Physics-informed soft actor-critic for constraint-aware combustion control of methanol/diesel dual-fuel engines

Cycle-resolved combustion control of methanol/diesel dual-fuel engines under dynamic operating conditions is challenging due to strong nonlinearity, actuator coupling, and combustion safety constraints. To address these challenges, a physics-informed Soft Actor-Critic (PISAC) framework is proposed by integrating a physics-informed neural network (PINN) surrogate, an SAC-based control policy, and a constraint-aware reward formulation. The PINN embeds thermodynamic constraints into neural-network training to achieve physically consistent prediction of in-cylinder pressure and combustion indicators, providing a reliable interaction environment for reinforcement learning. Based on the surrogate environment, the SAC agent learns coordinated multi-variable control actions through offline policy optimization. Results show that the proposed PINN achieves a pressure reconstruction RMSE of 3.2 bar, while the R2 values above 0.98 for CA50, IMEP, MPRR, and COVIMEP prediction. Under dynamic operating conditions, the proposed controller reduces CA50 and IMEP RMSE from 0.075°CA to 0.271 bar to 0.026°CA and 0.099 bar, respectively, compared with a conventional PI controller. Meanwhile, MPRR violations are eliminated and COVIMEP decreases from 3.41% to 2.97%, indicating improved combustion robustness and stability. Compared with MPC, the proposed framework further improves transient response and actuator coordination under rapidly varying operating conditions. In addition, the trained policy achieves real-time feasibility with an inference time of 3 ms per cycle. These results demonstrate that the proposed PISAC framework provides an accurate, robust, and computationally efficient approach for constraint-aware combustion control of nonlinear dual-fuel engines.