On the role of error metrics in the assessment of chemical kinetic mechanisms

Methods intended to decrease the computational cost of chemistry, such as mechanism reduction, neural-network surrogates, or genetic optimization, rely on metrics to assess agreement with reference datasets. However, the role of error definitions in the design of these metrics remains poorly examined. This work presents a systematic analysis of error metrics and fitness functions commonly used for the assessment and optimization of chemical kinetic mechanisms. Fitness functions are constructed based on elementary design choices, including quantities of interest, error formulations, data transformations, and aggregation strategies. Their behavior is first analyzed a priori using controlled perturbations of time-resolved species profiles. The methodology is then evaluated using zero-dimensional ammonia/hydrogen combustion simulations, comparing detailed and reduced mechanisms. Novelty and significance statement The results demonstrate that standard relative-error metrics introduce significant biases across operating conditions. Conversely, an absolute error formulation applied to logarithmically transformed and min–max normalized profiles ensures a robust and consistent evaluation. This metric configuration is shown to be essential for genetic algorithms to converge toward highly accurate and generalizable reduced mechanisms. This work emphasizes the influence of the fitness function on the outcome of methods reducing the cost of chemistry integration, an aspect often treated as a secondary technical choice. By decomposing commonly used assessment criteria into elementary design features and analyzing their behavior under controlled perturbations, this study reveals fundamental sensitivities and biases that directly affect mechanism comparison and optimization outcomes. A key strength of this contribution lies in its generality: the proposed analysis and conclusions apply to reduction, optimization, and data-driven strategies alike.