Machine Learning-Based Modelling of Soil Properties for Geotechnical Design: Review, Tool Development and Comparison

Abstract

Machine learning (ML) holds significant potential for predicting soil properties in geotechnical design but at the same time poses challenges, including those of how to easily examine the performance of an algorithm and how to select an optimal algorithm. This study first comprehensively reviewed the application of ML algorithms in modelling soil properties for geotechnical design. The algorithms were categorized into several groups based on their principles, and the main characteristics of these ML algorithms were summarized. After that six representative algorithms are further detailed and selected for the creation of a ML-based tool with which to easily build ML-based models. Interestingly, automatic determination of the optimal configurations of ML algorithms is developed, with an evaluation of model accuracy, application of the developed ML model to the new data and investigation of relationships between the input variables and soil properties. Furthermore, a novel ranking index is proposed for the model comparison and selection, which evaluates a ML-based model from five aspects.Soil maximum dry density is selected as an example to allow examination of the performance of different ML algorithms, the applicability of the tool and the model ranking index to determining an optimal model.

Back-calculation of soil parameters from displacement-controlled cavity expansion under geostatic stress by FEM and machine learning

“…The use of ML in geotechnics has seen exponential growth over the last decade [50]. ML has indeed proven to be a useful tool to provide estimations of soil properties and/or limit loads for geotechnical applications where no closedform analytical solutions exists.…”

“…For a given application, there is usually no consensus on which ML algorithm or data preparation methodology works best, and therefore some authors have written reviews that compare the advantages/limitations of several approaches. For instance, Lary et al [17] reviewed applications of ML in geo-sciences and remote sensing, Zhang et al [51] summarized applications of ML in the constitutive modeling of soils, and Wang and Sun [50] reviewed applications targeted toward modeling of soil properties. Although the range of applications is wide, most studies can be grouped into three categories: i) Estimation of mechanical properties of a system for specific loading conditions and soil type, ii) Estimation of a set of design parameters (e.g., limit load, factor of safety) from the response of a soil to stimuli and iii) Generation and/or calibration of constitutive models.…”

“…The model computes the average / in the training set, and then directly uses it as prediction of all samples in the validation set and test set. Linear regression is a popular and interpretable approach to study the relationship between soil properties and cavity behavior [50]. We applied Lasso regularization [41] to all of our linear regressions, as it can shrink the coefficients of less predictive features to zero and improve model generalizability.…”

Machine Learning Models for Predicting Shear Wave Velocity of Soils

“…Some commonly used methods for the in situ measurement of Vs include seismic refraction measurements, surface waves, and cross-hole and downhole techniques [2][3][4]. Owing to the favourable characteristics of machine learning (ML), many researchers in the literature have suggested the use of ML approaches for determining geotechnical parameters [5][6][7][8][9].…”

Ready-to-use deep-learning surrogate models for problems with spatially variable inputs and outputs

“…Incorrect operation in any component would result in unpredictable results. (4) The computational time for coupled multi-physics problems is time-consuming, which makes stochastic analysis and optimisation practically impossible because these tasks all require many such analyses [32].…”