Explainable Machine Learning for Soilcrete UCS Predictions-Video

Video Transcription

Video Summary

Catherine Chen presented an explainable machine-learning model for predicting the unconfined compressive strength (UCS) of soil-cement mixes. Using data from 2,226 samples across 15 California projects, she trained an XGBoost model with an 80/20 train-test split. The model achieved an R² of 88% on the test set. SHAP explanations showed that specimen age was the most influential feature: older samples generally had higher predicted strength. Clay-like soils were associated with lower strength, while core samples had a small positive effect. Strength predictions also tended to decline later in project timelines, potentially reflecting adjustments to mixes as projects progressed. Individual explanations helped identify both typical predictions and cases where the model’s estimates differed substantially from measured values. The findings aligned with established soil-mixing knowledge and may guide targeted field sampling. Chen noted that the model can only account for variables included in its dataset, underscoring the importance of selecting relevant inputs.

Keywords

explainable machine learning

soil-cement strength prediction

XGBoost model

SHAP explanations

unconfined compressive strength