Concrete mix proportion design and performance prediction based on deep learning and multi-objective optimization

Abstract

[Objective] Concrete, as the cornerstone of national economic construction, necessitates the accurate prediction of its compressive strength for the design and safety of engineering structures. This study aims to predict concrete compressive strength using Deep Neural Network(DNN) models and proposes the RF-NSGA-II algorithm to optimize concrete mix proportions, achieving dual optimization of compressive strength and cost. [Methods] Fifteen DNN model architectures with different hidden layers and neuron numbers were constructed and evaluated for performance, selecting the best model. Hyperparameter optimization strategies and Bayesian optimization were employed to enhance the predictive performance of the DNN model. The performance of the DNN model was compared with Support Vector Regression(SVR) and Random Forest(RF) models. The RF-NSGA-II algorithm was used to optimize concrete mix proportions to meet strength requirements and cost control. [Results] The result showed that the optimal model had 3 hidden layers and 64 neurons(3L-64u). After optimization, the DNN model′s MAE and MSE decreased by 18% and 27%, respectively. Compared to the SVR and RF models, the optimized DNN model reduced MAE and MSE by 4% and 12%, and 11% and 15%, respectively. [Conclusion] Case validation demonstrated that the DNN3-L64u-BOP model′s predictions aligned well with experimental values, and the RF-NSGA-II algorithm effectively reduced costs while meeting engineering strength requirements. The Bayesian-optimized DNN model successfully predicted concrete compressive strength, and the RF-NSGA-II algorithm exhibited excellent performance in multi-objective optimization of concrete mix proportions, showing significant practical value in engineering applications.

Keywords

concrete / DNN / compressive strength / prediction / optimization / mechanical properties / influencing factors / deep learning