Machine learning based optimization of valve timing and intake charge boosting in a hydrogen port fuel injection engine

Abstract

Hydrogen-fuelled internal combustion engines (H₂ICEs) present a promising near- term pathway for decarbonizing transport; however, their widespread adoption is constrained by abnormal combustion phenomena such as backfire and by trade-offs between performance and emissions. To address the above issues, current study developed a computational framework that integrates high-fidelity Computational Fluid Dynamics (CFD) simulations with machine learning-based surrogate modeling to optimize valve timing and intake charge boosting in a hydrogen port fuel injection (PFI) engine. CFD simulations were performed in the commercial code CONVERGE v4.1. The CFD model captured ignition kernel formation, flame propagation, and backfire evolution accurately. A Design of Experiments (DOE) using Latin Hypercube Sampling (LHS) quantified the influence of valve overlap timing (VOT) and intake pressure on indicated power, NOx emissions, and backfire risk. General trends showed that boosting strongly increases power however, it exacerbates NOx and backfire risk, while extended overlap reduces backfire but increases NOx. To reduce computational cost, an ensemble of surrogate models—comprising Random Forest, Support Vector Regression, Ridge Regression, Extreme Gradient Boosting, and Neural Networks— was trained, achieving R² values exceeding 0.95 across all objectives. These models enabled efficient optimization via a merit function framework, where tunable weights (α, β) balance competing objectives. Instead of conventional fixed-weight optimization, continuous trade-off surfaces were constructed across the (α, β) space, providing global insight into how weightings affect outcomes. This approach revealed key sensitivities and feasible regions under emission and backfire constraints, allowing informed selection of operating strategies. The developed CFD–ML framework and trade-off surface methodology provide a computationally efficient and generalizable strategy for multi-objective optimization in hydrogen engine design