A Hybrid ReaxFF–Cheminformatics–Machine Learning Architecture for Modelling Pyrolysis Pathways

Abstract

Artificial intelligence (AI) and machine learning (ML) can revolutionize chemical reaction research by providing data-driven methods that accelerate the analysis of complex datasets. This is particularly important for reactive molecular dynamics simulations, where the large number of reaction events generated makes mechanistic and kinetic analysis challenging. This work introduces the first hybrid framework that combines ReaxFF (Reactive Force Field) molecular-dynamics simulations, cheminformatics, and machine learning to automate and accelerate the modeling and analysis of pyrolysis pathways. The proposed novel methodology integrates four components: (i) ReaxFF simulations, (ii) pathway construction from annealing simulations through cheminformatics-based filtering and graph-based pathway analysis, (iii) automated assignment of chemical transformations through a supervised machine-learning model, and (iv) determination of reaction kinetics. Two case studies were investigated during framework development, including cyclohexanone, a promising second-generation biofuel. Ethene, ethenone, and ethyne were identified as the main products of cyclohexanone pyrolysis. Subsequent steps, all implemented in Python, constructed the dominant pathways leading to these products. For labelling all reactions, training and evaluation were performed with cross-validation on 16 reaction classes. Linear SVC achieved the highest macro F1-score of 0.99, while test-set validation attained a macro F1-score of 0.989. The labelled pathways revealed that cyclohexanone decomposition proceeded mainly via ring-opening α-, β-, and γ scissions, accounting for around 85% of the observed events, followed by Isomerization to enol and H-abstraction. Furthermore, cyclohexanone decomposed at Ea = 60.8 kcal mol-1 and Ao = 1.7 × 103 ps-1. To assess the workflow’s validity, a second case study was carried out on the pyrolysis of vegetable glycerin (VG), yielding labelled full mechanisms and kinetic constants consistent with previous work. The pre-trained classifier retained strong performance on VG reaction events, with a macro F1-score of 0.992. This validated framework provides a scalable template for rapid and accurate mechanism generation and kinetic analysis across a wide range of chemical processes.