Probability integrated hybrid machine learning models for predicting surface settlement of PVD-treated soft soil

Predicting surface settlement of soft soil under the combined prefabricated vertical drain (PVD)–embankment system remains a critical challenge in geotechnical engineering due to the complex and time-dependent nature of soil behavior. This study leverages in situ data collected from various real-world projects to develop a hybrid machine learning model that incorporates prediction uncertainty (confidence interval). Twelve input variables are determined based on conventional theories of radial soil consolidation under embankment, encompassing soil properties, PVD parameters, and loading conditions. Different machine learning algorithms are extensively evaluated with the Categorical boosting (CATB), which emerges as the most reliable and accurate algorithm for predicting the settlement of PVD-treated soft soil. The CATB performance is further enhanced by the adaptive step random search (ASRS), proving its exceptional predictive accuracy and robustness. The developed model is applied to an independent case from the field-scale embankment test at Australia’s National Field Testing Facility (NFTF), yielding promising outcomes. Importance-based sensitivity analysis revealed the dominant influence of key parameters, including time, pre-consolidation pressure, recompression index, and PVD characteristics, on settlement behavior, offering actionable insights for optimized design. The probabilistic parameters are incorporated into the data-driven model, facilitating the assessment of prediction confidence, advancing practical design of soft soil improvement, and bridging the gap between theoretical modeling and real-world geotechnical design needs.