The parameters of geotechnical are randomly distributed in space. To better reflect the actual engineering geological conditions, considering the uncertainty of soil in the study of foundation bearing capacity, establishing a bearing capacity prediction model holds significant engineering value. Incorporating spatial variability of geotechnical parameters based on random field theory into the study of coupling beam pile foundation, a two-dimensional random finite element model is established using numerical methods to analyze the bearing capacity of coupling beam pile foundation and pile foundation, validated against model test results. Subsequently, a convolutional neural network is employed to establish a model between the random field images of soil parameters and the ultimate bearing capacity of foundations for bearing capacity prediction, and the impact of different parameters is studied based on the prediction model. The results indicate that considering the spatial variability of soil, the foundation′s bearing capacity is in basic agreement with experimental results, with random results consistently higher than deterministic analysis. Under random conditions, both coupling beam pile and pile foundations exhibit normally distributed bearing capacities. The accuracy of the bearing capacity prediction model established using convolutional neural networks is high and can be utilized for parameter analysis. The bearing capacity of foundations increases with increasing soil parameters and decreases with increasing coefficient of variation. Under random conditions, the bearing capacity of coupling beam pile foundation is higher than that of pile foundation, effectively leveraging soil strength to withstand bearing capacity loss caused by parameter uncertainties.