Apart from spatial variability of soil properties, a geotechnical system can have many other sources of uncertainties. To efficiently analyze such a system in a probabilistic manner, many strategies have been proposed in the literature. This paper presents a deep learning technique for an efficient geotechnical reliability analysis with multiple uncertainties. The proposed method involves using convolutional neural networks (CNNs) as metamodels of the physics-based simulation model of a geotechnical system. In the present study, the spatially variable soil properties and the external loads are simultaneously considered in the analysis of a geotechnical system. The proposed neural network method configures these uncertainties to form a multi-channel “image.” CNNs can then simultaneously learn high-level features that contain information about the multiple uncertainties. With an appropriate architecture and adequate training, the trained CNNs can replace the computationally demanding physics-based simulation model for Monte Carlo simulations. Application of the neural network method is illustrated using a synthetic geotechnical example. The results reveal that the proposed neural network method effectively handles multiple uncertainties and efficiently predicts a failure probability value that is in good agreement with the benchmark result obtained using direct Monte Carlo simulations.

In recent years, the random field finite-element method (FEM) has been used increasingly in geotechnical engineering to carry out analyses that account for the inherent spatial variability in the physical and mechanical properties of both natural and treated soils. However, this method, which usually is performed in tandem with Monte Carlo simulation (MCS), requires significantly greater computational resources than deterministic finite-element analysis. Metamodeling is one of the techniques commonly adopted to alleviate the computational burden. This paper proposes a novel and computationally efficient metamodeling technique that involves the use of convolutional neural networks (CNNs) to perform random field finite-element analyses. CNNs, which treat random fields as images, are capable of outputting FEM predicted quantities with learned high-level features that contain information about the random variabilities in both spatial distribution and intensity. CNNs, after being trained with sufficient random field samples, could be used as a metamodel to replace the expensive random field finite-element simulations for all subsequent calculations. The validity of the proposed approach was illustrated using a synthetic excavation problem and a synthetic surface footing problem. The good agreement between the CNN outputs and the FEM predictions demonstrated the promising potential of using CNNs as a metamodel for reliability analysis in spatially variable soils.

Understanding the variation of geological interfaces plays a crucial role in the analysis and design of infrastructure systems. Generally, there are two classes of techniques for predicting geological interfaces, for example, interpolation/regression-based techniques and machine-learning-based techniques. In this paper, a Multi-scale Meta-learning Model (M³) methodology is proposed. The new methodology improves the current state-of-the-art techniques by fusing two levels of information: (i) generic characteristics of the sampling locations, for example, coordinates, and (ii) location-specific characteristics, for example, local-scale predictions. The implementation starts from using an array of classic interpolation/regression-based techniques as base learners to provide first-level predictions at a local scale. These predictions are then combined with generic characteristics to train a meta-learner following the stacking ensemble learning framework. In this manner, the location-specific information from the base learners can be simultaneously considered with the generic information in the training process. The variation of rockhead elevation is predicted using the M³ methodology and a comprehensive borehole dataset in Singapore. A detailed comparative study involving several existing methods is also carried out to rigorously validate the M³ methodology. The results show that the M³ methodology achieves 20% improvement in the model performance compared to existing methods, indicating its promising potential in geotechnical site characterization.