Compressive strength prediction of one-part alkali activated material enabled by interpretable machine learning

In recent years, alkali activated material (AAM) or geopolymer has emerged as a sustainable and eco-friendly alternative to cement. It is owing to its low power consumption and greenhouse gas emissions, as well as good mechanical and durability features. However, due to the nature and diversity of available source materials, developing an AAM mix to attain desirable fresh properties, sufficient strength characteristics, and touted environmental benefits is quite challenging. It demands a precise selection of input material and mix proportions based on several trials, which requires a large quantity of material, time, and effort. Therefore, employing machine learning techniques could facilitate and accelerate the development of one-part AAM binder with the desired properties. This study evaluates the performance of various machine learning models (Ridge regression, RF, LightGBM, and XGBoost) for accurate compressive strength prediction of one-part AAM binder. Extreme Gradient Boost (XGBoost) outperformed all other algorithms in terms of prediction efficacy and accuracy. In addition, SHapley Additive exPlanations (SHAP) is also used to interpret the predicted compressive strength through XGBoost and the effect of various parameters, independently and in relation with other parameters, is evaluated and discussed in detail. The interpretable ML strategy used in this study will aid in the production and performance tuning of durable and sustainable one-part AAMs for widespread applications.