Accurate estimation of soil properties is crucial for reliability-based design in engineering practices. Conventional empirical equations and prevalent data-driven models rarely consider uncertainty quantification in both measurement and modelling processes. This study tailors three uncertainty quantification methods including Bayesian learning, Markov chain Monte Carlo and ensemble learning into data-driven modelling, in which support vector regression is selected as the baseline algorithm. The compression index of clay is adopted as an example for model training and testing. In this context, Bayesian learning and Markov chain quantify uncertainty by considering the distribution of function and hyper-parameters, respectively, while different sampled data are employed to explore model uncertainty. These models are evaluated in terms of accuracy, reliability and cost-effectiveness and also compared with Gaussian process regression, etc. The results reveal that based on built-in structural risk minimization, sparse solution and uncertainty quantification, developed models can capture more accurate and reliable correlations from actual measured data over other methods. Their practicability and generalization ability are also verified on a new creep index database. The proposed probabilistic methods are also compiled into a user-friendly platform, showing a significant potential to enrich the data-driven modelling framework and be applied in other geotechnical properties.

Seasonally ice-bound ground are subjected to cyclic freeze-thaw processes, which can significantly degrade their mechanical properties, including static strength (Ss). To accurately characterize and predict the Ss of seasonally frozen soils, this research employs advanced machine learning techniques. Specifically, the study utilizes the Least Square Support Vector Regression (LSSVR) method, which is known for its robust performance in nonlinear regression tasks. A critical aspect of the LSSVR model is the appropriate selection of its hyperparameters, namely the penalty agent (c) and the breadth of the kernel function (g). To determine these parameters with high precision, the research integrates the LSSVR model with two novel optimization techniques: the Flow Direction Algorithm (FDA) and the Artificial Rabbit Optimization (ARO). The resulting hybrid models, denoted as LS(ARO) and LS(FDA), are designed to outperform the previously published Artificial Neural Network (ANN) approach in predicting the Ss of seasonally frozen soils. The Implementation of the proposed hybrid approaches is assessed by a comprehensive database of 120 soil samples collected from relevant published studies. The input parameters used in the frameworks include water content, negative temperature, confining stress, freeze-thaw processes, thawing time, and compaction ratio. The results demonstrate the superiority of the hybrid models, with the LS(ARO) network achieving remarkable R2 amounts of 0.9924 and 0.9976 during the train and test steps, respectively. Moreover, the LS(ARO) model outperformed the LS(FDA) and the previously reported ANN model in terms of other performance metrics, such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). The results of the present research expand the recognizing and predictive capabilities of Ss in seasonally frozen soils, which is crucial for infrastructure design and construction in cold regions. The integration of the LSSVR technique with the novel FDA and ARO optimization algorithms represent a significant advancement in the field of hybrid regression analysis for geotechnical engineering applications.

Accurate settlement forecasting is essential for preventing severe structural and infrastructure damage. This paper investigates predicting tunneling-induced ground settlements using machine learning models. Empirical methods for estimating settlements are often imprecise and site-specific. Developing novel, accurate prediction methods is critical to avoid catastrophic damage. The umbrella arch method constrains deformations for initial stability before installing primary support. This study develops machine learning models to forecast settlements solely from umbrella arch parameters, disregarding soil properties. Multilayer perceptron artificial neural networks (MLP-ANN) and support vector regression (SVR) are applied. Results demonstrate machine learning outperforms empirical methods. The MLPANN surpasses SVR, with R2 of 0.98 and 0.92, respectively. Strong correlation is observed between umbrella arch configuration and settlements. The suggested approach effectively estimates surface displacements lacking mechanical properties. Overall, this study supports machine learning, specifically MLP-ANN, as an efficient, reliable alternative to empirical methods for predicting tunneling-induced ground settlements from umbrella arch design.

In cold regions, the dynamic compressive strength (DCS) of rock damaged by freeze-thaw weathering significantly influences the stability of rock engineering. Nevertheless, testing the dynamic strength under freeze-thaw weathering conditions is often both time-consuming and expensive. Therefore, this study considers the effect of characteristic impedance on DCS and aims to quickly determine the DCS of frozen-thawed rocks through the application of machine-learning techniques. Initially, a database of DCS for frozen-thawed rocks, comprising 216 rock specimens, was compiled. Three external load parameters (freeze-thaw cycle number, confining pressure, and impact pressure) and two rock parameters (characteristic impedance and porosity) were selected as input variables, with DCS as the predicted target. This research optimized the kernel scale, penalty factor, and insensitive loss coefficient of the support vector regression (SVR) model using five swarm intelligent optimization algorithms, leading to the development of five hybrid models. In addition, a statistical DCS prediction equation using multiple linear regression techniques was developed. The performance of the prediction models was comprehensively evaluated using two error indexes and two trend indexes. A sensitivity analysis based on the cosine amplitude method has also been conducted. The results demonstrate that the proposed hybrid SVR-based models consistently provided accurate DCS predictions. Among these models, the SVR model optimized with the chameleon swarm algorithm exhibited the best performance, with metrics indicating its effectiveness, including root mean square error (RMSE) = 3.9675, mean absolute error (MAE) = 2.9673, coefficient of determination (R2) = 0.98631, and variance accounted for (VAF) = 98.634. This suggests that the chameleon swarm algorithm yielded the most optimal results for enhancing SVR models. Notably, impact pressure and characteristic impedance emerged as the two most influential parameters in DCS prediction. This research is anticipated to serve as a reliable reference for estimating the DCS of rocks subjected to freeze-thaw weathering. (c) 2024 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/ licenses/by-nc-nd/4.0/).