Adaptive multi-objective optimization of engine combustion via RL-enhanced NSGA-III and HDMR surrogates

This work proposes an adaptive multi-objective optimization framework that integrates a high-dimensional model representation (HDMR) surrogate with a Q-learning-enhanced NSGA-III algorithm (RLNSGA-III) for efficient combustion strategy optimization in internal combustion engines. The framework addresses two key challenges in engine optimization: the high computational cost of detailed CFD-based evaluations and the strong dependence of evolutionary algorithm performance on manually tuned parameters. A validated KIVA4-CHEMKIN II model of an n-octanol-fueled compression-ignition engine is used to generate the training database for constructing the HDMR surrogate. Reinforcement learning is employed to adaptively control crossover and mutation operations in NSGA-III, enabling the algorithm to automatically balance exploration and exploitation during the evolutionary process. Benchmark studies on ZDT, DTLZ, and UF test functions demonstrate that RLNSGA-III achieves superior convergence accuracy and competitive hypervolume performance relative to classical multi–objective algorithms. When applied to engine combustion optimization, the proposed HDMR-RLNSGA-III framework identifies injection strategies that reduce indicated specific fuel consumption by 6% and decrease CO emissions by 93.2% compared with the baseline. The results confirm that coupling surrogate modeling with reinforcement-learning-driven evolutionary optimization provides an effective and computationally efficient approach for clean and high-efficiency combustion strategy design.