<p>The multi-objective mechanical property optimization of steel products represents a critical challenge in enhancing both product reliability and economic efficiency. Conventional trial-and-error approaches and physical models suffer from inherent limitations, including low efficiency, high costs, and difficulties in balancing multiple performance targets. This study proposes a novel data-driven inverse design framework that integrates Just-In-Time Learning (JITL) dynamic modeling, Light Gradient Boosting Machine (LGBM), and Multi-objective Bayesian optimization (MOBO) to achieve synergistic optimization of multi-objective mechanical properties and manufacturing costs in hot-rolled strip. First, relevant samples with similar mechanical properties to the target are selected to construct local models using seven distinct machine learning algorithms. Comprehensive model evaluation demonstrates LGBM's superior performance, achieving R² values of 0.95 (tensile strength), 0.92 (yield strength), and 0.87 (elongation) through feature engineering incorporating Mantel correlation analysis and recursive feature elimination. Subsequently, the integration of LGBM with MOBO generates a Pareto front for multi-objective optimization. Finally, the TOPSIS method was used for dual-criteria screening (performance and cost), identifying optimal manufacturing conditions that meet the target mechanical properties while minimizing production costs. SHAP value analysis further elucidates the mechanistic relationships between input features and mechanical properties. Experimental validation confirms that the optimized manufacturing parameters not only enhance steel's mechanical properties and stability but also verify the model's predictive accuracy and interpretability. This work establishes an innovative framework for intelligent design and cost-effective manufacturing in addressing complex multi-objective optimization challenges within the steel industry.</p>