A machine learning method to predict rate constants for various reactions in combustion kinetic models

Accurate prediction of temperature-dependent reaction rate constants is essential for the development of combustion kinetic models. However, the computational expense associated with calculating rate constants using high-level quantum chemistry methods becomes infeasible as the complexity of the kinetic models grows, and alternative approaches relying on analogies can exhibit significant inaccuracies. In recent times, as the field of combustion has generated a vast volume of kinetic data, the utilization of data-driven approaches, specifically machine learning, holds great promise in facilitating the development of kinetic models. In particular, natural language processing (NLP) models, such as ChatGPT, have become very useful. Here, we propose a deep neural network-based model to predict rate constants, and to explore the potential of machine learning methods to facilitate combustion kinetic model development. A diverse and high-quality dataset has been compiled concerning high-pressure limit reaction rate constants from nine important reaction classes. As the common representation of chemical reactions forms a language, we use the BERT transformer from that is part of common NLP techniques to generate reaction fingerprints from reaction SMILES. The model employs these reaction fingerprints as input to predict the three modified-Arrhenius parameters, i.e. the log of the frequency parameter (ln A), temperature exponent (n), and activation energy (Ea). A joint loss function is introduced to ensure that the rate constants calculated from the predicted Arrhenius parameters jointly provide good accuracy and to avoid overfitting. The final model achieves coefficients of determination (R2) of 0.74, 0.71, and 0.96 for the predictions of ln A, n, and Ea, respectively. The calculated rate constants, based on the predicted Arrhenius parameters, exhibit an R2 value of 0.95 across the temperature range of 500–2000 K. Additionally, the model's ability to predict rate constants in reaction mechanisms for different fuels is evaluated through species-based cross-validation.