Research on modeling the thixotropic properties of cementitious systems using regression methods in machine learning

Abstract

In this study, the rheological properties of cementitious systems were investigated through modeling studies on structural build-up and breakdown area. The area values were calculated using Herschel Bulkley analysis and hysteresis area method. The properties were examined by varying the composition of the cementitious system (cement fineness, C4AF, C3S, C2S, C3A, equivalent alkali and metakaolin ratio) and changes made in the rheological measurement processes (applied shear rate, maximum shear rate and duration). For this purpose, cement paste mixtures were prepared by substituting metakaolin at four different ratios (3%, 6%, 9%, and 12%) into cements with varying C3A content (2.13, 3.60, 6.82, 9.05%). The modeling study of the obtained results was conducted using three different learning methods: Linear Regression Analysis (LR), AdaBoost, and K Nearest Neighbor (KNN), encompassing machine learning and ensemble learning techniques. It was determined that the most dominant parameter affecting the rheology and thixotropic properties of the mixtures is the metakaolin usage ratio. The pre-shear rate was dominant over the duration and maximum shear rate parameters. Effect of the C3A content on dynamic yield stress and viscosity becomes more pronounced with an increase in the applied shear rate. The KNN method has yielded the best results in all experimental modeling studies. Euclidean distance criterion was used in the KNN method. Although the AdaBoost method obtained results close to the KNN method, the opposite situation was observed depending on the number of data. Logcosh, MAE and RMSE metrics were used to evaluate the experimental results. When the results for 3 different metrics in all modeling studies were examined, the success order of the metrics was found to be Logcosh, MAE and RMSE.