Exploring chemistry and catalysis by biasing skewed distributions via deep learning.

Abstract

The automated discovery of chemical and catalytic reactions remains a major challenge in computational chemistry, particularly in complex systems where conventional methods struggle to identify optimal searching directions. Here, we propose Loxodynamics, a machine-learning-driven approach for reaction exploration via biased molecular dynamics. By leveraging the skewness of local probability distributions, Loxodynamics dynamically determines low-energy barrier directions, efficiently guiding the system toward metastable states. The core of our framework is Skewencoder, an Autoencoder augmented with a skewness-based loss function that extracts reaction coordinates from minimal sampling data. Through iterative sample-and-search cycles, the system adaptively maps the free energy surface, capturing finite-temperature effects critical to complex reactive environments. We validate our method across model potentials, gas-phase reactions (S_N 2 and Diels-Alder), and catalytic alcohol dehydration in acidic chabazite under operando conditions. Loxodynamics provides a systematic framework for reaction discovery that addresses key limitations of conventional techniques, notably by allowing acceleration without elevated temperatures or a priori knowledge of collective variables.