Optimizing CNN, SVM, and MLP for Prediction of Compressive Strength of Concrete using Grid Search Optimization

Abstract

In civil engineering, the accurate prediction of concrete compressive strength (CS) is crucial for evaluation of material performance and structural design. In the present study, the main objective is to optimize the performance of three machine learning (ML) models including Support Vector Machine (SVM), Convolutional Neural Network (CNN), Multi-Layer Perceptron Neural Network (MLP) using Grid Search Optimization technique for improving the prediction accuracy of CS of concrete. For doing this, a total of 236 data points were collected from the “Red River Surface Water Plant” project, a major infrastructure initiative in Vietnam were collected and used to create training (70%) and testing (30%) datasets used for training and testing the models. For validation and comparison of the models, the popular validation metrics such as R², RMSE, MAE, and Taylor diagram were used. In addition, the Partial dependence plots (PDP) technique was used to validate the importance of each input variable used in the modeling. Analysis of the results illustrates that the optimized CNN, SVM, and MLP models significantly outperformed the single CNN, SVM, and MLP models, especially the optimized CNN model is the best compared with the optimized SVM and optimized MLP models, achieving an R² of 0.92, RMSE of 3.86 (MPa), and MAE of 3.09 (MPa). PDP analysis further revealed that key variables including cement content, coarse aggregates, and water-cement ratio have the most influential effects on the CS. The finding of this study highlights the advantages of combining deep learning with systematic hyperparameter optimization to capture complex, nonlinear relationships in concrete mix designs.