Ammonia combustion kinetic model optimization using elementary rate coefficient and global combustion data

Ammonia combustion kinetic models differ in key rate coefficients, and achieving consistent agreement with measurements across diverse conditions remains challenging. We optimize a detailed ammonia kinetic model derived from C3MechV4.0.1 for pure NH3, NH3/H2, and NH3/NO mixtures. The optimization uses direct rate determinations from 142 publications and 11,232 indirect combustion measurements from 113 publications, covering species concentrations, ignition delay times, laminar burning velocities, and extinction strain rates. Random-effects meta-analysis combines direct determinations into a prior distribution accounting for individual source uncertainties and additional disagreement between sources. General Bayesian inference updates this prior using a generalized likelihood based on the curve matching (CM) score, which quantifies agreement between model predictions and indirect measurements in values and trends. Neural network surrogates approximate these predictions for efficient likelihood evaluation. Sensitivity-based clustering partitions the indirect datasets into seven kinetic groups with characteristic chemistry. Networks sharing hidden layers within each group reduce mean surrogate curve error by 60 % relative to independent networks trained on the same data. Relative to the nominal model, the optimized model improves agreement in all seven kinetic groups and 71 % of individual datasets. Its overall weighted mean CM score exceeds those of all 54 published models evaluated. Optimization with the direct data-informed prior yields indirect data agreement comparable to that obtained with a uniform prior, while achieving substantially closer agreement with direct data. Chemical analysis links the improved predictions to underlying rate changes, while the posterior quantifies remaining rate uncertainties and identifies NH2 decomposition, H2NO chemistry, and N2Hx pathways as measurement priorities. The optimized ammonia chemistry is integrated into C3MechV4.2.0.