Abstract: Traditional experimental design methods struggle to quantitatively express the complex relationships between material component parameters and performance indicators for thixotropic slurry used in large-section rectangular pipe jacking construction, and thus fail to provide accurate optimization results. This study proposes a multi-objective optimization scheme for thixotropic slurry mix proportions based on machine learning prediction algorithms. First, traditional controlled-variable experiments combined with orthogonal experiments were conducted to obtain performance indicator data (including filtration loss, apparent viscosity, and filter cake thickness) corresponding to different material component ratios (bentonite, sodium carboxymethyl cellulose (CMC-Na), soda ash, and water). The sensitivity and influence patterns of different material dosages on slurry performance were investigated. Subsequently, two prediction models—Support Vector Regression (SVR) and eXtreme Gradient Boosting (XGBoost)—were compared using machine learning. The results indicated that the SVR algorithm offered higher accuracy and efficiency. Then, with the objectives of controlling apparent viscosity, minimizing filtration loss, and minimizing cost, the NSGA-Ⅱ algorithm was used to generate a Pareto optimal solution set. The Entropy Weight TOPSIS method was then applied to select the optimal mix proportion: bentonite 26.61%, CMC-Na 0.35%, and soda ash 0.06%. Finally, material re-testing and direct shear tests confirmed that the slurry performance indicators met the requirements, and a drag reduction rate of 42.9% was achieved, demonstrating a significant drag reduction effect. This intelligent mix proportion design method provides valuable technical guidance for related engineering applications.