PaperlayerView exact revision on arXiv

Source-linked AI summary

3D Topology Optimization using Convolutional Neural Networks

Saurabh Banga, Harsh Gehani, Sanket Bhilare, Sagar Patel, Levent Kara

arXiv:1808.07440v1cs.LGphysics.comp-phstat.ML

TL;DR

Topology optimization is computationally demanding because finite element analysis is repeated as material distributions are evaluated. This paper uses a 3D CNN to predict final structures from intermediate optimization states, reporting reduced computation time and strong structural accuracy. The approach supplements conventional optimization rather than replacing it.

  • Problem

    Repeated finite element analysis makes topology optimization computationally demanding, and the data-driven framework remains dependent on conventional solvers and methods.

    Source: §1 · p. 1, §2 · p. 2

  • Method

    A 3D CNN maps intermediate topology-optimization densities and their gradients to final structures, with input combinations and training strategies compared.

    Source: §1 · p. 1, §4.3 · p. 7, §6 · p. 13

  • Results

    The best strategy reduced computation time by 40% over the conventional solver approach, with 96.2% average binary accuracy and 79.7% average RMS accuracy on the test set.

    Source: §6 · pp. 13–14

  • Takeaways & Limitations

    The CNN can predict final outputs from intermediate structural inputs, offering a way to reduce finite element calculations or provide starting points for conventional iterations.

    Source: §6 · p. 13

    The framework depends on the robustness of conventional solvers and methods.

    Source: §2 · p. 2

Abstract

from arXiv · show

Topology optimization is computationally demanding that requires the assembly and solution to a finite element problem for each material distribution hypothesis. As a complementary alternative to the traditional physics-based topology optimization, we explore a data-driven approach that can quickly generate accurate solutions. To this end, we propose a deep learning approach based on a 3D encoder-decoder Convolutional Neural Network architecture for accelerating 3D topology optimization and to determine the optimal computational strategy for its deployment. Analysis of iteration-wise progress of the Solid Isotropic Material with Penalization process is used as a guideline to study how the earlier steps of the conventional topology optimization can be used as input for our approach to predict the final optimized output structure directly from this input. We conduct a comparative study between multiple strategies for training the neural network and assess the effect of using various input combinations for the CNN to finalize the strategy with the highest accuracy in predictions for practical deployment. For the best performing network, we achieved about 40% reduction in overall computation time while also attaining structural accuracies in the order of 96%.

1. INTRODUCTION

Topology optimization relies on costly finite element analyses, and large-scale deployment remains constrained by computational time and resources. This work explores a data-driven 3D CNN that predicts final structures from intermediate optimization states, within a deliberately simplified problem scope.

1. INTRODUCTION

  • Finite element analysis makes topology optimization computationally demanding, limiting large-scale deployment through high computational complexity, resource needs, and time requirements.The analysis evaluates objective functions and iteratively guides material distribution.Source: §1 · p. 1
  • A 3D CNN predicts final structures from intermediate material-distribution solutions, complementing conventional optimization with a data-driven approach.The study uses a fixed discretized domain and one material type while varying point-force location, number, and magnitude, along with displacement boundary conditions.Source: §1 · p. 1
  • The proposed method is not a substitute for conventional optimization; it trades offline training-data costs and potential inaccuracy for fast predictions that may reduce finite-element calculations or seed downstream iterations.Once trained, CNN prediction is described as virtually instantaneous.Source: §1 · p. 1
  • The study focuses on density-based compliance minimization, representing a continuous Young’s-modulus field with pseudo-densities and minimizing compliance subject to a specified volume fraction.It investigates whether earlier optimization iterations can estimate the final structure.Source: §1 · p. 2

2. RELATED WORK

Prior topology-optimization research spans manufacturing-oriented methods, computational improvements, and data-driven predictions. This paper extends the latter line by evaluating a 3D CNN, varied inputs, and deployment strategies for mapping intermediate to optimized structures.

2. RELATED WORK

  • Prior topology-optimization work addresses additive manufacturing applications, alternative optimization methods, and numerical issues such as checkerboards and material islands.The cited work also considers support structures, lattice structures, uncertainties, mesh resolution, buckling, and temperature fields.Source: §2 · p. 2
  • The framework supplements rather than modifies established methods, and its performance therefore depends on the robustness of conventional solvers.The authors position it as an out-of-the-box efficiency approach built on conventional-process data and outputs.Source: §2 · p. 2
  • Earlier data-driven studies explored topology prediction with generative networks using load and displacement-boundary inputs [29], while Ulu et al. used PCA and a fully connected network to map loads to optimal topologies.Other work addressed force-location uncertainty and data-driven prediction challenges.Source: §2 · p. 2
  • Compared with the 2D encoder-decoder approach of Sosnovik and Oseledets [30], this study examines 3D structures, adds loads and boundary conditions as candidate inputs, and evaluates deployment choices and RMS accuracy.The paper presents a comparative analysis intended to guide computational strategy selection.Source: §2 · p. 3

3. THEORY

The section develops a SIMP-based minimum-compliance framework and a 3D CNN workflow that predicts a final structure from an intermediate optimization state. It also maps iteration-wise structural progress to identify a practical solver-to-network transfer point.

3. THEORY

  • SIMP seeks a physical density distribution that minimizes strain energy under prescribed supports and loads, subject to volume and density bounds.A neighborhood density filter smooths the design-variable distribution before optimization.Source: §3 · p. 3

4.1 Data Generation

  • The study generated 6,000 problem statements for a rectangular beam domain sized 24 x 12 x 12; each took about 70–100 conventional topology-optimization iterations to converge.The domain contains 3,456 elements.Source: §4.1 · pp. 5–6

4.2 Network Architecture

  • The 3D CNN uses convolutional encoder and decoder layers, with ReLU between layers and tanh at the output.Its loss combines binary cross-entropy against TopOpt ground truth with a volumetric-constraint penalty.Source: §4.2 · p. 7, Figure 5 · p. 7

4.3 Training the Network

  • Training compares density, inter-iteration density differences, and force and boundary-condition channels across datasets sampled from different SIMP stages.The 6,000 samples are split into 4,500 training, 500 validation, and 1,000 test samples; transfer iteration m is sampled uniformly or from Poisson distributions with λ = 5, 10, or 30.Source: §4.3 · pp. 7–8

4.4 Process Mapping

  • The process-mapping study evaluates 1,000 test samples using binary accuracy and a neighborhood-filtered spatial-gradient measure.The gradient measure compares each voxel’s density with a weighted neighborhood average; the direct density gradient was not useful for mapping spatial change.Source: §4.4 · p. 9, §4.4.1 · p. 9, §4.4.2 · p. 9

4.4.1 Binary Accuracy:

  • Binary accuracy exceeds 90% within the first fifth of SIMP progress, while later accuracy gains require significantly more time.The authors use the CNN after this initial structural resolution.Source: §4.4.1 · p. 9, Figure 6 · p. 9

4.4.2 Gradient Norm of neighborhood filtered density distribution:

  • The neighborhood-filtered spatial gradient tends toward zero after 20–30% of iterations, as local density changes continue without much change to the global structural outlook.Its change point nearly coincides with the point where binary accuracy exceeds 90%; the gradient can be computed incrementally, and transfer is set at a norm ≤ 0.05.Source: §4.4.2 · p. 10, Figure 7 · p. 10

5. RESULTS AND DISCUSSION

The experiments compare CNN input and training strategies for predicting optimized structures from intermediate topology-optimization data. Higher-iteration density inputs and nearby-iteration density gradients offer a strong accuracy–runtime trade-off.

5. RESULTS AND DISCUSSION

  • The study compared binary accuracy, based on equality of binarized output voxels, and RMS accuracy, based on differences between predicted and ground-truth float densities.Both metrics evaluate network predictions against the ‘TopOpt’ solver output.Source: §5 · pp. 10–11

5.1 Performance of networks trained on 4 datasets obtained through different input sampling strategies

  • Four training datasets of 6,000 samples each used uniform density-iteration sampling or Poisson sampling with λ = 5, 10, or 30.Their networks were evaluated on a common test set using iteration 20 for density input and iteration 15 for the spatial-gradient input.Source: §5.1 · p. 11
  • 96.2% binary accuracy and 79.7% RMS accuracy were achieved by the network trained with Poisson sampling at λ = 30, making it the best-performing network.Uniform sampling was only marginally inferior; datasets 2 and 3 were harder to learn from because their early-phase input structures were poorly resolved.Source: Figure 8 · p. 11, §5.1 · p. 12

5.2 Training on transformed data

  • A transformed dataset was formed by randomly rotating 40% of training samples along the x, y, and z axes, and its network was compared with one trained without transformation on the same test set.The passage describes the comparison setup but does not provide its outcome.Source: §5.2 · p. 12, Table 3 · p. 12

5.3 Investigating performance of varying combinations of density data, gradient data and boundary conditions as input

  • Density and gradient data alone formed the best input combination, while boundary conditions added to both did not appear beneficial.Omitting density data was detrimental, and boundary conditions alone performed worst because they lacked structural resolution.Source: §5.3 · p. 12, Table 4 · p. 12

5.4 Investigating performance of varying combinations of iterations for density data & gradient data as input

  • Nearly 40–45% average overall computation-time reduction was achieved using density data from iteration 20 and a gradient computed with respect to iteration 15, with reasonably good spatial accuracy.Higher density iterations improve performance but require longer solver runs; the chosen inputs balance spatial accuracy and computation time.Source: §5.4 · p. 13, Table 5 · p. 13, Table 6 · p. 13

6. CONCLUSIONS

The study finds that a 3D CNN can predict final topology-optimization outputs from intermediate structures, with the selected strategy reducing computation time while maintaining reported structural accuracy. The approach complements rather than replaces conventional optimization, and the paper identifies several directions for further work.

6. CONCLUSIONS

  • The findings support predicting final topology-optimization outputs from intermediate structures, with reported RMS and thresholded binary accuracies supporting the method’s applicability to 3D problems.The paper describes the predicted and solver outputs as identical after thresholding.Source: §6 · p. 13
  • The best test-set network uses density and density-gradient inputs, with density iterations sampled using a Poisson distribution with λ = 30.The authors report no random input augmentation because training and test accuracies differed only marginally.Source: §6 · p. 13
  • 40% less time than using the solver alone, with average test-set binary accuracy of 96.2% and RMS accuracy of 79.7%.The reported strategy uses spatial gradients of filtered densities to predict the final output from initial iterations.Source: §6 · pp. 13–14
  • Future work includes handling more complex problems beyond the MBB beam, incorporating forces and constraints, and checking predicted structures for islands.Other listed directions include user control of volume fraction, improved penalization, and feeding CNN outputs back into optimization.Source: §6 · p. 14
  • Table 7 compares ‘TopOpt’ [31] and CNN outputs at the 20th input iteration.The paper points to visual comparisons of predictions against conventional solver outputs.Source: Table 7 · p. 14, §6 · p. 14

A. Training progress

Training accuracy rises and loss falls across epochs, while the averaged plots smooth noisy, oscillatory SGD progress. The process mapping also indicates high binary accuracy early in optimization.

A. Training progress

B. Process Mapping Gradient Norm:

Across 1,000 test samples, the material-density gradient norm drops steeply in early iterations and then gradually tends toward zero as optimization converges.

B. Process Mapping Gradient Norm:

  • The gradient norm falls steeply during the first iterations, then changes gradually toward zero until convergence.Figure B.1 maps the Frobenius norm of iteration-wise density gradients against optimization progress for 1,000 samples.Source: B. Process Mapping Gradient Norm · p. 19, Figure B.1 · p. 19
  • The late-stage gradient norm does not capture the early high spatial resolution, because densities continue changing locally and marginally before convergence.The passage relates the trend to gradient-based algorithms such as Optimality Criteria and Sequential Quadratic Programming.Source: B. Process Mapping Gradient Norm · p. 19

C. Network Outputs

Appendix Table C.1 presents an exhaustive set of network predictions alongside solver ground truth and prediction-accuracy values.

C. Network Outputs

  • Table C.1 collects network predictions, solver ground truth, and network prediction accuracy across an exhaustive set of examples.Source: Table C.1 · p. 19
Loading 1808.07440v1…