Explainable Machine Learning for Soilcrete UCS Predictions-PPT Presentation

PPT Presentation

Pdf Summary

The presentation examines explainable machine learning for predicting the unconfined compressive strength (UCS) of soilcrete, a soil–cement mixture used for ground improvement. Because UCS depends on nonlinear interactions among factors such as soil type, moisture, binder content, age, and mixing methods, it can be difficult to estimate using experience or standard guidelines alone.<br /><br />Researchers applied XGBoost to data from 15 Keller projects in California, comprising 2,226 soil-mix samples. The dataset included specimen age, binder content, mixing and soil measurements, CPT data, sampling method, and UCS. The model explained 88% of the variance in UCS. SHAP analysis was used to interpret how features influenced predictions, both across the dataset and for individual samples.<br /><br />Specimen age was the most influential factor: older samples generally had higher predicted UCS, while predictions tended to decrease over longer timescales. Soil type also mattered, with clay-like soils associated with lower predicted strength. Sampling method had little overall influence, though core samples showed slightly higher UCS values.<br /><br />The identified influential features were consistent with existing literature. The authors suggest that machine-learning analysis can help guide targeted field sampling and support design decisions, while noting future work could add or transform input features and investigate feature interactions.

Keywords

explainable machine learning

soilcrete

unconfined compressive strength

XGBoost

SHAP analysis

ground improvement

soil-cement mixture

specimen age

soil type

field sampling