Intelligent Topology Optimization via Multi-Stage Contrastive Generative Adversarial Networks

Topology optimization faces challenges such as computational inefficiency and limited expressiveness in engineering material design. To enhance optimization accuracy while maintaining computational efficiency, we propose Multi-Stage Contrastive conditional Generative Adversarial Networks (MSC-GANs). Our model innovatively incorporates contrastive learning into the discriminator design, significantly improving the model’s ability to capture subtle structural features and relationships in complex topological structures. The approach utilizes physical information such as loads and volume fractions as input conditions and minimizes iteration losses through the generator network. By leveraging the contrastive learning mechanism, our method distinguishes high-quality structures from suboptimal ones more effectively than traditional discriminators. Experimental results demonstrate that MSC-GANs achieves a nearly $36 \%$ reduction in average prediction error (MSE) on the test set compared to traditional methods. The contrastive learning component proves particularly critical, with ablation studies showing significant performance degradation when this component is removed, while also enabling a $30 \%$ improvement in computational efficiency. Our approach provides new insights for intelligent design in CAE software through enhanced discrimination capabilities in generative models.