Source-linked AI summary
Deep learning for determining a near-optimal topological design without any iteration
Yonggyun Yu, Taeil Hur, Jaeho Jung, In Gwun Jang
TL;DR
Topology optimization is computationally costly because it uses many design variables and iterations. The paper proposes a deep-learning pipeline that predicts near-optimal designs from boundary conditions and optimization settings, with refinement across resolutions, and reports near-optimal pixel values and compliance with negligible computational cost.
Problem
Topology optimization can be computationally expensive because it involves many design variables and iterations.
Method
The method trains a low-resolution CNN encoder-decoder and connects it to a two-stage cGAN refinement using paired low- and high-resolution optimized structures.
Results
The integrated network determines near-optimal structures in terms of pixel values and compliance with negligible computational cost.
Takeaways & Limitations
The approach predicts near-optimal topological designs for given boundary conditions and optimization settings without iteration.
Takeaways & Limitations
The study considers simplified and limited boundary conditions and optimization settings, constraining its scope for further real-world applications.
Abstract
from arXiv · showhide
In this study, we propose a novel deep learning-based method to predict an optimized structure for a given boundary condition and optimization setting without using any iterative scheme. For this purpose, first, using open-source topology optimization code, datasets of the optimized structures paired with the corresponding information on boundary conditions and optimization settings are generated at low (32 x 32) and high (128 x 128) resolutions. To construct the artificial neural network for the proposed method, a convolutional neural network (CNN)-based encoder and decoder network is trained using the training dataset generated at low resolution. Then, as a two-stage refinement, the conditional generative adversarial network (cGAN) is trained with the optimized structures paired at both low and high resolutions, and is connected to the trained CNN-based encoder and decoder network. The performance evaluation results of the integrated network demonstrate that the proposed method can determine a near-optimal structure in terms of pixel values and compliance with negligible computational time.
1. Introduction
Topology optimization can produce feasible conceptual designs, but its many design variables and iterative procedures create substantial computational cost. This motivates deep-learning methods that use boundary conditions and optimization settings to predict near-optimal structures without iteration.
- Topology optimization determines material distributions under boundary conditions to optimize system performance while satisfying design constraints.
- Large numbers of design variables and optimization iterations make topology optimization computationally expensive.
- Earlier machine-learning approaches reduced design-variable dimensionality, approximated sensitivities, or mapped intermediate structures to optimized structures.
- A prior deep-learning topology-optimization model did not incorporate boundary conditions and optimization settings essential for performing topology optimization.
- The proposed method learns from paired boundary-condition, optimization-setting, and optimized-structure data, then applies two-stage GAN refinement to predict near-optimal designs without iteration.
2. Overview of machine learning and deep learning
Machine learning uses trained datasets to support prediction or reasoning, while deep learning is a machine-learning subfield based on artificial neural networks. CNNs process image data, and generative models learn data distributions to produce new data, with GANs using competing generators and discriminators.
- Machine learning enables trained systems to predict, judge, or reason about new input data.
- Deep learning is a part of machine learning that stems from artificial neural networks inspired by biological neural networks.
- Artificial neural networks consist of nodes that transform multiple inputs into outputs through activation functions, with training adjusting weights to minimize a loss function.
- Convolutional neural networks train filter parameters with the classifier and use convolution operations to extract features from image data.
- Generative models approximate input-data probability distributions and use latent variables to generate new data.
- GANs train a generator and discriminator adversarially: the generator produces fake data, while the discriminator estimates whether inputs are real.
3. Proposed Method
The proposed method predicts near-optimal topological designs without iterative optimization by combining a low-resolution CNN encoder–decoder with a high-resolution cGAN refinement stage.
- Step 3: cGAN refinement: The low-resolution CNN output is refined to high resolution by a conditional GAN trained on paired optimized structures.The cGAN uses low-resolution optimized structures as conditional variables and generates predicted high-resolution structures without latent random variables.
- Step 4: Integrated network: The integrated network connects the trained CNN encoder–decoder and cGAN to predict a near-optimal high-resolution structure without iteration.The proposed pipeline uses a two-stage refinement process to determine higher-resolution structures from the learned low-resolution output.
- Step 2: CNN encoder–decoder: The CNN encoder–decoder learns low-resolution topology optimization from encoded force, geometric boundary, passive-zone, and scalar conditions.The network encodes boundary conditions into latent variables and directly inputs mass fraction, then reconstructs relative-density structures.
- Step 4: Performance evaluation: The proposed method requires only 0.06% of the computational time of same-resolution topology optimization at 128 × 128 resolution.High-resolution predicted structures had pixel-wise MAE below 2.72% and mass fraction error below 0.51%.
- Step 4: Performance evaluation: Visually similar structures can have different compliances, motivating future networks that directly consider structural connectivity or strength.The paper identifies compliance-related error as a limitation of relying on visually and pixel-wise similar structures.
4. Conclusions
The proposed deep learning method predicts near-optimal topological designs for given boundary conditions and optimization settings without iteration. It combines CNN-based encoding and decoding with cGAN refinement, achieving near-optimal pixel values and compliance at negligible computational cost while retaining important scope limitations.
- Contribution: The method determines a near-optimal topological design for given boundary conditions and optimization settings without iteration.It is presented as a deep learning-based design method.
- Method: Boundary-condition images are encoded into latent variables, decoded into near-optimal structures, and refined to higher resolution using a connected cGAN.The workflow uses a CNN-based encoder-decoder followed by progressive cGAN refinement.
- Results: Numerical examples show near-optimal structures in terms of pixel values and compliance with negligible computational cost.The evaluation demonstrates the method’s feasibility for efficient structure prediction.
- Limitations: Larger finite element models make it more difficult to prepare sufficiently large training datasets for larger ANNs.The paper notes that current GANs handle maximum output sizes of 1024 × 1024 for two-dimensional cases and 64 × 64 × 64 for three-dimensional cases.
- Limitations: The study considers only simplified, limited boundary conditions and optimization settings, so more general conditions require more datasets and larger ANNs.This is identified as a limitation for further real-world applications.
- Future directions: The method currently handles fixed regular meshes, although recent deep learning studies suggest that non-Euclidean meshes could address this limitation.The paper also proposes combining predicted structures with conventional topology optimization to improve convergence and reduce computational burden.