Bayesian-Optimized Fully Connected Neural Network For Enhanced Prediction Accuracy In Concrete Compressive Strength Estimation

Abstract Concrete compressive strength serves as a critical quality control metric in construction engineering. Traditional empirical formulas and regression models struggle to capture multifactorial nonlinear relationships, while neural networks often exhibit hyperparameter sensitivity and overfitting issues. This study develops a Bayesian-optimized fully connected neural network (BOFCNN) that employs Gaussian process surrogate modeling to dynamically identify optimal hyperparameters (e.g., learning rate, neuron count), enhancing generalization capability. Utilizing 1,015 experimental datasets from the UCI repository, feature selection integrated Pearson-Spearman correlation with random forest importance ranking. The optimized model achieved state-of-the-art performance: MSE = 21.0745 MPa 2 , R² = 0.9080, MAE = 3.5028 MPa. Within the primary strength range (10–60 MPa), predictions showed strong agreement with experimental data, confirming the efficacy of Bayesian optimization in enhancing predictive accuracy. This establishes a robust computational framework for concrete strength prediction, enabling sustainable mix design optimization.