Uncertainty Quantification and Computational Cost Reduction of a Chemical Reactor Network for a Hydrogen Combustor

This work presents an integrated framework to improve the predictive capability and computational efficiency of Chemical Reactor Network models for lean premixed swirl-stabilized hydrogen combustion, using the AHEAD burner as a reference test case. Six optimization algorithms benchmarked over twenty seeds under a common budget of about 2000 evaluations identify the Covariance Matrix Adaptation Evolution Strategy as the most reliable calibration method, with a 95% success rate. Skeletal mechanism reduction via the Directed Relation Graph with Error Propagation yields compact mechanisms with standalone speedups up to 2.06 and maximum nitrogen oxide errors below 0.71%. A Flow-Map In Situ Adaptive Tabulation scheme delivers per-equivalence-ratio speedups up to 24.88 with mean errors below 6 x 10^-4 %. Embedded in the calibration loop, the combined acceleration reduces the median wall-clock time of a twenty-seed campaign from about 70 to less than 5 hours, without significant loss in calibration quality. A Polynomial Chaos Expansion uncertainty analysis attributes about 87% of the joint variance to operating mass-flow perturbations. The resulting 95% uncertainty band brackets six of the eight experimental measurements, with the residual gap concentrated at the leanest points, where localized hot spots and ignition-reignition events not captured by the homogeneous network likely contribute to nitrogen oxide formation.