Machine learning assisted particle swarm optimization of NH3/H2 combustion mechanism for laminar flame speeds

• Generated a dataset with 100,000 samples for machine learning (ML) model training. • Developed an ANN-XGBoost ML model to predict NH 3 /H 2 laminar flame speeds (LFSs). • Proposed an efficient ML-assisted particle swarm optimization for kinetic model. • The optimized NH 3 /H 2 mechanism accurately reproduces LFSs across wide conditions. Ammonia (NH 3 ) and hydrogen (H 2 ) are promising carbon-free fuels for future power generation. Laminar flame speeds (LFSs) are important fundamental characteristics crucial for NH 3 /H 2 combustion and conversion. However, existing NH 3 /H 2 mechanisms often exhibit limited accuracy in simulating LFSs across wide conditions, and their optimization remains computationally intensive due to the need for extensive premixed flame simulations. To address these challenges, an efficient optimization framework that couples machine learning (ML) models with a global optimization algorithm is proposed to optimize the NH 3 /H 2 mechanism for LFSs. A hybrid ML model combining neural networks and an extreme gradient boosting model was trained on 100,000 samples to accurately predict NH 3 /H 2 LFSs (R 2 = 0.9996, average relative error = 1.06% on the test set). Coupling the ML model with improved particle swarm optimization (ML-PSO), it outperformed Bayesian optimization (ML-BO), with the resulting mechanism reducing the squared prediction error for 1,050 collected experimental LFSs by 24.79% compared to the original mechanism. This improved accuracy stems from chemical effects that adjust combustion reactivity. The optimized mechanism also reproduces ignition delay times and species profiles data well. These results demonstrate the practicality of the proposed ML-PSO framework in NH 3 /H 2 mechanism optimization and its potential applicability to other fuels and combustion scenarios.