Abstract: To ensure the safe operation of existing railways and the successful construction of new tunnels during shield undercrossing, a small sample data-driven optimization method combining numerical simulation and ma? chine learning is proposed for construction parameter optimization. First, a refined numerical model is established based on an actual project, and orthogonal experiments are conducted to analyze the impact of shield construction parameters such as support pressure, cutterhead face friction torque, grouting pressure, and shield shell friction on the deformation of the existing tunnel. Then, the main influencing factors (support pressure and grouting pressure)are selected as variables, and Latin Hypercube Sampling is used to generate representative working condition data.A surrogate model is constructed based on the corresponding numerical simulation results. Finally, the surrogate model is used for ergodic prediction of working conditions in running tunnel to determine the optimal combination of shield construction parameters. The results show that excavation face support pressure and grouting pressure are the key factors affecting the deformation of the existing tunnel. The numerical simulated working conditions obtained through optimized Latin Hypercube Sampling are highly representative, and the constructed surrogate model meets the engineering requirements in terms of prediction accuracy. Field practice shows that the optimized construction parameters not only effectively suppress the deformation of the existing tunnel but also significantly reduce computational costs, demonstrating higher efficiency and practicality compared to traditional methods.