Application of machine learning algorithms to the simplification of the performance evaluation process of cement dispersants

Abstract

The performance of cement dispersants and their influence on the rheological behavior of concrete can be evaluated using a variety of methods, including slump test or ICAR rheometer viscosity measurements

however, they are considered labor, time, and resource- ineffective. Thus, unsupervised machine learning algorithms are suggested to be used to develop an algorithm capable of proposing candidates for further testing, avoiding the necessity to test numerous cement dispersants under investigation. A model is based on Principal Component Analysis (PCA) and K-means clustering algorithms, the results of which are used to detect the cement dispersants located on the established data boundary. Its accuracy is evaluated based on the Type II error and J-score quantification that allows selecting a dataset built upon the testing procedure with lower experimental costs. As a result, the combination of Fourier-Transform Infrared Spectroscopy (FT-IR) and Gel Filtration Chromatography (GFC) test results provided the highest accuracy in identifying special PCE-based admixtures, reducing the number of cement dispersants proposed as candidates for prospective analysis by more than a half.