With the accelerating development of urban underground space, caissons, as a key construction technology, are widely applied under complex geological and restricted site conditions. However, traditional caisson construction highly relies on manual experience for attitude control, leading to low observation accuracy, lagging response, and high safety risks. To solve these challenges, this paper proposes a novel method for real-time prediction and decision support for caisson sinking posture by integrating numerical simulation with a neural network-based surrogate model. Firstly, a parametric finite element model of the sinking process of a circular caisson was established. By screening the key excavation parameters as input variables and taking the displacement and inclination of the key points of the caisson as output, a series of training samples were generated. Secondly, a neural network surrogate model was constructed and trained to efficiently replace the time-consuming finite element calculation. Finally, the accuracy and reliability of the established model were verified through the measured engineering data. The results show that this surrogate model can predict the sinking posture of the caisson quickly and accurately, significantly shorten the time required for traditional numerical analysis, achieve real-time feedback and decision support for the construction process, and provide effective guidance for ensuring the safety of caisson construction.