Machine Learning Approaches for Predicting Flow and Consistency Retention Performance in Cementitious Mixtures with Diverse Grinding Aid Dosages and Types.

Grinding aids (GAs) are commonly utilized to enhance energy efficiency during clinker grinding and improve cement properties. However, alongside their benefits, GAs may lead to challenges such as reduced flow performance, loss of consistency retention, and increased demand for water or water-reducing admixtures in cementitious systems. Therefore, it is crucial to assess the flow properties and consistency retention behavior (time-dependent flow) in mixtures prepared with GAs. Given that such experiments are labor-intensive and time-consuming, predicting these performance metrics through machine learning offers substantial advantages. In this study, 29 different cements, incorporating various types and dosages of GAs, were produced. Time-dependent flow and compressive-strength tests were conducted on mixtures made with these cements. Additionally, the experimental results were compared with predictions using four machine learning models: random forest, adaptive boosting (AdaBoost), gradient boosting and multilayer perceptron. The computational results indicated that random forest outperformed the other machine learning algorithms in predicting compressive strength, whereas multilayer perceptron achieved the best overall performance for flow value prediction. Furthermore, it was concluded that random forest, adaptive boosting and gradient boosting provided acceptable performance in compressive strength prediction, while random forest and multilayer perceptron demonstrated comparatively better performance in predicting the flow value. However, the study is limited to specific types and dosages of GAs and does not include the influence of temperature, humidity, or long-term durability metrics, which may affect the generalizability of the machine learning models.