<p>The era of big data has profoundly transformed mechanics research, with data-driven approaches playing a vital role in modeling and optimization. This study focuses on tunnel boring machine (TBM), where the thrust-torque ratio is a key determinant of their tunneling energy efficiency. However, due to the complexity of experiments and the testing requirements, obtaining sufficient high-quality data under varying geological conditions remains a major challenge in optimizing the tunneling energy efficiency of TBM. To address this, multi-cutter rotary cutting machine experiments and numerical simulations were conducted on 22 different rock types. Comprehensive datasets of normal and rolling forces were systematically collected. Using specific energy (SE) as the rock-breaking efficiency metric, we integrated physical and numerical data through a CatBoost-based fusion framework. The predictive model was initially trained on simulation data to capture the relationships among penetration, uniaxial compressive strength, tensile strength, and SE, and was subsequently fine-tuned with experimental data to develop the final fused model. Compared to models trained solely on experimental or simulated data, the fused model reduced RMSE by 37.1% and 58.6%, respectively, and improved R² by 19.0% and 44.6%, thereby enhancing both prediction accuracy and generalization capability. Furthermore, Bayesian optimization was employed to minimize SE and identify the optimal penetration. The results indicate that as rock strength increases, the optimal penetration decreases, while the corresponding minimal SE increases. These findings provide theoretical and engineering insights for improving TBM energy efficiency and parameter optimization, while establishing a robust data fusion framework for mechanical data analysis.</p>