An Integrated Machine Learning Framework for Predictive Assessment of Soil Bearing Capacity and Settlement Using Multisource Geotechnical and Environmental Data

Abstract

Abstract Reliable foundation design requires simultaneous control of ultimate bearing capacity and serviceability settlement, yet conventional prediction methods commonly treat soil parameters as static and omit rainfall, moisture, matric suction and groundwater fluctuations. This article develops a machine-learning-driven framework that integrates geotechnical, foundation and environmental data for joint prediction of shallow-foundation bearing capacity and settlement. A transparent proof-of-concept benchmark comprising 480 physically constrained synthetic records from 12 heterogeneous sites was created from accepted strength, stiffness and load-deformation relationships. Ridge regression, support vector regression, random forest, extra trees and gradient boosting were evaluated using site-grouped cross-validation and a holdout test containing entirely unseen sites. Gradient boosting achieved the best holdout performance, with R² values of 0.906 for ultimate bearing capacity and 0.893 for settlement. Adding environmental variables increased bearing-capacity R² from 0.894 to 0.906 and settlement R² from 0.772 to 0.893. Permutation analysis identified friction angle, footing width, cohesion and embedment as dominant bearing-capacity predictors, whereas soil modulus, service pressure, footing width, moisture and recent rainfall governed settlement. The framework combines data-quality control, explainability, uncertainty screening and engineering checks, offering a reproducible pathway for site-calibrated decision support rather than replacement of geotechnical judgement