Dynamic surrogate-assisted particle swarm optimization for chemical kinetic mechanisms

Optimizing chemical kinetic mechanisms is challenging due to the large number of uncertain rate parameters and the high computational cost of kinetic simulations. While existing methods often rely on static surrogate models to approximate the kinetic response across the entire uncertainty space, they typically suffer from limited robustness and poor generalization in complex regions. This study proposes a dynamic surrogate-assisted particle swarm optimization (PSO) framework that adaptively refines the surrogate model during optimization to improve both accuracy and efficiency. The framework operates iteratively through three stages: active sampling, surrogate training, and swarm update. A radial basis function neural network (RBFNN) is employed as a surrogate model to predict kinetic responses. Based on the predictions and associated uncertainty, an active sampling strategy selects informative samples for kinetic simulations to obtain real responses. These samples are then used to incrementally retrain the surrogate, enhancing prediction accuracy in regions of interest. Finally, guided by the retrained surrogate, a truncated learning PSO algorithm updates the samples to efficiently find the optimal rate parameters. The proposed method is evaluated on an ammonia combustion mechanism and demonstrates superior optimization accuracy with orders-of-magnitude acceleration compared to state-of-the-art baselines. It also supports inverse uncertainty quantification for posterior analysis and uncertainty reduction. Furthermore, its scalability is validated by including more rate parameters for optimization, leading to a 550-dimensional space. These results highlight the effectiveness of the dynamic optimization strategy and its potential for advancing data-driven mechanism development in complex reacting systems.