[Paper Review] Data files for "Neural reparameterization improves structural optimization"

Stephan Hoyer|arXiv (Cornell University)|Sep 10, 2019

Topology Optimization in Engineering28 references51 citations

TL;DR

The paper reparameterizes structural optimization by optimizing a neural network that outputs material densities, yielding better designs across 116 tasks and 50% more often the best designs than baselines.

ABSTRACT

Structural optimization is a popular method for designing objects such as bridge trusses, airplane wings, and optical devices. Unfortunately, the quality of solutions depends heavily on how the problem is parameterized. In this paper, we propose using the implicit bias over functions induced by neural networks to improve the parameterization of structural optimization. Rather than directly optimizing densities on a grid, we instead optimize the parameters of a neural network which outputs those densities. This reparameterization leads to different and often better solutions. On a selection of 116 structural optimization tasks, our approach produces the best design 50% more often than the best baseline method.

Motivation & Objective

• Motivate how parameterization affects solution quality in structural optimization.
• Propose a CNN-based reparameterization to output element densities for topology optimization.
• Demonstrate the approach on 116 diverse structural optimization tasks.
• Compare CNN-based reparameterization against pixel-based and traditional baselines using gradient-based optimization.
• Analyze qualitative and quantitative gains and scalability across problem size.

Proposed method

• Reparameterize the optimization problem by replacing direct pixel densities with a CNN that outputs densities.
• Enforce volume and density constraints via a differentiable forward model with a constrained sigmoid mapping.
• Use implicit differentiation to backpropagate through the constraint transformation.
• Train the CNN parameters per task using a gradient-based optimizer (L-BFGS) to minimize compliance.
• Implement the forward pass as a differentiable TensorFlow graph with an adjoint-style physics solver.
• Compare CNN-LBFGS and Pixel-LBFGS against MMA and OC baselines on the 116 tasks.

Experimental results

Research questions

• RQ1 Does CNN-based reparameterization yield better structural designs than pixel-based parameterization and traditional optimization baselines?
• RQ2 How does the performance of neural reparameterization scale with problem size (small vs large grids)?
• RQ3 What qualitative differences in the resulting structures emerge under CNN parameterization compared with conventional methods?
• RQ4 Can the approach satisfy physical constraints (volume, densities) while improving design quality?

Key findings

• CNN-based reparameterization matches MMA on small problems and outperforms it on large problems.
• Across 116 tasks, CNN-LBFGS produced better designs than OC and Pixel-LBFGS in most cases.
• CNN-parameterized solutions tended to be simpler and showed fewer spiderweb artifacts than baselines.
• Large problems benefited more from CNN reparameterization due to multi-scale optimization effects and mesh-dependency considerations.
• Best-of-ensemble CNN designs were best or near-best in 99 of 116 tasks, outperforming MMA in qualitative examples.
• Reparameterization with neural networks often yielded qualitatively different design patterns, such as treelike pillars and fewer supports, compared to baseline structures.