A Surrogate Model for Uncertainty Quantification for the Reinforced Soil Footing Problem

Surrogate modeling, also known as metamodeling, has several applications in the analysis of reinforced soil structures. Reinforced soil design often involves dealing with uncertainties related to material properties and project conditions. Surrogate models can facilitate probabilistic analysis by allowing the rapid evaluation of different scenarios. This helps in assessing the reliability and safety of a candidate design under uncertain conditions. In recent times, there has been an increased emphasis on understanding the influence of soil spatial variation on soil-structure interactions in geotechnical engineering problems. Traditional probabilistic theories often fall short in accurately representing spatial variability due to computational demand. As a result, there is growing interest in stochastic numerical modeling by generating a large number of realizations of soil parameters and incorporating them into numerical analyses to predict stability outcomes. However, this method is computationally demanding, especially using Monte Carlo simulations to predict uncertainty and probability of failure. To confidently estimate the probability of failure, or reliability index, in reliability-based analysis and design, a sufficient number of realizations is necessary. However, this is often impractical due to computational demand. A solution to reduce the number of required Monte Carlo simulations to achieve a confident estimate of reliability index is to use artificial intelligence, particularly artificial neural networks (ANNs). This study adopts an efficient ANN algorithm, specifically the Radial Basis Function (RBF), to predict the ultimate bearing capacity of a shallow footing placed on a geogrid-reinforced granular fill overlying a very soft clay deposit using a large synthetic database of footing load-settlement results previously reported by the writers. Assuming the foundation soil is spatially variable, the goal was to forecast the associated ultimate load outcomes across numerous realizations. The study underscores the potential of surrogate modeling to enhance the efficiency and accuracy of stochastic soil-structure interaction modeling.

Keywords: Artificial Neural Network (ANN), Surrogate modeling, Radial Basis Function (RBF), Soil variability, Reliability-based design, Reinforced soil footing.