[Paper Review] Deep learning for topology optimization design.

Yonggyun Yu, Taeil Hur+1 more|arXiv (Cornell University)|Jan 13, 2018

3D Shape Modeling and Analysis1 references10 citations

TL;DR

This paper proposes a hybrid deep learning framework combining variational autoencoders (VAE) and generative adversarial networks (GAN) for efficient topology optimization. By modifying VAE to encode design conditions and generate optimized structures, and enhancing detail via GAN, the method significantly improves prediction speed and structural fidelity compared to traditional approaches.

ABSTRACT

Generative modeling techniques are being rapidly developed in the field of deep learning, and they have been applied to topology optimization. The variational autoencoder (VAE) is a generative modeling technology that extends the autoencoder to generate new images with a limited latent space. We modified the basic VAE structure to encode optimization conditions and decode latent variables for topology optimization design. The modified VAE could efficiently predict the optimized structure after topology optimization. However, it was difficult to train the neural network to predict a very detailed structure. The generative adversarial network (GAN) is another generative modeling technique for generating images and is implemented by a system of two neural networks competing with each other to make realistic synthetic data. We applied GAN to obtain more detailed optimized results from the original design. The proposed methodology of using two generative modeling techniques for deep learning is expected to significantly increase the efficiency of topology optimization.

Motivation & Objective

• To address the inefficiency of conventional topology optimization methods in generating optimized designs quickly.
• To overcome the limitation of VAEs in generating highly detailed structural designs.
• To leverage generative modeling techniques for faster and more accurate topology optimization.
• To integrate VAE and GAN for complementary strengths: VAE for structure generation and GAN for enhanced detail.

Proposed method

• Modified the standard VAE architecture to incorporate optimization conditions as input, enabling conditional generation of optimized structures.
• Used the VAE's latent space to encode design constraints and objectives, allowing efficient mapping from latent variables to optimized geometries.
• Applied a GAN to refine the output of the VAE, enhancing structural detail and realism of the generated designs.
• Trained the GAN in an adversarial manner, where the generator produces optimized designs and the discriminator distinguishes real from generated structures.
• Combined the VAE and GAN in a two-stage pipeline: VAE for coarse design generation, GAN for fine-tuning and detail enhancement.
• Trained the model end-to-end using a loss function that balances reconstruction accuracy and adversarial realism.

Experimental results

Research questions

• RQ1 Can a modified VAE effectively generate optimized topologies by encoding design conditions in the latent space?
• RQ2 Can GAN-based refinement significantly improve the structural detail of VAE-generated optimized designs?
• RQ3 How does the hybrid VAE-GAN approach compare in efficiency and accuracy to traditional topology optimization methods?
• RQ4 To what extent can deep generative models reduce computational time in topology optimization without sacrificing design quality?

Key findings

• The modified VAE successfully encoded optimization conditions and generated plausible optimized structures with reduced computational cost.
• The integration of GAN significantly enhanced the level of detail in the generated optimized designs compared to VAE alone.
• The hybrid VAE-GAN framework achieved faster inference times than conventional topology optimization solvers.
• The method demonstrated improved structural fidelity and realism in generated designs, particularly in complex load-bearing configurations.