Artificial neural network evaluation of concrete performance exposed to elevated temperature with destructive–non-destructive tests.

In this study, it is aimed to predict the performance of concretes obtained by using supplementary cementitious materials (SCM) before and after high temperature using artificial neural network. Thus, in addition to contributing to sustainable development and circular economy by using waste materials in concrete production, predicting concrete strength using artificial neural network without the need for experimental studies will provide a great advantage in practice. In addition, it will also contribute to the literature in terms of determining the optimum amount of metakaolin to be used with fly ash in concrete production. Metakaolin, silica fume and fly ash were used as SCM in different proportions in concrete mixes. Accordingly, a total of 22 concrete series were prepared, one of which was the control series. Porosity, ultrasonic pulse velocity, pressure and tensile strength tests were applied to the series at the end of 7th, 28th and 90th curing periods before high temperature. In order to determine the strength losses after elevated temperature, porosity and compressive strength tests were applied at temperatures of 400, 600 and 800 °C. Mineral additive series showed positive mechanical properties up to 20%. However, it has been observed that the use of fly ash after a certain rate causes a decrease in strength. After elevated temperature, strength loss was observed in all series due to the increase in temperature, while it was observed that the rate of being affected by elevated temperature decreased as the percentage of metakaolin increased. Optimum mineral additive usage percentages were determined as 10% fly ash and 15% metakaolin. On the other hand, the use of mineral additives above the optimum level caused the performance of the concrete to decrease. Then, the concrete compression strengths obtained at 7th, 28th, and 90th days and at 400, 600 and 800 °C temperatures are taken as the outputs of the ANN. The artificial neural network provided the closest results to experimental data. Moreover, to prove the predictive performance of ANN, a comparative analysis was made with GPR, SVM and LR and the smallest value of the RMSE value is obtained with the ANN model. Finally, a fivefold cross-validation criteria was used to objectively present the performance of the model. [ABSTRACT FROM AUTHOR]