Slope stability machine learning predictions on spatially variable random fields with and without factor of safety calculations. (January 2023)

• Record Type: Journal Article
• Authors: Aminpour, Mohammad; Alaie, Reza; Khosravi, Sajjad; Kardani, Navid; Moridpour, Sara; Nazem, Majidreza
• Is Part Of: Computers and geotechnics. Volume 153(2023)
• Publication Date: 2023-01
• Subjects: Machine learning -- Slope stability -- Monte Carlo -- Surrogate models -- Anisotropy -- Heterogeneity; Engineering geology -- Data processing -- Periodicals; Soil mechanics -- Data processing -- Periodicals; Rock mechanics -- Data processing -- Periodicals; 624.1510285
• DOI: 10.1016/j.compgeo.2022.105094
• Languages: English
• ISSNs: 0266-352X

Abstract

Abstract: Random field Monte Carlo (MC) reliability analysis is a robust stochastic method to determine the probability of failure. This method, however, requires a large number of numerical simulations demanding high computational costs. This paper explores the efficiency of machine learning (ML) models and Artificial Neural Networks used as surrogate models trained on a limited number of random field slope stability simulations in predicting the results of large datasets. The paper explores the efficiency of the predictions on the probability of failure using databases with and without factor of safety (FOS) computations. An extensive range of soil heterogeneity and anisotropy is examined on unstratified and layered slopes. On datasets requiring only the examination of failure or non-failure class of slopes (without FOSs), the performance of ML classification of the random field slope stability results generally reduces with higher anisotropy and heterogeneity of the soil. However, using the probability summation method proposed here, ML prediction of the probability of failure is shown to be highly accurate for the whole range of soil heterogeneity and anisotropy. The errors in the predicted probability of failure using 5% of MC data is only 0.46% in average for the prediction of the remaining unseen 95% of data. Offering such accuracies, the approach accelerates the computations for about 100 folds. The models also proved similarly efficient in predicting FOSs for stratified random field anisotropic heterogenous slopes.