Compressive strength prediction of reactive powder concrete using fuzzy interface system model

Ultra-high-performance concrete is an extremely heterogeneous material with large variations of materials. The strength of the concrete is majorly dependent on the mix design and its material properties. In this research, an attempt is made to develop a model for compressive strength prediction and optimized mix design of reactive powder concrete (RPC) by a Mamdani-based Fuzzy Interface System (FIS). A total set of 8 main parameters influencing the strength of the concrete were considered as input variables. 130 experimental data were used, in which 100 datasets were collected from various works of literature and 30 datasets were collected through self-conducted experimental investigations. 150 rules were set based on the permutation and combination of various material inputs. 30 test results were considered to check the efficiency of the FIS output by applying various membership functions and defuzzification methods. The predicted results show the potential efficiency of FIS in the prediction of the compressive strength and mix design of reactive powder concrete. The experimental analysis was found accurate to the analytical prediction, thus proving a perfect correlation between the experimental analysis and the soft computing techniques. The average accuracy was found to be 98-99% for Triangular, Gaussian, Pimf, D-Sigmoidal, and G-Bell functions. Whereas S-Curve, Z-Curve, and sigmoid curve failed miserably by providing inconsistent and inaccurate prediction when used. But the same curves proved their efficiency of 98-99% when used as hybrid functions with the highly precised membership functions like Triangular, Gaussian, Pimf, D-Sigmoidal, and G-Bell membership functions. Further to the predicted results obtained from soft computing, these were checked with the experimental results (with same mix design and input parameters). The R2 values obtained by using Triangular, Gaussian, Pimf, D-Sigmoidal, and G-Bell functions with various defuzzification methods were found out to be 0.99, which ensures a highly accurate prediction analysis. Whereas the curve like S-Curve, Sigmoidal, and Z-Curve resulted the R2 values in the range of 0.31 to 0.81, showing highly inconsistent and inaccurate values. But, when combined as hybrid, it resulted in R2 value of 0.96 to 0.99, which is highly accurate value.