The steel industry urgently needs efficient waste heat recovery from sintered ore to reduce carbon emissions. Traditional optimization methods, relying on single-variable computational fluid dynamics (CFD) or experiments, fail to resolve the complexities of high-dimensional parameter interactions and face difficulties in solving multi-objective optimization problems involving both the amount of heat transfer ( \(Q\) ) and the amount of exergy destruction ( \({E}_{{x},{d}}\) ). An integrated framework combining optimized Latin hypercube sampling (OLHS), metaheuristic-optimized surrogate models, explainable artificial intelligence (XAI), and NSGA-II multi-objective optimization for a novel sintered dual-stage cooling unit was proposed. OLHS generates spatially uniform six-dimensional training data, enabling high-fidelity CFD response modeling with minimal simulations. Hybrid support vector regression models achieve exceptional accuracy (R2 > 0.999 for both \(Q\) and \({E}_{{x},{d}}\) ), validated by tenfold cross-validation. SHapley additive exPlanations and partial dependence plot analyses reveal that the particle mass flow rate and the gas inlet temperature dominate \(Q\) , while the gas inlet temperature and volume flow rate of cooling air to pre-cooling unit critically influence \({E}_{{x},{d}}\) , with multi-parameter synergy driving trade-off. NSGA-II resolves the \(Q\) – \({E}_{{x},{d}}\) conflict, yielding a Pareto front with 24.4% hypervolume improvement, and achieves a multi-objective optimization of Q = 65.26 MW and Ex, d= 37.43 MW, balancing 81% peak heat transfer and 54% lower exergy destruction.
