In the service period of an immersed tunnel, large differential settlement may lead to serious safety problems. The settlement prediction of an immersed tunnel faces several uncertainties, including the complex stratigraphic distribution problem, errors of measurement, back-silting and dredging, etc. In this study, based on the simplification of taking a tunnel as a multi-beam model, a physics-informed machine learning method was proposed to analyse the foundation modulus distribution of the tunnel and predict its settlement. A case study with the designed parameters showed that the proposed method can overcome 20% of noise while using at least three groups of settlement data as training set. Furthermore, with seven to ten groups of training data, accurate and stable results can be obtained. For the uncertainties in the problem, the Markov Chain Monte Carlo algorithm and Bayesian theory were introduced to calculate the credible interval of the prediction, which further improves the accuracy of the analysis. The data analysis on the immersed tunnel of the Hong Kong–Zhuhai–Macao Bridge proved that the proposed method is applicable to practical engineering.