Multi-objective optimization of engineered cementitious composite based on machine learning and generative adversarial network

This study aims to establish a novel framework for mixture design optimization of engineered cementitious composite (ECC) by first collecting two original datasets of ECC's tensile stress and strain from the extensive and credible literature. The datasets comprise a wide range of variables including cementitious ingredients, 9 types of fiber and characteristics, admixtures, and experimental conditions. The data augmentation is then performed using a tuned constraints-modified Conditional Tabular Generative Adversarial Network (Tuned-CTGAN) to increase the model accuracy and generalizability. The fitness functions of tensile stress and strain are established based on four machine learning models with the hyperparameters tuned by the Hunger Games search (HGS) algorithm. After the data augmentation, the values of R2 in their test sets are increased from 0.874 to 0.925 and from 0.772 to 0.889, respectively. Subsequently, the third objective function (cost) is computed by polynomials and four classes of constraints (Min-max, volume, ratio, and fiber) are set up to define the variable's search space. A non-dominated sorting genetic algorithm based on reference-point strategy (NSGA-III) is introduced to optimize the mixture proportions of ECC by simultaneously optimizing tensile stress, tensile strain, and cost. This paper combines the results of data augmentation, model prediction, and multi-objective optimization for complex ECC design, which aims to provide a basis for practical application.