Abstract: Accurate prediction of mechanical properties in thermo-mechanical control process (TMCP) X80 pipeline steels is essential for process design and quality control. To meet this need, an interpretable multi-target prediction framework is developed for TMCP X80 pipeline steels. The framework simultaneously predicts yield strength (YS), tensile strength (TS), elongation (EL), and drop-weight tear test (DWTT) percent shear area in production settings. A layered stacked ensemble is adopted to learn shared representations, and a dedicated meta-learner is assigned to each mechanical property, thereby separating tasks and enabling selective information sharing. The non-dominated sorting genetic algorithm III (NSGA-III) multi-objective evolutionary algorithm is employed to jointly optimize hyperparameters and to construct a Pareto front that reflects different process trade-offs, while explainable artificial intelligence is used to provide insight at both global and instance levels. The proposed framework is evaluated on real production data, and superior predictive performance over multi-output baseline models is demonstrated. Under the NSGA-III compromise solution, the root-mean-square errors for YS, TS, EL, and DWTT are reduced to 6.69 MPa, 5.17 MPa, 1.31%, and 1.28%, corresponding to relative reductions of about 19.5%, 17.3%, 13.9%, and 18.1%, respectively. Shapley additive explanations based interpretability analysis further clarifies the influence mechanisms of key thermal history and cooling parameters on each mechanical property and enhances the industrial credibility of the framework.