Physics-Assisted Data-Driven Methods for Predicting the Performance of Deep Excavation in Soft Clays
Journal
International Journal of Geomechanics
Journal Volume
26
Journal Issue
7
Start Page
04026127
ISSN
15323641
Date Issued
2026-07-01
Author(s)
Abstract
This paper investigates the use of physics-assisted data-driven methods for predicting the performance (such as wall deformation and factor of safety against basal heave) of deep excavation in soft clay. Physics-assisted prediction methods with different levels of physical analysis assistance are investigated: the Level 1 method adopts multiple physical analysis parameters such as excavation dimension and support system stiffnesses as inputs; the Level 2 method relies on the wall deformation calculated by one-dimensional (1D) FEM as the main input; the Level 3 method relies on the wall deformation calculated by three-dimensional (3D) FEM as the main input. The degree of physical analysis assistance is minimal for the Level 1 method and maximal for the Level 3 method because the Level 1 method does not require FEM, but the Level 3 method requires 3D FEM. A new database comprising 100 clay sites is compiled to train the three data-driven methods, and their prediction capacities are compared using the extensive leave-one-out cross-validation method. The results show that the Level 3 method provides the most accurate wall deformation predictions, achieving the highest coefficient of determination (R2 = 0.73) and coefficient of efficiency (COE = 0.72) and the lowest mean absolute error (MAE = 0.42), with the narrowest 95% confidence intervals, whereas the Level 2 method exhibits weaker performance (R2 = 0.58; COE = 0.58; MAE = 0.51). Accordingly, wall deformation prediction should rely on the Level 3 method when 3D FEM is feasible. When 3D FEM is not viable, the Level 1 method (R2 = 0.61; COE = 0.61; MAE = 0.50) is preferred over the Level 2 method because FEM-free physical parameters are less complicated and more informative than 1D FEM outputs. For factor of safety prediction, the Level 1 method performs better than other methods due to multiple physical input parameters. Therefore, the Level 3 method is recommended for wall deformation prediction when 3D FEM is available, while the Level 1 method is recommended for factor of safety prediction and for practical applications when 3D FEM is unavailable.
Subjects
Data-driven geotechnics
Deep excavation
Factor of safety
Physics-assisted machine learning
Wall deformation
Publisher
American Society of Civil Engineers (ASCE)
Type
journal article