[Paper Review] Slope stability predictions on spatially variable random fields using machine learning surrogate models

Mohammad Aminpour, Reza Alaie+3 more|arXiv (Cornell University)|Apr 4, 2022

Geotechnical Engineering and Analysis4 citations

TL;DR

This paper proposes a machine learning (ML) surrogate modeling approach to accelerate slope stability analysis on spatially variable random fields, using only 0.47% of Monte Carlo simulation data (500 samples) to train models that predict failure or non-failure with 85% accuracy and 91% AUC. The method reduces computational time from 306 days to under 6 hours while maintaining high reliability, with Random Forest and Bagging Ensemble showing superior performance among tested algorithms.

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 different machine learning (ML) algorithms used as surrogate models trained on a limited number of random field slope stability simulations in predicting the results of large datasets. The MC data in this paper require only the examination of failure or non-failure, circumventing the time-consuming calculation of factors of safety. An extensive dataset is generated, consisting of 120,000 finite difference MC slope stability simulations incorporating different levels of soil heterogeneity and anisotropy. The Bagging Ensemble, Random Forest and Support Vector classifiers are found to be the superior models for this problem amongst 9 different models and ensemble classifiers. Trained only on 0.47% of data (500 samples), the ML model can classify the entire 120,000 samples with an accuracy of %85 and AUC score of %91. The performance of ML methods in classifying the random field slope stability results generally reduces with higher anisotropy and heterogeneity of soil. The ML assisted MC reliability analysis proves a robust stochastic method where errors in the predicted probability of failure using %5 of MC data is only %0.46 in average. The approach reduced the computational time from 306 days to less than 6 hours.

Motivation & Objective

• To reduce the high computational cost of random field Monte Carlo (MC) reliability analysis for slope stability.
• To evaluate the effectiveness of various machine learning models as surrogate predictors for slope failure in spatially variable soil fields.
• To determine the minimum training data size required for accurate ML-based prediction of failure probability.
• To assess model performance under varying levels of soil heterogeneity and anisotropy.
• To validate the ML-assisted MC method by comparing predicted failure probabilities with reference results.

Proposed method

• Generate a large synthetic dataset of 120,000 finite difference simulations for slope stability under spatially variable soil properties.
• Define failure as binary outcome (failure or non-failure), avoiding time-intensive factor of safety calculations.
• Train nine different machine learning classifiers—including Random Forest, Bagging Ensemble, and Support Vector Machines—on a small subset (500 samples) of the dataset.
• Use cross-validation and hyperparameter tuning to optimize model performance on unseen data.
• Apply the trained ML models to classify the entire 120,000-sample dataset, significantly reducing computational demand.
• Validate the ML-assisted MC method by comparing predicted failure probabilities with reference results using 5% of MC data.

Experimental results

Research questions

• RQ1 Can machine learning surrogate models accurately predict slope failure on spatially variable random fields using only a small fraction of full Monte Carlo simulations?
• RQ2 Which machine learning algorithms perform best for classifying slope stability outcomes in heterogeneous and anisotropic soil fields?
• RQ3 How does increasing soil heterogeneity and anisotropy affect the generalization performance of ML surrogate models?
• RQ4 To what extent can ML models reduce computational time while preserving accuracy in reliability analysis?
• RQ5 How accurate is the ML-assisted Monte Carlo method when estimating the probability of failure compared to the full reference simulation?

Key findings

• The Bagging Ensemble and Random Forest classifiers achieved the highest performance, with 85% accuracy and 91% AUC on the full 120,000-sample dataset when trained on only 500 samples (0.47% of data).
• The computational time for 120,000 simulations was reduced from 306 days to less than 6 hours using the trained ML surrogate model.
• The average error in predicted probability of failure using only 5% of MC data was 0.46%, demonstrating high reliability of the ML-assisted approach.
• Model performance decreased with increasing soil heterogeneity and anisotropy, indicating higher uncertainty in complex soil conditions.
• The binary classification approach (failure/non-failure) effectively bypassed the need for factor of safety computation, significantly reducing per-simulation cost.
• The study confirms that ML surrogates can replace computationally expensive MC simulations in slope stability reliability analysis with minimal loss of accuracy.