USE OF ARTIFICIAL NEURAL NETWORK IN HIGH PERFORMANCE CONCRETE STRUCTURAL APPLICATIONS

Abstract

Flat plate systems are widely used in reinforced concrete structures for their appealing architectural properties. Using high-strength concrete (HSC) and, to a lesser extent, high performance concrete has been recently common for reinforced concrete structures at whole; not to mention flat slabs. Within this study, a thorough literature review has been conducted on the punching shear capacity of slab-column connections wherein the concrete used was either normal strength or high strength concrete (HSC).

In this study, an extensive data set of 495 punching shear samples were collected from earlier studies; used to both compare and recalibrate existing models/formulas, available in codes and/or research papers. Out of the former dataset, 150 samples are of high strength concrete (HSC); whilst considering HSC of a compressive strength (fcu) greater than 50 MPa; less than this value – 345 samples - are considered as normal strength concrete (NSC).

In this study, nine punching-shear-capacity models (in codes and earlier research efforts) were compared - in their original form - in light of popular statistical parameters; the most important of which is the “Normalized Root Mean Square Error” (NRMSE) as well as the “Coefficient of Determination” (R2). This is in addition to the amelioration/enhancement of the former models using the “Matlab optimization toolbox”; wherein the NRMSE is minimized through a computerized iteration process; in turn yielding the optimal coefficients (whether multipliers or exponents). This enhancement process was conducted for each individually: (i) Normal strength concrete (NSC); (ii) High strength concrete (HSC) and (iii) the entire dataset comprising both NSC and HSC. Consequently, the optimized model of best performance was chosen to conduct a parametric study that illustrates all influencing parameters versus the output/objective punching-shear-capacity.

In light of the above, a new empirical model was proposed where NRMSE is inferior to any of the compared models, whether in codes or displayed in earlier research studies. Furthermore, an Artificial Neural Network (ANN) model was proposed; wherein the output - typically displayed in a complicated matrix-based form - is demonstrated in a simplified manner. The results of the latter were compared to that of the former proposed empirical model. Proximity was evident in both models, where the former shows an NRMSE value lower than the latter by only 13%.