Source-linked AI summary
On the use of Artificial Neural Networks in Topology Optimisation
Rebekka V. Woldseth, Niels Aage, J. Andreas Bærentzen, Ole Sigmund
TL;DR
Topology optimisation is computationally expensive because iterative finite-element and optimisation procedures must repeatedly solve physics-based problems, motivating AI-based alternatives. This review categorises neural-network applications, evaluates their evidence and limitations, and identifies both problematic and promising directions. It finds that direct design often produces poor or restricted solutions, while progress depends on evaluating structural performance, computational savings, and robustness rather than relying on image similarity alone.
Problem
Topology optimisation requires costly iterative analysis, while AI research has produced many applications but few convincing breakthroughs and limited evidence of broad practical benefit.
Method
The paper critically reviews AI applications in topology optimisation, analyses their evaluation practices and limitations, and formulates recommendations for future research.
Results
Direct-design models are reported to produce poor designs, require expensive training data, and remain restricted in problem variety and mesh resolution, while image-based errors can miss structural-performance differences.
Takeaways & Limitations
Meaningful progress should be demonstrated through solution quality, actual computational speed-up, broader viable problem scope, and task-specific reliability.
Takeaways & Limitations
Neural-network approaches can fail unpredictably outside their training distribution, and small boundary-condition changes may produce substantially different optimal structures.
Abstract
from arXiv · showhide
The question of how methods from the field of artificial intelligence can help improve the conventional frameworks for topology optimisation has received increasing attention over the last few years. Motivated by the capabilities of neural networks in image analysis, different model-variations aimed at obtaining iteration-free topology optimisation have been proposed with varying success. Other works focused on speed-up through replacing expensive optimisers and state solvers, or reducing the design-space have been attempted, but have not yet received the same attention. The portfolio of articles presenting different applications has as such become extensive, but few real breakthroughs have yet been celebrated. An overall trend in the literature is the strong faith in the "magic" of artificial intelligence and thus misunderstandings about the capabilities of such methods. The aim of this article is therefore to present a critical review of the current state of research in this field. To this end, an overview of the different model-applications is presented, and efforts are made to identify reasons for the overall lack of convincing success. A thorough analysis identifies and differentiates between problematic and promising aspects of existing models. The resulting findings are used to detail recommendations believed to encourage avenues of potential scientific progress for further research within the field.
1 Introduction
Topology optimisation distributes material within constrained domains to improve structural performance, but iterative PDE-based analysis makes large and detailed problems computationally expensive. Artificial neural networks have therefore attracted interest as data-driven approximators, although their capabilities and limitations must be assessed carefully.
- Topology optimisation: Topology optimisation distributes material within a defined domain under physical and geometric constraints to maximise structural performance.Its increased design freedom has supported both academic development and industrial use.
- Computational challenge: Conventional topology optimisation repeatedly maps problem characteristics to structures while solving governing partial differential equations at each iterate.The nested analysis-and-design procedure requires intermediate state solutions before continuing optimisation.
- Computational challenge: Increasing problem size, complexity, and finite-element resolution raises computational cost and limits topology optimisation for real-life designs.Solution accuracy and detail depend strongly on element size in finite-element analysis.
- Artificial intelligence: Machine learning uses data-driven function approximations for difficult input-target mappings, while deep learning and neural networks have shown strong pattern-recognition capabilities.Image-analysis advances helped motivate applying neural networks to topology optimisation because regular finite-element meshes resemble pixel grids.
- Review scope: The review examines AI applications in topology optimisation while noting that rapidly expanding literature may prevent complete coverage, despite aiming for a representative selection.It introduces relevant machine-learning concepts and organises the reviewed specialised models around their topology-optimisation applications.
- Neural networks: Neural networks map inputs through layered nonlinear functions whose architectures, weights, biases, and activation functions are learned from data.Common architectures include feedforward networks, convolutional networks for regular grids, and generative models such as variational autoencoders.
2 Literature Review
AI applications in topology optimisation were initially driven by deep-learning success in image analysis and generation. The dominant aim has been to replace iterative optimisation with direct prediction of structural images from problem descriptions.
- Motivation: Deep-learning success in image analysis and generation motivated neural-network models intended to produce topology-optimised structures directly from problem-descriptive inputs.These approaches seek to replace conventional iterative optimisation with iteration-free prediction.
2.1 Overview
The reviewed AI applications in topology optimisation fall into five categories spanning direct prediction, computational assistance, structure modification, compact representations, and multiple design generation. Each category targets a distinct way of changing the conventional optimisation workflow.
- Classification: The review further organises the five groups into more specific subcategories to describe trends across the literature.Table 1 provides the detailed sorting of reviewed articles.
- Direct design: Direct design predicts an optimal structure from problem characteristics with the aim of achieving iteration-free topology optimisation.The intended output is an optimised structure produced directly rather than through repeated optimisation iterations.
- Acceleration: Acceleration supplements conventional iterative methods by approximating finite-element analysis or skipping selected optimisation iterations.The goal is to reduce computational cost while retaining the conventional optimisation framework.
- Post-processing: Post-processing modifies conventionally optimised or homogenised structures to improve manufacturability or replace de-homogenisation steps.Examples include shape changes, microstructure configuration, boundary smoothing, and manufacturability-oriented adjustments.
- Reduction: Reduction reparameterises topology with a compact model, decreasing design variables and speeding iterative optimisation.The approach resembles model-order reduction but learns the representation rather than explicitly programming it.
- Design diversity: Design diversity generates multiple candidate structures for the same topology-optimisation problem, providing alternatives with different desired characteristics.This category is related to exploring a Pareto front in multi-objective optimisation.
2.2 Categorisation
The review categorises neural-network applications in topology optimisation and identifies recurring constraints in direct design, acceleration, reduction, and post-processing approaches. Across these categories, computational cost, limited generalisation, structural-performance mismeasurement, and mesh dependence restrict current progress, although some approaches show promise.
- Direct design: Mesh dependence increases network parameters, memory use, training cost, and structural complexity as resolution rises; reviewed models typically use fewer than 4,000 elements, at most 26,000, versus two billion in state-of-the-art topology optimisation.The review also finds that published direct-design models do not fully exploit CNN generality across input image dimensions.
- Direct design: Image-reconstruction losses can overlook structural disconnections: MAE stayed below 0.4% while compliance more than doubled after thresholding the black-and-white designs.The example shows that pixel-wise similarity does not reliably represent structural performance.
- Direct design: Thresholding can expose full disconnections hidden by grey-scale representations, so solid-void designs are advised for fair structural comparisons.The review reports that thresholding may substantially alter compliance and reveal the true structural performance.
- Direct design: Direct-design models require many optimised training structures, forcing most studies toward coarse meshes and limited boundary-condition variation.Generating one conventional topology-optimisation solution for each training case makes broad, high-resolution datasets expensive.
- Acceleration: Acceleration methods can reduce computational effort through multi-level mappings, but coarse meshes impose fine-grid length-scale constraints that may offer little advantage over interpolation.The reported online approach shows promising fine-grid sensitivity accuracy and speed-up, while the shared two-scale limitation reduces the distinct benefit of ANNs.
- Reduction and post-processing: Reduction and post-processing applications remain less established: deep representation learning supports detailed 3D geometries, whereas CAD extraction with machine learning is described as under-studied.The review presents detailed-geometry representation as encouraging further exploration but notes limited comparative evaluation and unresolved post-processing needs.
3 Assessments
The review assesses AI-based topology-optimisation frameworks by balancing computational cost, applicability, and solution quality. It finds promising approaches, but also substantial limitations in generalisation, evaluation, training cost, and reported evidence.
- 3.1 Computational cost and applicability: Evaluation should consider generalisation ability and computational cost before structural solution quality, because expensive, narrowly applicable methods may provide no practical benefit.The review argues that solution quality is decisive only when a method offers a sufficient balance of speed-up and generality.
- 3.1 Computational cost and applicability: Training-data generation can eliminate claimed speed-ups: Nakamura and Suzuki [2020] used 333,000 optimised structures, requiring at least 333,001 comparable applications to break even.This issue applies to direct-design models trained with supervised or semi-supervised target samples.
- 3.1 Computational cost and applicability: Large or narrow training sets may encourage memorisation rather than learning, producing infeasible predictions when test loads shift beyond the represented examples.For Yan et al. [2022], some test predictions omitted material at applied loads, suggesting failure to learn the significance of load positions.
- 3.1 Computational cost and applicability: Generality is judged subjectively across mesh dependence, problem-definition variety, and training–test similarity, with higher scores requiring broader applicability.The review emphasizes different boundary conditions, mesh dimensions and resolutions, and diverse loading conditions as key dimensions.
- 3.1 Computational cost and applicability: Direct-design methods generally combine high breakeven thresholds with low generalisation, whereas acceleration and upscaling methods range from similarly poor to near-zero-threshold, highly general approaches.The best-scoring example is Chi et al., with a breakeven threshold close to zero and a generality score of 36 through online learning without pre-training cost.
- 3.1 Computational cost and applicability: Reduction-based methods can achieve broad applicability through online training and reparameterisation, but many are excluded from comparisons because their computational costs are insufficiently reported.Zhang et al. [2021b] applied such a framework to compliance, stress-constrained, compliant-mechanism, heat-conduction, and nonlinear-elasticity problems.
- 3.2 Solution quality: Solution-quality comparisons are weakened because most papers omit fair quantitative metrics, while averages or illustrative examples can conceal poor outliers and training–test overlap.The review therefore calls for evaluation that covers the full performance distribution rather than selected examples alone.
- 3.2 Solution quality: Reduced geometric representations may smooth structures and remove fine features, although deeper networks or post-processing can partly recover detail and improve manufacturability.The review notes that filtering conventional SIMP results can also reduce fine features, so this advantage is not unique to reduction methods.
4 AI limitations
The review identifies brittleness, limited physical understanding, and fragile generalisation as central limitations of current neural-network approaches to topology optimisation. Small changes in loading conditions can produce substantially different optimal structures, making apparently successful mappings unreliable outside their learned settings.
- Model brittleness: Current deep-learning models can fail unpredictably on unfamiliar domains or after small input perturbations, often with high-confidence incorrect predictions.The review links this brittleness to models learning statistical mappings rather than human-like understanding of salient features.
- Direct design limitations: Generative and convolutional approaches may map boundary conditions to structures through a low-dimensional latent representation that cannot capture the full problem-to-design relation.The review argues that direct mappings from problem conditions to mechanical structures are therefore fundamentally problematic for general topology optimisation.
- Sensitivity to boundary conditions: A load-angle perturbation collapses the first optimised structure, whereas a structure with an additional thin support bar performs similarly under both load cases.The two load cases differ only slightly: Fx=0, Fy=1 versus Fx=0.0099998, Fy=0.99995.
- Physical understanding: Using iterative element-density histories can produce plausible structures for many problems, but the approach cannot guarantee that those histories encode the physical properties needed for all cases.The review presents this as another consequence of neural networks lacking physical understanding.
- Evaluation concerns: Reported performance is often difficult to assess because studies may use restricted training and test distributions, image-based metrics, or incomplete benchmark comparisons.These practices can hide poor outlier cases and fail to measure structural performance adequately.
5 Recommendations
The review recommends designing AI-aided topology-optimisation methods around realistic computational benefits, physically meaningful representations, and rigorous evaluation. It particularly emphasises generalisation beyond training cases, fair comparisons with conventional methods, and benchmarks that reflect practical complexity.
- Motivation and scientific value: AI methods should reduce memory or computational cost without compromising the other, and should expand the size, complexity, or speed of viable applications.The review treats convincing comparisons with state-of-the-art methods and demonstrated problem coverage as necessary for scientific progress.
- When designing the model: Model inputs should encode all problem properties relevant to the output, including relational information about the environment in which the model operates.The review cautions that neural networks should not be expected to infer such relationships magically.
- Data and loss design: Training-data generation should be counted as a substantial computational cost when assessing whether an AI framework becomes more efficient than conventional topology optimisation.The review connects data formatting, breakeven analysis, and model versatility to the overall computational assessment.
- Data and loss design: Loss functions should reflect physical properties and use FEA-based evaluation more than image-based prediction error, while models should support adaptable settings and inexpensive post-processing.The recommended direction is away from mesh- and boundary-condition-specific models toward frameworks that handle more problem types without retraining.
- Evaluation and benchmarking: Testing should report representative performance distributions, use structurally meaningful metrics, and compare against conventional methods under matched length scales, thresholds, and test conditions.Random test cases drawn from the same restricted pool as training data are insufficient for assessing generalisation.
- Benchmark cases: For sub-task models, evaluation should cover both the specific prediction and the resulting optimisation, including sensitivity accuracy or intermediate designs where relevant.Otherwise, apparent speed-ups may arise from simple pixel rounding or moving-average changes rather than a superior learned method.
- Benchmark cases: Benchmark problems should extend beyond overly simple linear-volume-constraint formulations because such formulations simplify learning and avoid variation in material properties and final volume fractions.The review notes that direct-design learning becomes harder when volume fractions vary and constrained responses require quantitative evaluation.
- Benchmark cases: The ultimate goal should be outperforming conventional methods on practical problems or enabling new, more complex problems, rather than stopping at proof-of-concept demonstrations.The review acknowledges proof-of-concept value but reserves claims of scientific progress for broader demonstrated capability.
6 Conclusion
The review finds that most AI approaches for topology optimisation have not delivered convincing progress, especially direct iteration-free models. It identifies more promising directions in accelerating iterative computation, post-processing, and critical evaluation, while recommending transparent assessment and better understanding of AI capabilities.
- Current status for AI in TO: Direct iteration-free topology optimisation models often produce poor designs, are expensive to obtain, and support limited problem and mesh variety.The review argues that inter-iteration computations, rather than iteration itself, are the main obstacle to large-scale topology optimisation.
- Promising applications: The review identifies acceleration of iterative optimisation and post-processing of optimised structures as more promising application areas than direct design models.Suggested acceleration routes include reducing finite-element analyses, partially replacing them, and using machine learning for coarse-to-fine approximations.
- Conclusion: The review concludes that AI research in topology optimisation is at an early stage: some applications are promising, but many studies hold unrealistic expectations about what models can learn.The authors call for stronger knowledge of AI capabilities and more critical interpretation of results.
- Recommendations: The review recommends evaluating AI systems through transparent reporting of achieved performance, generalisation, worst-case behaviour, computational gains, and relevant benchmarks.It presents six questions and treats breakeven thresholds, generalisation ability, solution quality, and fair comparisons as minimum assessment requirements.
- Future promise: Physics-informed neural networks may approximate governing partial differential equations with fewer training samples and potentially support more efficient or accurate substitutes for finite-element analysis.Their combination with topology optimisation remains outside the review’s scope, and the authors caution that optimisation can prioritise numerical errors over physics.
- Promising applications: Large-scale topology optimisation may benefit from re-parameterising models to reduce design variables or structural information, provided finite-element analysis is also removed.The review also notes that manufacturing-oriented post-processing and de-homogenisation remain underrepresented opportunities.
A Fig. 6 - Upscaling procedure
The procedure first generates a coarse topology-optimised design, then resizes it to a finer grid and applies volume-preserving thresholding to obtain a discrete structure.
- Upscaling procedure: A SIMP-derived coarse design is resized from a 60×20 grid to a 120×40 fine grid using bicubic interpolation.The coarse design uses volume fraction 0.5, penalisation 3.0, and minimum filter radius 2.0.
- Upscaling procedure: Volume-preserving thresholding sorts fine-grid densities and assigns solid or near-void values to produce the discrete structure.The threshold retains the prescribed volume fraction while setting selected entries to 1 and the remainder to 1e-9.