Adaptive neural network controller for CA50 optimization under multiple constraints in hydrogen internal combustion engines

Hydrogen-fuelled internal combustion engines are a promising solution for sustainable mobility, but their efficiency strongly depends on combustion phasing, commonly quantified by the crank angle at 50% mass fraction burned (CA50). This work presents an adaptive CA50 controller that maximizes efficiency while maintaining reliability with respect to knock intensity and peak cylinder pressure (Pmax). Artificial neural networks (ANNs) are used to model the dependence of brake thermal efficiency, Pmax and knock intensity on CA50. To ensure robust performance under varying or aged conditions, ANN parameters are continuously updated online using real-time combustion data. The updated ANN outputs are processed through a desirability function to determine the optimal CA50 target. Simulation results show improved efficiency under off-design conditions: a 5% increase in EGR is compensated by advancing CA50 by about 4°, recovering most of the baseline efficiency while respecting Pmax and knock constraints. Real-time feasibility was confirmed on embedded hardware. • The combustion process of H2ICE is managed using the cylinder pressure information. • CA50–performance trends are determined using ANNs and used to set the target CA50. • ANNs parameters are updated online, adapting the optimal CA50 to actual conditions. • The optimization process takes into consideration efficiency, knock and Pmax. • Real-time feasibility demonstrated on embedded hardware up to 20,000 rpm.