Explainable Artificial Intelligence for predicting the compressive strength of soil and ground granulated blast furnace slag mixtures

Abstract

Weak soil causes significant challenges during infrastructure development, necessitating soil stabilization to enhance its engineering properties. The pozzolanic properties of Ground Granulated Blast Furnace Slag (GGBS) have led to its widespread use as an effective stabilizer in soil improvement. This study aims to predict the UCS of soft soil stabilized with GGBS using various machine learning models. A database of 200 samples was compiled from the literature, and six ML models—linear regression, decision trees, random forest, artificial neural networks, gradient boosting, and extreme gradient boosting were developed and evaluated. The study highlights the performance of these models and employs SHAP and LIME analysis to evaluate feature importance. The XGB model emerged as the most effective predictor of unconfined compressive strength for soil treated with GGBS, accounting for over 90% of the variance explained by independent factors. The curing period, optimal moisture content, and maximum dry density served as critical variables influencing UCS, demonstrating the model's capacity to recognize underlying patterns and generate precise predictions. In addition to being more appropriate for complicated models, SHAPE is more accurate than LIME. SHAPE suggests that OMC has a detrimental impact on UCS in the current investigation, but LIME suggests the opposite. SHAPE results are in agreement with lab experiment results.

Keywords

• Machine Learning
• Ground granulated blast slag
• Unconfined compressive strength
• SHAP analysis
• Xtreme Gradient Boosting