Data-driven topology-fiber shape optimization method for CFRP-winding buckle arrestor based on MG-cGAN

Abstract

To enhance the performance of deep-sea pipeline CFRP-winding buckle arrestors, this paper innovatively proposes a CFRP arrestor through joint topology-fiber shape optimization (TFSO). For the high cost of traditional joint optimization, a Multi-Generator conditional Generative Adversarial Network (MG-cGAN) is proposed to enable rapid TFSO prediction without iteration under limited high-cost TFSO dataset. Considering CFRP arrestor’s structural characteristics, the Bi-directional Evolutionary Structural Optimization (BESO) and Nondominated Sorting Genetic Algorithm III (NSGA-III) methods are sequentially employed for topology optimization (TO) and fiber shape optimization (FSO) to yield an improved structure form. Next, MG-cGAN method is used to construct a TFSO prediction model. In offline phase, TO and FSO prediction models are developed using Enhanced Structural Optimization Prediction Residual Network (ESOP-ResNet) based on single-form optimization results. In online phase, a TFSO prediction model is developed by combining TO and FSO predictions, with the model outputs treated as fake and limited serial TFSO results treated as real for adversarial training. Case studies demonstrate that the jointed optimized CFRP arrestor achieves a 25 % increase in arresting efficiency while reducing 40 % volume. Furthermore, MG-cGAN, coupled with ESOP-ResNet, significantly enhances optimization efficiency while maintaining high prediction accuracy, avoiding the substantial cost of constructing large TFSO result datasets.