A Bayesian-physical informed conditional tabular generative adversarial network framework for low-carbon concrete data augmentation and hyperparameter optimization

Data shortage, unbalanced data distribution and multi-factor coupling mechanism of materials all increase the difficulty of design. This paper proposed a physical constraint-conditional generative adversarial network (PI-CTGAN) to solve the above problems. Firstly, residual layers are added to the generator to enhance model stability. The continuous differentiable function and wasserstein_distance were constructed to embed physical loss functions into the generator, including water-cement ratio, supplementary cementitious materials (SCMs) ratio, and aggregate water absorption ratio. Based on this, Bayesian optimization (BO) was used to optimize the hyperparameters of PI-CTGAN. The results showed that BO effectively optimized the model's hyperparameters, reducing the total of Kolmogorov-Smirnov distribution (K-S tot ) of the generated dataset by 27.9 %. Additionally, applying physical loss to the optimized model can improve the model's data recognition capability, with generation accuracy increasing by 16.2 %. The influence of physical weight (W PI ) and activation functions on data generation quality was compared. Revealing that K-S tot initially decreased and then increased with W PI . The model using the rectified linear unit exhibited the best generation accuracy, with a K-S tot of 0.37 and anomaly data ratios (water-binder ratio and supplementary cementitious materials/binder ratio) of 11.67 % and 3.2 %, respectively. The generated data and the experimental data show statistical similarity and conform to the physical law. By constructing the target dataset and related physical constraints, the proposed PI-CTGAN can effectively solve the issues of multi-source data sets with data shortage and imbalance, thereby providing numerous datasets to guide engineering design.