Probabilistic and Data-Efficient Modelling of Soil Properties: Review, Tool Development and Practical Applications

Reliable estimation of soil properties is crucial for geotechnical design. While machine learning (ML) offers significant potential in this domain, its application remains challenging for practitioners lacking expertise in data-driven modelling, including data acquisition, model development, evaluation, and deployment. Moreover, conventional data sampling in geotechnical practice is often random and inefficient, leading to excessive costs and suboptimal model performance. This study provides a comprehensive review of ML algorithms applied to soil properties modelling and introduces five representative probabilistic methods: quantile random forest (QRF), variational inference-based evolutionary polynomial regression (VIEPR), ensemble-based support vector regression (ESVR), Bayesian neural networks (BNN), and Gaussian process regression (GPR). These algorithms are integrated into a generic and user-friendly graphical interface platform, ErosMLM, enabling seamless model development, optimization, evaluation, and application through intuitive click-based operations. Furthermore, an active learning strategy is proposed to prioritize informative data acquisition, reducing data demands and associated costs. The platform and strategy are validated through a case study on soil creep index prediction, demonstrating high accuracy, reliability, and practical utility. This work highlights the potential of ML to advance geotechnical modelling while lowering the barrier to its adoption.