A Machine Learning-Based Approach for Predicting Structural Settlement on Layered Liquefiable Soils Improved with Densification

In this paper, we propose a machine learning-based approach for predicting foundation settlement on liquefiable soils improved through ground densification. The model considers variations in the properties of the soil profile, foundation, 3D structure, mitigation design (in terms of densified depth and width), and ground motion. A numerical data set from 770, 3D, fully coupled, effective-stress, finite element analyses was developed initially with a statistically determined range of input parameters (through quasi-Monte Carlo sampling). The numerical models were themselves calibrated and validated with centrifuge model studies. Subsequently, the numerical database and an additional 15 centrifuge experiments were used to train a gradient boosting model (tree-based, supervised, machine learning method, GB) for predicting foundation's settlement. In general, the data-driven GB model could better predict settlement compared to the classical regression model (by about 14%). This is because the non-functional form model could better capture the nonlinear trends in permanent foundation settlement as observed in the numerical and experimental database. However, when evaluating a very limited existing field case history database in the literature, the data-driven GB model only slightly improved the settlement predictions compared to the regression model. This is because the GB model cannot take the impact of model features on foundation settlement in a continuous manner (due to the inherent shortcoming of a decision-tree framework in GB), leading to a dramatic increase in model uncertainty when the input parameters are outside the ranges considered in the database. The insight from the presented GB model aims to guide the development of future data-driven predictive models for a more reliable estimation of engineering demand parameters related to soil-foundation-structure systems.