Based on the salvage project of the Yangtze Estuary No. 2 shipwreck, this study investigates an intelligent decision-making method that integrates physical priors and data-driven models. The aim of the study is to improve the accuracy of thrust prediction in variable-depth underwater curved pipe jacking and solve the over-reliance on manual experience for tunneling parameter optimization under multiple complex constraints. A series of 2 698 sets of valid construction data, including key parameters such as penetration rate and field penetration index, are extracted, and a standard dataset is constructed using a hybrid outlier cleaning strategy. Random forest (RF) and genetic algorithm-neural network (GA-NN) are used to build multi-input, single-output soil-machine mapping models for thrust prediction, and their generalization abilities are evaluated via fivefold cross-validation. Taking the safety control ranges of thrust (≤4 000 kN）and cutterhead rotational speed as physical constraint boundaries, and maximizing the advance rate as the objective, the dung beetle optimizer (DBO) is introduced for global parameter optimization. For thrust prediction, the RF mapping model achieves a mean absolute error (E_MA) of 68.41 kN and an R²of 0.912 on the testing set, notably outperforming GA-NN (E_MA=86.12 kN) and traditional theoretical models. These results demonstrate the superiority of ensemble learning in capturing nonlinear features in underwater composite strata. For parameter optimization, the hybrid intelligent model (RF+DBO) successfully increases the mean advance rate of the test samples from the actual 30.44 mm/min to the design upper limit of 35.00 mm/min without exceeding the safety thrust limit, achieving an average efficiency improvement of 14.99%. Influenced by stratum mechanics, the optimization improvement space in the soft bluish-gray mud stratum (~17.2%) is markedly higher than that in the hard iron-sand stratum (~11.3%). This physics-data-driven model accurately depicts the nonlinear dynamic response laws of variable-depth curved pipe jacking. This method corrects the conservatism of manual operations through intelligent optimization under extremely high environmental safety standards, thereby achieving an optimal tradeoff between safety and efficiency.