This study developed an optimized neural-network framework for predicting the factor of safety (FOS) against soil liquefaction using shear wave velocity and relevant geotechnical and seismic parameters. Three models were evaluated: artificial neural network (ANN), ANN optimized with a genetic algorithm (ANN-GA), and particle swarm optimization (ANN-PSO). A dataset of 300 randomly generated cases was developed using seven input variables: stress-normalized shear wave velocity, soil deposit depth, groundwater table depth, fines content, bulk density, peak ground acceleration, and earthquake magnitude. Model performance was assessed using the coefficient of determination (R²), and root mean square error (RMSE). Among the investigated models, ANN-PSO exhibited the best predictive performance. During training, it achieved an R² = 0.998 and RMSE = 0.004, while the testing phase yielded an R² = 0.972 and RMSE = 0.010. Its predictive capability was further evaluated through rank analysis, regression analysis, reliability assessment, error-matrix evaluation, and Taylor diagrams. The proposed models were also compared with conventional liquefaction assessment indices, including the liquefaction potential index and liquefaction severity index. Overall, the results showed that ANN-PSO provides an accurate, efficient approach for estimating liquefaction-related FOS and has potential for rapid seismic risk assessment and practical geotechnical engineering applications.
This study investigated the static and seismic performance of cantilever retaining walls incorporating shear keys and relief shelves using three-dimensional finite element analysis. Four configurations were examined: a conventional wall, a wall with a shear key, a wall with a relief shelf, and a wall incorporating both modifications. Sandy backfill was modeled using the Mohr-Coulomb constitutive model, while concrete behavior was represented using the Concrete Damaged Plasticity model. The numerical model was validated against Coulomb’s analytical active earth-pressure solution, and seismic response was evaluated using the 1999 Chi-Chi earthquake record. Under static loading, the combined shear-key and relief-shelf configuration exhibited the lowest lateral displacement and surrounding-soil deformation, although localized stress and strain concentrations developed near the shelf-stem connection. Under seismic loading, the relief-shelf-only configuration exhibited the lowest pressure stress, shear strain, wall displacement, and surrounding-soil deformation. The results indicated that the effectiveness of geometrical modifications varied with loading conditions and highlighted the importance of evaluating retaining wall configurations under both static and seismic loading.
A systematic experimental method was developed to measure the angle of repose of cohesionless sand under plane strain and axisymmetric conditions. Two sands of different relative densities were tested using a planar tank and a pipe filled with sand to account for the plane strain and axisymmetric nature of the experimental investigation. Under axisymmetric conditions, the effects of diameter and lifting velocity of pipe, and relative density of sand on the measured angle of repose were investigated. The angle of repose measured under plane strain and axisymmetric conditions was compared with the friction angle obtained from direct shear tests. The relationships between the angle of repose and friction angle were established for the investigated sands and relative densities. The experimental results showed that the measured angle of repose varied with the testing configuration and relative density. Under axisymmetric conditions, the measured angle of repose was influenced by pipe diameter and pipe lifting velocity. Differences were observed between the plane strain angle of repose, axisymmetric angle of repose, and friction angle obtained from direct shear tests. The results provided an experimental basis for evaluating the relationship between angle of repose and friction angle of cohesionless sand under different testing conditions.
This study develops a hybrid soft-computing model for slope stability prediction and probabilistic risk assessment based on extreme gradient boosting (XGB) optimised by the Addax optimisation algorithm (AOA) and Hippopotamus optimisation (HO). A combined Latin hypercube sampling–Monte Carlo simulation (LHS-MCS) approach generated 10,000 synthetic slope cases, and the factor of safety (FOS) was calculated using the Bishop Simplified method. Subsequently, the XGB-AOA and XGB-HO models were used as surrogates to model the FOS of slopes. According to the results, the XGB-AOA model exhibited superior predictive performance with R² of 0.9727 in the testing phase. Then, the XGB-AOA-estimated FOS was used for probabilistic analysis. For each combination of the coefficient of variation (COV), fifty independent datasets were generated, and the First Order Second Moment method was adopted for estimating the reliability index and probability of failure (Pf). The probabilistic results indicate that with increasing COV and horizontal seismic coefficient, reliability index decreases and Pf increases. Explainable machine learning analysis also points to slope angle, height, and cohesion as important parameters controlling the FOS. The proposed framework provides an efficient way for FOS prediction, enabling quick probabilistic slope stability assessment.
This study developed and critically validated an explainable physics-guided machine-learning framework for predicting the ultimate axial compressive capacity ( Q_u1 ) of bored piles in layered cohesive–frictional soils. A database of 337 static load tests was collected across four states in Nigeria and was analysed using uncalibrated and non-negative least-squares (NNLS)-calibrated physics models, seven standalone machine-learning algorithms, a physics-guided residual model, and a stacking ensemble. All preprocessing, calibration, residual learning, and meta-learning operations were confined to the development data. Robustness was assessed using repeated stratified cross-validation, learning curves, split- and outlier-sensitivity analyses, and leave-one-stratigraphic-group-out validation. The results showed that NNLS calibration improved the alpha–beta model from R^2=-1.0278 to R^2=0.9154 . The physics-guided hybrid achieved the highest hold-out R^2 (0.9933), lowest root-mean-square error (165.9 kN), and lowest mean absolute error (95.2 kN). Its repeated-cross-validation mean R^2 was 0.9914, with a standard deviation of 0.0045. Transfer to excluded stratigraphic groups was strongly group-dependent. The hybrid provides accurate and auditable within-database prediction, supporting more reliable and resource-efficient foundation design for sustainable infrastructure development.
