Toward Interpretable Graph-Based Machine Learning for Combustion Kinetic Model Generation
- 1Department of Mechanical Engineering, University of Minnesota–Twin Cities, 111 Church Street SE, Minneapolis, MN, USA, 55455
Abstract
This work presents an atom-conserving and interpretable graph neural network (GNN) for combustion kinetic model generation, where atoms are represented as nodes, covalent bonds are represented as edges, and reactions are represented as bond-rearrangement operations. Its artificial-intelligence (AI) contribution is a reaction generator that extends a standard Graph Isomorphism Network with Edges (GINE) architecture through reaction-context embeddings, reaction branching & third-body reaction inputs, Arrhenius-rate-parameter prediction, pairwise decoding, and stochastic inference over evolving reactant pools. Trained on the Li–Dryer hydrogen combustion kinetic model (8 species and 21 reactions), the GNN reproduces trained chemistry and generates element-balanced new reactions and intermediate species outside the training dataset, many of which have experimental analogues. When trained on a subset of the Li–Dryer dataset, the model recovered most held-out reactions and revealed relationships between samples (reactions) in the model’s latent space. The results suggest that the model learns transferable local reaction structures and generates valid species and reactions. As with any reaction generators, the generated reaction models require thermochemical and kinetic verification; the intended engineering application is low-cost reaction-pathway exploration for new fuels and prioritization of experiments and first-principles calculations to develop new combustion kinetic models.
Related
- Kinetic Mechanism Development and Pollutant Chemistry in NH3/H2/CH4 Combustion
- Interpretable Machine Learning Disentangles Pore Confinement Across Multiple Zr-MOF CO2 Capture Targets
- An Investigation into the Kinetic Persistence of TiO2 Polymorphs Using Machine-Learning-Driven Pathfinding in Crystal Configuration Space.
- Interpretable Machine Learning-Based QSPR Modeling for Predicting Hydrogen Adsorption Performance of Ester Oil Impregnating Agents
- Physics-inspired feature engineering and interpretable machine learning for thermoelectric properties of multicomponent materials
BibTeX
@misc{du2026toward,
title = {{Toward Interpretable Graph-Based Machine Learning for Combustion Kinetic Model Generation}},
author = {James H. Du and Suryanarayan Ramachandran and Suo Yang},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010767/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010767/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Toward Interpretable Graph-Based Machine Learning for Combustion Kinetic Model Generation
AU - James H. Du
AU - Suryanarayan Ramachandran
AU - Suo Yang
PY - 2026
DA - 2026/10/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010767/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010767/v1
AB - This work presents an atom-conserving and interpretable graph neural network (GNN) for combustion kinetic model generation, where atoms are represented as nodes, covalent bonds are represented as edges, and reactions are represented as bond-rearrangement operations. Its artificial-intelligence (AI) contribution is a reaction generator that extends a standard Graph Isomorphism Network with Edges (GINE) architecture through reaction-context embeddings, reaction branching & third-body reaction inputs, Arrhenius-rate-parameter prediction, pairwise decoding, and stochastic inference over evolving reactant pools. Trained on the Li–Dryer hydrogen combustion kinetic model (8 species and 21 reactions), the GNN reproduces trained chemistry and generates element-balanced new reactions and intermediate species outside the training dataset, many of which have experimental analogues. When trained on a subset of the Li–Dryer dataset, the model recovered most held-out reactions and revealed relationships between samples (reactions) in the model’s latent space. The results suggest that the model learns transferable local reaction structures and generates valid species and reactions. As with any reaction generators, the generated reaction models require thermochemical and kinetic verification; the intended engineering application is low-cost reaction-pathway exploration for new fuels and prioritization of experiments and first-principles calculations to develop new combustion kinetic models.
ER -