Understanding slope stability is crucial for effective risk management and prevention of slides. Some deterministic approaches based on limit-equilibrium and numerical methods have been proposed for the assessment of the safety factor (SF) for a given soil slope. However, for risk analyses of slides of earth dams, a range of SFs is required due to uncertainties associated with soil strength properties as well as slope geometry. Recently, several studies have demonstrated the efficiency of artificial neural network (ANN) models in predicting the SF of natural and artificial slopes. Nevertheless, such techniques operate as black-box models, prioritizing predictive accuracy without suitable interpretability. Alternatively, multivariate polynomial regression (MVR) models offer a pragmatic interpretability strategy by combining the analysis of variance with a response surface methodology. This approach overcomes the difficulties associated with the interpretability of the black-box models, but results in limited accuracy when the relationship between independent and dependent variables is highly nonlinear. In this study, two models for a quick assessment of slope SF in earth dams are proposed considering the MVR and the ANN models. Initially, a synthetic dataset was generated considering different soil properties and slope geometries. Then, both models were evaluated and compared using unseen data. The results are also discussed from a geotechnical point of view, showing the impact of each input parameter on the assessment of the SF. Finally, the accuracy of both models was measured and compared using a real-case database. The obtained accuracy was 78% for the ANN model and 72% for the MVR one, demonstrating a great performance for both proposed models. The efficacy of the ANN model was also observed through its capacity to reduce false negatives (a stable prediction when it is not), resulting in a model more favorable to safety assessment.

This study utilizes a combined approach of Finite Element Method (FEM) simulation and Artificial Neural Network (ANN) modeling to analyze and predict the load-displacement relationship of bored piles in clayey sand. FEM is applied to simulate the nonlinear relationship between load and vertical displacement, with input parameters including load and the mechanical properties of the soil. The results obtained from FEM are used as input data for the ANN model, enabling accurate predictions of vertical displacement based on these parameters. The findings of this study show that the predicted ultimate bearing capacity of the bored piles is highly accurate, with negligible error when compared to field experiments. The ANN model achieved a high level of accuracy, as reflected by an R2 value of 0.9992, demonstrating the feasibility of applying machine learning in pile load analysis. This research provides a novel, efficient, and feasible approach for analyzing and predicting the bearing capacity of bored piles, while also paving the way for the application of machine learning in geotechnical engineering and foundation design. The integration of FEM and ANN not only minimizes errors compared to traditional methods but also significantly reduces time and costs when compared to field experiments.

Physics-Informed Neural Networks (PINNs) have shown considerable potential in solving both forward and inverse problems governed by partial differential equations (PDEs) for a wide range of practical applications. This study leverages PINNs for modeling nonlinear large-strain consolidation of soft soil, including creep behavior. The inherent material and geometric nonlinearities associated with soft soil consolidation pose challenges for PINNs, including precision and computational efficiency. To address these issues, we introduce self-adaptive physics-informed neural networks (SA-PINNs), featuring an adaptive loss function weighting and a slope scaling method for the activation functions. Additionally, a sensitivity analysis exploring the influence of monitoring data on the parameter inversion accuracy is presented. Two engineering case studies are used to benchmark the settlement prediction capabilities of the present SA-PINN method with traditional techniques, demonstrating the superior prediction accuracy and consistency of the SA-PINN approach. The findings highlight the significant potential of SA-PINN in practical geotechnical engineering problems.

Physics-informed neural networks (PINNs) are increasingly employed for surrogate modelling of soil behaviour. Existing surrogate models for unsaturated soil only account for seepage in rigid soil, neglecting the complex coupling between deformation and seepage in unsaturated soil. This study develops a new surrogate model for hydro-mechanical coupling in unsaturated soil using the PINN approach. Dimensionless governing equations, including mass balance and force balance equations, are derived and adopted for physical constraints. With absence of explicit constitutive relations, this new surrogate model utilises sparse measured data to identify pore water pressure, effective stress and deformation in unsaturated soil. Separate neural networks are employed to facilitate efficient back-propagation for coupled problem involving multiple outputs. The newly developed model is then applied to simulate two cases with sparse measurements in unsaturated soil. The results illustrate that the newly developed surrogate model successfully learns the elasto-plastic constitutive relation of suction-induced volume change from experimental data. Meanwhile, model predictions regarding both water flow and stress distribution align within the 95 % confidence interval of theoretical values, demonstrating interpretability of PINN model. Furthermore, by adhering to physical constraints, the relative error in predicting soil deformation from neural networks significantly reduces from 49 % to less than 10 %. These findings suggest PINN model with separate networks is capable to simulate unsaturated soil considering both deformation and seepage, even with sparse measured data and incomplete physical constraints.

More attention has been paid to integrating existing knowledge with data to understand the complex mechanical behaviour of geomaterials, but it incurs scepticism and criticism on its generalizability and robustness. Moreover, a common mistake in current data-driven modelling frameworks is that history internal state variables and stress are known upfront and taken as inputs, which violates reality, overestimates model accuracy and cannot be applied to modelling experimental data. To bypass these limitations, thermodynamically consistent hierarchical learning (t-PiNet) with iterative computation is tailored for identifying constitutive relations with applications to geomaterials. This hierarchical structure includes a recurrent neural network to identify internal state variables, followed by using a feedforward neural network to predict Helmholtz free energy, which can further derive dissipated energy and stress. The thermodynamic consistency of t-PiNet is comprehensively validated on the synthetic data generated by von Mises and modified Cam-clay models. Subsequently, the potential of t-PiNet in practice is confirmed by applying it to experiments on kaolin clay. The results indicate neural networks embedded by thermodynamics perform better on the loading space beyond the training data compared with the conventional pure neural network-based modelling method. t-PiNet not only offers a way to identify the mechanical behaviour of materials from experiments but also ensures it is further integrated with numerical methods for simulating engineering-scale problems.

The number of studies concerning the shear strength of resedimented alluvial soils is extremely limited compared to the studies conducted on fine-grained marine sediments, since alluvial soils are generally tested in remolded or reconstituted state especially in the studies investigating their liquefaction potential. In this study, estimation models were developed to predict cohesion (c) and internal friction angle (phi) parameters of a fine-grained alluvial soil using resedimented samples. A total of 60 undisturbed soil samples were obtained from Bafra district of Samsun province (Turkiye) by core drilling. A cone penetration test with pore water pressure measurement (CPTu) was also carried out alongside each borehole to determine the over-consolidation ratios of the samples. Physical-index property determinations and triaxial tests were conducted on the undisturbed samples. 20 sample sets were created with known physical, index, and strength characteristics. The samples are classified as CH, CL, MH, and ML according to the Unified Soil Classification System, with liquid and plastic limits ranging from 31.6-75% and 19.3 to 33.6% respectively. The c and phi values of the samples varied from 4.1 to 46.1 kPa and 26 to 35 degrees respectively. The samples were then resedimented in the laboratory under conditions reflecting their original in-situ properties, and triaxial tests were repeated. The c and phi values of the resedimented samples ranged from 5.3 to 24.5 kPa and 28 to 32 degrees respectively. The results indicate that the c values of the resedimented samples are generally lower than those of the undisturbed samples, whereas upper and lower bounds for phi values are similar. Multivariate regression analyses (MVR) were utilized to develop estimation models for predicting c and phi using strength and physical properties of 20 soil samples as independent variables. Three estimation models with R-2 values varying between 0.723 and 0.797 were proposed for c and phi which are statistically significant for p <= 0.05. Using artificial neural networks (ANN), the estimation models developed by MVR were replicated to validate the models. ANN yielded very similar results to the MVR, where the R-2 values for the correlations between c and phi values predicted by both methods varied from 0.852 to 0.955. The results indicate that c and phi values of undisturbed samples can be estimated with acceptable accuracy by determining basic physical and index properties of the disturbed samples and shear strength parameters of the resedimented samples. This approach, which enables the reuse of disturbed soil samples, can be used when undisturbed soil samples cannot be obtained from the field due to economic, logistical, or other reasons. Further research on the shear strength parameters of resedimented alluvial soils is needed to validate the estimation models developed in this study and enhance their applicability to a wider range of alluvial soils.

This study investigated the stabilization of fine-grained soil from the Indo-Gangetic plain using nano-silica (NS) and predicted the unconfined compressive strength (UCS) using advanced machine learning techniques. Experimental investigations were conducted on 118 UCS samples with NS contents varying from 0.5 to 4%. The results showed significant improvements in the soil plasticity, compaction characteristics, and UCS with NS incorporation. NS acted as a reinforcing agent, filling void spaces and improving interlocking between soil particles, leading to increased maximum dry density, reduced optimum moisture content, and notable improvements in the UCS. Microstructure analysis revealed the positive impact of NS on soil properties, attributed to enhanced durability, reduced swell strains, and improved strength due to the synergistic effects of NS particles. Furthermore, five innovative hybridized models based on artificial neural networks (ANN) and nature-inspired optimization algorithms were developed to predict the UCS of NS-stabilized fine-grained soils. The models demonstrated high accuracy, with R2 values exceeding 0.96 and 0.89 for the training and testing dataset. The ANN-Firefly algorithm (ANN-FF) model emerged as the most proficient predictor. This study highlights the importance of input parameters in model development and suggests that further research should focus on expanding experimental data to enhance model flexibility. The proposed approach offers significant implications for cost and time savings in experimental sample preparation and demonstrates the high capability of ANN to determine optimal values for soil stabilization techniques in the Indo-Gangetic plains.

Precise and rapid simulation of a material's mechanical response is crucial in engineering. Conventional numerical schemes, such as finite element methods, face computational hurdles due to the intricate analysis required for path-dependent elastoplastic behavior. Accurately computing mechanical behavior under continuous loading necessitates tracking yield surface evolution through iterative mapping algorithms. This study introduces a short sequence machine learning approach to quickly predict the constitutive behavior of sandy materials and obtain the mechanical response of engineering materials under continuous loading. Initially, advanced constitutive models and their variants are employed to synthesize a dataset, accounting for variations in intrinsic features. The performance of various machine learning model frameworks is evaluated through mean absolute error percentage and maximum error percentage. The findings demonstrate that an incremental strategy machine learning constitutive model showed poor performance in predicting the mechanical behavior of granular materials. However, the full sequence strategy using Multilayer Perceptron(MLP) and long short-term memory(LSTM) machine learning models demonstrates the ability to learn and rapidly predict the irreversible, history-dependent phenomena in sandy materials. Notably, LSTM performs optimally when the time step is 4. This work offers valuable insights for enhancing the computational efficiency of numerical schemes.

Many experiments and computational techniques have been employed to explain the mechanical properties of frozen soils. Nevertheless, due to the substantial complexity of their responses, modeling the stress-strain characteristics of frozen soils remains challenging. In this study, artificial neural networks (ANNs) were employed for modeling the mechanical behavior of frozen soil, while different testing strategies were carried out. A database covering stress-strain data from frozen sandy soil subjected to varying temperatures and confining pressures, resulting from triaxial tests, was compiled and employed to train the model. Subsequently, different artificial neural networks were trained and developed to estimate the deviatoric stress and volumetric strain, while temperature, axial strain, and confining pressure were considered as the main input variables. Based on the findings, it can be indicated that the models effectively predict the stress-strain behavior of frozen soil with a significant level of accuracy.

An approach based on a Physics-Informed Neural Network (PINN) is introduced to tackle the two-dimensional (2D) rheological consolidation problem in the soil surrounding twin tunnels with different cross-sections, under exponentially time-growing drainage boundary. The rheological properties of the soil are modelled using a generalized viscoelastic Voigt model. An enhanced PINN-based solution is proposed to overcome the limitation of traditional PINNs in solving integral-differential equations (IDEs) equations. In particular, two key elements are introduced. First, a normalization method is employed for the spatio-temporal coordinates, to convert the IDEs governing the consolidation problem into conditions characterized by unit-duration time and unit-area geometric domain. Second, a conversion method for integral operators containing function derivatives is devised to further transform the IDEs into a set of second-order constant-coefficient homogeneous linear partial differential equations (PDEs). By using the TensorFlow framework, a series of PINN-based models is developed, incorporating the residual adaptive sampling method to address the 2D consolidation equations of soft soils surrounding tunnels with different burial depths and cross-sections. Comparative analyses between the PINNbased solutions, and either finite element or analytical solutions highlight that the aforementioned normalization stage empowers PINNs to solve the PDEs across different spatial and temporal scales. The integral operator transformation method facilitates the utilization of PINNs for solving intricate IDEs.