Akademik Veri Yönetim Sistemi
Processes, cilt.13, sa.7, 2025 (SCI-Expanded, Scopus)
This study aimed to identify the most reliable prediction model for estimating the compressive strength of concrete by conducting a comparative analysis of Particle Swarm Optimization (PSO) and Artificial Neural Network (ANN) methodologies. The modeling process utilized 92 experimental data points for training purposes and allocated 28 data points for testing validation. PSO was employed to optimize coefficients within mathematical equations used for concrete compressive strength prediction, facilitating the development of appropriate models based on various error metrics. Specifically, PSO models optimized to minimize Weighted Root Mean Square Error (WRMSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) criteria were evaluated against the highest-performing model developed using ANN. Model A, optimized using a constant term and the WRMSE loss function within a PSO-ANN framework, achieved the highest performance, with a correlation coefficient exceeding 0.99 and low error values on the training dataset. The same model also demonstrated strong predictive accuracy and low error on the test dataset, indicating excellent generalization capability. In contrast, the standalone ANN model exhibited near-perfect accuracy on the training data (R2 = 0.9994) but suffered a significant drop in performance on the test data (correlation ≈ 0.60). This highlights the impact of overfitting and underscores the importance of regularization techniques for improving generalizability. Through comprehensive statistical and visual assessments using Taylor diagram analysis, PSO-based models demonstrated significantly superior accuracy compared to the ANN model. Furthermore, the constant-term WRMSE model exhibited optimal generalization performance and provided the most reliable predictions among all tested models. It has been observed that highly accurate predictions can be made even for values outside the range of the data used. The results obtained in this study indicate that reliable predictive models for concrete production can be developed using both the available data and information from the literature. In cases where data are lacking, it is also possible to establish these models by conducting a sufficient number of experiments.