Prediction of early-age mechanical properties of high strength concrete with pozzolans by using statistical methods

Abstract

The developments in concrete technology are becoming more important and effective with the help of innovative approaches on materials and computer sciences and their applications. With advanced calculation methods, computing programs/softwares and supercomputers, the mechanical behavior of concrete is better understood in many aspects, today. In addition, the materials used in concrete technology are now much more diverse, more useful, and much more effective than in the past by the opportunities provided from the industry. On the other hand, this level of development and effectiveness still depends on specific needs of concrete. However, this natural limitation does not prevent performance improvement, durability, sustainability, environmental and budget-friendly expectations of concrete in a planned service life. Accordingly, while cement types, aggregates, moisture contents of aggregates, and air contents in concrete mixtures maintain their importance, the concrete mixture designs can be rearranged by weight and/or concrete mixing ratios according to the relevant pioneer test results, and new concrete matrices can be obtained by using fly ash, micro silica, nano silica, ground blast furnace slag, fiber, glass, wood, etc. Moreover, recyclable materials such as water, aggregate, glass, fiber, wood, etc. and even living organic materials are the topics that the concrete industry has recently focused on. In this context, the idea of using new construction materials may arise depending on relevant test results of special concretes produced for special projects. However, willing to change the concrete mixture designs and/or building materials based on test results can be quite difficult, because of time and budget concerns. For this reason, the most used type of concrete in the ready mixed concrete world is normal weight concrete (NWC), which is adapted by the concrete industry. Considering this fact, despite all the possibilities, determining a right concrete mixture design still differs in many ways depending on time, material, and external factors. In this idea, in general, specimens of hardened concrete in the form of cubes, cylinders, and rectangular prisms are tested at an early age to obtain results of mechanical properties such as compressive strength, splitting tensile strength, and modulus of elasticity so that further investigations and predictions of the concrete can be made. According to these test results, statistical methods come to the fore in many cases in terms of time and cost efficiency, and deep analysis to predict results of concrete performance depending on time and material to decide whether these concrete mixture designs comply with standards and regulations. Because, in regression analysis, which is one of these statistical methods, it is possible to predict a mechanical property of concrete without using destructive or non-destructive methods with enough concrete samples. In this way, the gains are obtained in terms of space, time, and cost. As a further step from the regression analysis, the use of machine learning methods such as Neural Net Fitting (NNF) to predict a data has become quite common today in the concrete world. Before statistical estimation of a data set, the concrete mixture designs should be cared for their validations. Furthermore, the atmospheric conditions at work sites where the concrete is casted are very important to obtain realistic test results from the concrete casting process. Therefore, the experiments such as slump, flow, unit weight, air content, ambient temperature, bleeding, adiabatic process, setting time etc. for fresh concrete samples can be carried out in the work fields. For this thesis, fresh concrete samples were taken for 33 different concrete mixture designs in 150X300 mm cylindrical sample containers in the numbers allowed by national standards and regulations. Besides, two distinct types of fine aggregates (FA) and three diverse types of coarse aggregates (CA) were used in these mixture designs with fly ash (FA) + micro silica (MS), ground granulated blast furnace slag (GGBS), and five different cement (C) types were used as binding material for these designs. The samples prepared within this framework were also kept in safe places in the worksites for the first setting process of the concrete, right after the sampling process was completed. Subsequently, the concrete samples, when the initial setting process were completed, were transferred to the laboratory environment for the hardened concrete tests in the international standards for 0.5, 1, 2, 3, 7, 14 and 28 days. And, the samples were prepared for the compressive strength, splitting tensile strength and modulus of elasticity tests for statistical analysis and estimations. In this thesis, as one of the statistical analysis models, regression analysis based on convergence of the obtained estimation results to real data (drawing curves) are used. The properties such as age of concrete samples (time), unit weights of mixture components, unit volumes of mixture components, mixing ratios and/or coefficients of an estimation methods etc. were analyzed individually and cumulatively. Accordingly, the relations of the predicted data with the concrete mixture designs are studied with linear or non-linear equations in univariate and multivariate regression models. In addition to the equations used for the estimation of the test results, other statistical results such as R (Correlation of Coefficient), R² (Coefficient of Determination), R²adj (Adjusted Correlation of Determination), Sum of Squared of Errors (SSE), Mean Square Error (MSE), and Root Mean Square Error (RMSE) were obtained. The relationships between the actual test results, and predicted results were examined at the end. Due to the nature of the models used in the univariate regression analysis, only one variable was considered, and the results were estimated accordingly. The number of variables taken into consideration was analyzed individually for each mixture design. Although such individual analyzes were possible, many sequential studies on the actual, and estimated results had been the cost of time. Therefore, predicting the actual results required more complex analyzes like the multivariate regression analysis in this study. Before the more complex analyses, the variables were studied one-by-one and/or in combinations for the multiple regression analyses. The substantial number of these combinations let the study to the machine learning process, and the effect of hidden layers between the input (mixture designs) values and the target (test) values four output values (algorithm results) were observed in the machine learning process. Although it was really complicated to detect these hidden layers by the individual calculations, only the input values, and target data values were chosen in the machine learning procedure without stepping directly into the hidden layers. On the other hand, it was understood that increasing the number of hidden layers deviated the estimation results from the target values. Therefore, to obtain more accurate results, the number of samples in the machine learning algorithms were changed as much as possible, while the number of hidden layers was increased. Yet, it was revealed that increasing the number of samples and/or hidden layers at the same time caused undesirable estimation results. It was also determined that an infinite number of experiments could be made with the machine learning to predict the target values. But, since it was not possible to conduct an infinite number of trials one-by-one, all trials were recorded first, and then evaluated from the best to the worst and/or in the Levenberg-Marquardt (LM) algorithm form the NNF machine learning process. In addition to this, R and MSE values in the NNF machine learning process, training, validation, test, and all correlation results were displayed in the x - y planes. Finally, in this framework, the best results were shared in association with the statistical results with physical meanings specific to mixture designs.