Explainable Machine Learning for Soilcrete UCS Predictions-Paper

Paper

Pdf Summary

The study applies explainable machine learning to predict the unconfined compressive strength (UCS) of soilcrete, a cement-treated soil produced by in-situ mixing. Researchers compiled 2,226 test results from 15 western U.S. projects, combining strength measurements with soil, site, binder, mixing, sampling, and project-timing data. They trained an XGBoost model and used SHAP to interpret its predictions.<br /><br />The model explained 88% of the UCS variation in the held-out test data. SHAP analysis found specimen age to be the strongest predictor: older samples generally had higher predicted strength. Predicted UCS also tended to decline over time, both within projects and across the dataset. The authors suggest this may reflect contractors’ increasingly precise mix designs and reductions in binder content after initially conservative installations. Sand-like soils generally contributed to higher predicted strength than clay-like soils, while higher grout specific gravity also increased predictions. Sampling method had comparatively little influence, although core samples sometimes contributed to higher predicted values than grab samples.<br /><br />For individual predictions, SHAP helped explain both accurate estimates and errors, including flagging a potentially mismatched cone penetration test record as a possible source of underprediction. The authors conclude that explainable machine learning can support UCS estimation while revealing patterns, biases, and data-quality issues that warrant further investigation. The model is an initial tool rather than a replacement for engineering judgment; future work could incorporate additional variables, test feature interactions, and collect targeted data to improve validation and performance.

Keywords

explainable machine learning

soilcrete

unconfined compressive strength

UCS prediction

XGBoost

SHAP analysis

cement-treated soil

binder content

specimen age

soil mixing