Machine learning-driven prediction of NH₂-n-alkane reactions for constructing high-precision ammonia-macromolecule hydrocarbon combustion models

The co-combustion of ammonia and large hydrocarbon fuels has been proven to be an effective strategy for reducing carbon emissions. However, the insufficient understanding of C-N cross-reactions poses significant challenges to combustion modeling. Due to the lack of data on the reaction NH₂ + macromolecular n-alkane, this reaction relys heavily on estimation derived from small alkanes. Neural networks (NN) provide viable alternative method for predicting rate constants for macromolecules. To establish a dataset for the NN model, the rate constants of NH₂ + C_nH_2n+2 (n= 1-4, 6, 8, 10, 12) were theoretically calculated at the DLPNO-CCSD(T)/CBS(T-Q)//M06-2X/def2-TZVP level. A variety of descriptors (including RDKIT descriptors and quantum chemistry descriptors) were comprehensively considered, and redundant descriptors were eliminated based on the stepwise regression algorithm and Pearson correlation coefficient. The optimal hyperparameter was obtained via OPTUNA code [1], and the repetitive training of different random seeds further proved the reliability and efficiency of the hyperparameters. Based on the partial reaction rate database (NH₂/C₁-C₁₀), a DNN model was trained and its ability to extrapolate to predict the reaction rates of larger molecule (NH₂/C₁₂) was validated. Training results show that the NN model exhibit good performance with high determination coefficient R² and small RMSE, but deviations are still observed in the thermodynamic limit (near 300 K or 2000 K) and the medium-high temperature regions. The predicted rate constants of NH₂ + C₁₆H₃₄ by NN model differ from RMG-py and rate rules, but overall they are distributed within the range of uncertainty. Kinetic modelling shows that when using the rate constants of NN model, the mechanism can simulate the ignition delay of the NH₃/C₁₆H₃₄ mixture. In summary, the proposed method is supposed to provide accurate and affordable rate constants prediction of macromolecular reaction systems, accelerating the development process of macromolecular mechanisms in real transportation fuel.