2024 : 5 : 22
Morteza Ghorbanzadeh Ahangari

Morteza Ghorbanzadeh Ahangari

Academic rank: Associate Professor
Education: PhD.
Faculty: Faculty of Technology and Engineering
Phone: 35305107


Experimental assessment of the efficiency of deep learning method in predicting the mechanical properties of polymer concretes and composites
Polymer concrete Deep learning Deep neural network Mechanical properties
Journal Journal of Building Engineering
Researchers Mostafa Hassani Niaki ، Matin Pashaian ، Morteza Ghorbanzadeh Ahangari


The current study is an attempt to investigate the usefulness of a deep learning-based method, backpropagation deep neural network (DNN) for the prediction of the mechanical properties of basalt fiber and nanoclay reinforced polymer concrete (PC) composite. To accurately evaluate the efficiency of the DNN model, instead of building the dataset from the experimental data available in the literature, all training and validation data are prepared from the experimental investigation conducted in this work. First, the mechanical properties of neat epoxy, and its combination with fly ash, silica sand, crushed basalt, basalt fiber, and nanoclay (67 different combinations) are experimentally obtained. The obtained results are used to construct three datasets. Then, three independent DNN models are presented so that the weight percentage of the ingredients is six input variables and compressive, flexural, and splitting tensile strengths of the PC are the output parameters. The DNN models including three hidden layers and the structures of 6-13-13-11-1 are selected as optimal configurations. The prediction accuracy of the validation specimens remains satisfying with the maximum relative errors of 3.49%, 3.45%, and 4.43% for compressive, flexural, and splitting tensile strength, respectively. The statistical criteria R2, RMSE, and MAE for tensile strength are calculated 0.9953, 0.4471, and 0.3219, respectively. The DNN model for tensile strength shows a higher accuracy than compressive and flexural strength. The obtained results have proven the effectiveness of the deep learning method for predicting the properties of PC, even with a small dataset (67 experimental data points). Finally, partial dependence plots (PDPs) are implemented within the presented DNN models to analyze the sensitivity of the predicted mechanical properties to each single input variable. Using the obtained results, it is possible to predict the mechanical properties of PC without fabricating and testing the specimens, as well as obtain the optimal weight percentage of resin and fillers to achieve the maximum compressive, flexural, and tensile strength. This information is useful in designing and optimizing the composition of the PC.