systems.
• Enables efficient exploration of reaction pathways by leveraging machine learning and biased molecular dynamics.