Accurate prediction of concrete compressive strength based on explainable features using deep learning

Recently, a number of machine-learning models have been proposed for the prediction of 28-day compressive strength of concrete using constituent material information as inputs. These models required a series of unexplainable features to be pre-proportioned and predetermined via experiments. Therefore, the a priori knowledge and experience of concrete engineers in terms of concrete formulation and proportioning are unfortunately neglected and wasted in this prediction logic, which might lead to serious predictive errors in concrete design and construction. In this study, a deep-learning based "factors-to-strength" approach that considers multiple explainable features and therefore takes advantage of existing job-site proportioning information is presented for concrete strength prediction. A deep convolutional neural network is proposed and trained using a data set consisting of 380 groups of concrete mixes. The accuracy and reliability of the model are validated by comparing with three models - SVM, ANN, and AdaBoost - using a data set prepared experimentally. The results show that the proposed model achieves high coefficients of determination (0.973 for the training set and 0.967 for the test set), demonstrating its excellent accuracy and generalization ability. This new model also reveals the interplay between varying explainable features in determining the compressive strength of concrete, hence facilitating an interactive experience for engineers to maneuver familiar and understandable factors for concrete strength design.