Abstract: The current method of land reclamation is still incomplete. Due to the insufficient understanding of the nature of the silt in the reclamation area by engineering personnel， the design parameters are not accurate enough， making it very easy for problems to occur during the construction process. Therefore， the study first modeled the foundation excavation in the reclamation area and simulated the finite element calculation of the excavation of the foundation excavation. Then the back propagation neural network （BPNN） model was used to establish the nonlinear functional relations between coating parameters and deformation values. Through dynamic back analysis， the BPNN model was used for dynamic back analysis. By utilizing the nonlinear relations between the mechanical parameters in the soil layer and the deformation value of the foundation excavation， four sets of silt layer parameters at different time periods were obtained. The BPNN prediction model was optimized by using Genetic Algorithm （GA）. During the research process， it was found that the structure and number of nodes contained in the hidden layer of the prediction model were directly related to their learning ability. For this purpose， the study used the least squares method to fit and calculate the relevant data， and obtained the optimal combination of hidden layer structures. From this， a deformation prediction model for underground foundation excavation based on silt layer parameters was obtained. Through experiments， the total absolute error of the model was 20.46 mm， and the average prediction accuracy was 98.48%. The deformation patterns of foundation excavation could be accurately predicted based on the parameters obtained from back analysis.