Intelligent optimization of steam-curing for precast concrete tunnel lining: Insight from sample data parameters analysis and interpretable machine learning

• Proposes a novel BO-CatBoost-NSGA-III framework for intelligent, multi-objective optimization of concrete steam-curing. • Integrates SHAP analysis to reveal thermostatic temperature, heating rate, and duration as critical, interpretable drivers. • Achieves significant performance gains: 8.55% higher strength, 16.45% lower permeability, and 1.01% reduced carbon emissions. • Demonstrates high robustness and accuracy (R 2 > 0.936), validated experimentally and via cross-dataset testing. Optimizing steam curing for precast concrete tunnel linings (PCTL) is vital for reconciling high performance with low carbon emissions (CE), yet existing methods lack holistic, interpretable multi-objective solutions. This study addresses this gap by proposing an enhanced, knowledge-driven hybrid framework integrating Bayesian Optimization (BO), CatBoost, and an improved NSGA-III, augmented with SHAP (SHapley Additive exPlanations) for full model interpretability. Firstly, energy conversion theory quantifies CE, while experimental data captures compressive strength (CS) and chloride ion permeability coefficient (CIPC). Secondly, BO optimizes CatBoost’s hyperparameters, establishing a highly accurate predictive model (R 2 > 0.936, training time: 61.1 s). Finally, an NSGA-III algorithm, initialized via chaotic mapping, performs multi-objective optimization (MOO) to optimize CS, CIPC, and CE. Crucially, SHAP analysis is embedded throughout to interpret feature importance and causal relationships. Our case study found that: (1) The framework identified optimal parameters (e.g., 60.5℃ thermostatic temperature) that increase 28-day CS by 8.55%, reduce CIPC by 16.45%, and lower CE by 1.01% versus baseline. (2) SHAP analysis revealed thermostatic temperature, heating rate, and duration as the most critical, interpretable drivers of performance and emissions. (3) Experimental validation confirmed prediction errors < 5%, and cross-dataset tests proved its robustness under material and environmental uncertainty. The implications are profound: this framework provides a validated, intelligent decision-support tool for the construction industry to achieve sustainable, high-performance concrete production. By embedding interpretability and robustness, it sets a new standard for data-driven, low-carbon process optimization in intelligent manufacturing.