<p>Deep excavations in urban environments require accurate prediction of lateral wall movements, particularly under spatially variable soil conditions. A Finite Element Method (FEM) often relies on deterministic soil parameters and therefore struggles to capture the inherent uncertainty in natural soil deposits. While the Random Finite Element Method (RFEM) improves realism by incorporating soil spatial variability, its reliance on large-scale Monte Carlo simulations (MCSs) results leads to high computational cost. This study proposes an enhanced computational framework that integrates RFEM with a Convolutional Neural Networks (CNNs) surrogate model to efficiently predict lateral wall movements in deep excavations. The stiffness modulus (E₅₀^ref) of the medium dense sand layer is modelled as a two-dimensional random field. RFEM simulations are performed for serious realizations to generate training, validation, and testing datasets for the CNN model. The framework is applied to a real deep excavation project, the Madison Project in Ho Chi Minh City, Vietnam, and predictions are compared against 2D FEM simulations and field inclinometer measurements. The trained CNN achieves high predictive accuracy (R² &gt; 0.85 across datasets) in estimating both maximum and depth-dependent lateral wall movements, while reducing computation time by more than an order of magnitude compared with RFEM. The CNN model also reliably reproduces the probability of required failure curves associated with excessive lateral wall movements, demonstrating strong potential for stability assessment. The results confirm that the proposed CNN-based framework provides an efficient and accurate tool for deformation prediction and reliability evaluation in complex deep excavations with spatially variable soil properties.</p>