Assessment of compressive strength of Ultra-high Performance Concrete using deep machine learning techniques

Abstract

The compressive strength of Ultra-High Performance Concrete (UHPC) is a function of the type, property and quantities of its material constituents. Empirically capturing this relationship often requires the utilization of intelligent algorithms, such as the Artificial Neural Network (ANN), to derive a predictive model that fits into an experimental dataset. However, its black-box nature prevents researchers from mathematically describing its contents. This paper attempts to address this ambiguity by employing two deep machine learning techniques – Sequential​ Feature Selection (SFS) and Neural Interpretation Diagram (NID) – to identify the critical material constituents that affect the ANN. 110 UHPC compressive strength tests varying based on the material quantities were compiled into a database to train the ANN. As a result, four material constituents were selected; mainly, cement, fly ash, silica fume and water. These material constituents were then employed into the ANN to compute more accurate predictions (r2=80.1% and NMSE = 0.012) than the model with all eight material constituents (r2=21.5% and NMSE = 0.035). Finally, a nonlinear regression model based on the four selected material constituents was developed and a parametric study was conducted. It was concluded that the utilization of ANN with SFS and NID drastically improved the accuracy of the model, and provided valuable insights on the ANN compressive strength predictions for different UHPC mixes.