A Deep Learning–Based Approach for the Evaluation of Tunneling-Induced Ground Surface Settlement Considering Spatial Variability of the Soil

Tunneling-induced ground surface settlement (GSS) is a critical concern in urban areas as it can lead to differential settlement of surface constructions and cracking of underground pipelines. While numerical methods and machine learning approaches are popular for evaluating GSS, they face challenges such as high computational cost and limited generalization ability. To address these issues, this paper proposes a pre-trained deep learning model for efficient and accurate prediction of the maximum GSS (Smax) in soils with spatial variability. This model employs an artificial neural network (ANN) to identify Smax as the output, using random fields of soil parameters, tunnel geometries, and tunnel convergence parameters as inputs. The mean absolute percentage error (MAPE) and coefficient of determination (R2) of the proposed ANN model on the testing dataset are 5.3 percent and 99.2 percent, respectively, indicating the robustness and reliability of the model. It accelerates the GSS analysis process and can be applied to various tunnel projects under different geotechnical conditions. Specifically, there are three innovations in this paper. First, a new technique is developed that allows for the use of nonuniform mesh in data acquisition. Second, the spatial variability of soil is considered when analyzing tunneling-induced GSS problem, with these random fields subsequently used as inputs for deep learning. Third, by integrating numerical simulations and deep learning, a model for predicting Smax is developed, which is applicable across different tunnel geometries under diverse geotechnical conditions. The model is trained on a large dataset comprising 45,000 simulations, covering common tunnel depths, tunnel diameters, soil properties, and their spatial variability. In this case, it is ready to be used without additional numerical simulations. To verify the effectiveness of the developed model, it is applied to real-world tunnel projects. The results demonstrate that the predicted Smax values are highly consistent with field measurements, while the discrete distribution of Smax due to random fields provides the probability analysis.