Interpretable physics-aware alkali-silica reaction expansion prediction

Abstract

The alkali-silica reaction (ASR) is a major contributor to the aging and degradation of infrastructure. Understanding ASR-induced expansion in concrete structures and accurately predicting its future progression are critical components of effective risk assessment frameworks. This paper presents a study focused on developing and interpreting an advanced machine learning model specifically designed to predict ASR expansion. The model strategically integrates two powerful algorithms to achieve this goal. A comprehensive database comprising 2000 samples of ASR expansion data with various attributes was used to train the model. The first algorithm, eXtreme Gradient Boosting (XGBoost), was employed to establish a predictive model for ASR expansion, achieving approximately 90% validation accuracy. The second algorithm, SHapley Additive exPlanations (SHAP), was applied to assess the relative importance of the factors influencing the XGBoost model’s predictions. This approach provided valuable physical and quantitative insights into the input–output relationships, which are often obscured in conventional machine learning methods. The study revealed that higher silica content, elevated alkali levels, and longer reaction times are strongly correlated with increased ASR expansion. In contrast, larger aggregate sizes and higher water-to-cement ratios were associated with reduced expansion. Additionally, the analysis identified alkali content as the most influential factor in determining the ultimate ASR expansion, while silica content emerged as the key parameter affecting the characteristic time of the ASR curve.