<p>The rapid expansion of underground infrastructure has intensified the demand for dynamic and accurate risk prediction in deep foundation pit engineering, particularly in soft soil environments. To address this, a data–physics-fused deep learning (DPF-DL) framework that integrates parameter back analysis with spatial variability characterization is developed to improve the predictive modeling of excavation-induced risks in soft soil excavations. More specifically, DPF-DL employs a convolutional neural network (CNN)-based surrogate model to replace the computationally expensive finite element model (FEM). Physical consistency is ensured through a genetic algorithm (GA)-based back analysis, augmented with spatial variability to improve soil behavior modeling and parameter calibration. These physically grounded components are subsequently fused with field monitoring data via a transformer model, facilitating dynamic and data-driven risk prediction. The proposed DPF-DL framework is validated in a representative soft soil project: the Shanghai Metro Line 9 Lantian Road Station deep foundation pit project. Results show that the CNN-based surrogate model reproduces retaining wall deformation with high accuracy (<i>R</i>² = 0.9406, <i>MSE</i> = 0.8793) and achieves prediction times under 1 s. GA-based back analysis further enhances parameter estimation by calibrating key soil properties, reducing the <i>MSE</i> from 44.208 to 4.428. Notably, the incorporation of spatial variability significantly improves the robustness of the GA optimization during early excavation stages, as indicated by an increase in the average optimal fitness from 0.6448 to 0.8115. The transformer model further delivers robust hazard state forecasts on retaining walls (<i>R</i>² = 0.9337, <i>MSE</i> = 0.00128). In summary, DPF-DL establishes a physically consistent and computationally efficient paradigm for real-time and data-informed risk management in complex subterranean environments.</p>