Methodology for generating diverse geotechnical datasets using Monte Carlo simulation and genetic algorithms

Abstract

The reliability of machine learning heavily depends on training data; however, in the field of geotechnical engineering, it is challenging to obtain diverse datasets due to economic and accessibility limitations. The aim of this study is to propose a method for generating data for use in the training phase of machine learning by combining Monte Carlo simulations and genetic algorithms. The original data sample is constructed using a 1 × 1 m grid for a slope, based on geotechnical properties measured in 23 regions, including soil cohesion, slope angle, soil density, soil depth, and friction angle. Based on the original sample, further predictions are made at an additional 1777 grid locations to estimate the spatial distribution of geotechnical properties across the entire slope. When a single variable is used as input, the log-likelihood values (e.g., –5.4 to –144.5) are used only as relative indicators, not as absolute measures. The results are also compared to those generated using existing algorithms such as the synthetic minority oversampling technique and adaptive synthetic sampling. The data generated using the proposed method exhibits fewer duplicate values, broader distribution ranges, and greater diversity. To ensure that the generated data closely aligns with the statistical characteristics of the actual data, the combination of input variables is configured to maximize the log-likelihood value. To achieve this, Pearson correlation values are referenced, and multivariate input variables are constructed using highly correlated factors. As a result of this approach, the log-likelihood value increased by 21% to 96%. This study demonstrates that the method combining Monte Carlo simulations and genetic algorithms generates data with more diverse distributions, compared to existing methods. It also highlights that constructing multivariable input data is preferable for improving reliability.