Title:
Artificial Intelligence Approach for Predicting Compressive Strength of Geopolymer Concrete
Author(s):
Muhammad Naveed, Asif Hameed, Ali Murtaza Rasool, Rashid Hameed, and Danish Mukhtar
Publication:
Materials Journal
Volume:
122
Issue:
3
Appears on pages(s):
51-66
Keywords:
artificial neural network (ANN); constituent ratios; fly ash (FA); geopolymer concrete (GPC) strength; green material; mathematical modeling
DOI:
10.14359/51746714
Date:
5/1/2025
Abstract:
Geopolymer concrete (GPC) is a progressive material with the
capability to significantly reduce global industrial waste. The
combination of industrial by-products with alkaline solutions initiates an exothermic reaction, termed geopolymerization, resulting
in a carbon-negative concrete that lessens environmental impact.
Fly ash (FA)-based GPC displays noticeable variability in its
mechanical properties due to differences in mixture design ratios
and curing methods. To address this challenge, the authors optimized the constituent proportions of GPC through a meticulous
selection of nine independent variables. A thorough experimental
database of 1242 experimental observations was assembled from
the available literature, and artificial neural networks (ANNs) were
employed for compressive strength modeling. The developed ANN
model underwent rigorous evaluation using statistical metrics such
as R-values, R2 values, and mean squared error (MSE). The statistical analysis revealed an absence of a direct correlation between compressive strength and independent variables, as well as a lack of correlation among the independent variables. However, the
predicted compressive strength by the developed ANN model aligns
well with experimental observations from the compiled database,
with R2 values for the training, validation, and testing data sets
determined to be 0.84, 0.74, and 0.77, respectively. Sensitivity
analysis identified curing temperature and silica-to-alumina
ratio as the most crucial independent variables. Furthermore, the
research introduced a novel method for deriving a mathematical
expression from the trained model. The developed mathematical
expressions accurately predict compressive strength, demonstrating
minimal errors when using the tan-sigmoid activation function.
Prediction errors were within the range of –0.79 to 0.77 MPa,
demonstrating high accuracy. These equations offer a practical
alternative in engineering design, bypassing the intricacies of the
internal processes within the ANN.