Constructing boundary-identical microstructures via guided diffusion for fast multiscale topology optimization

Jingxuan Feng, Lili Wang, Xiaoya Zhai, Kai Chen, Wenming Wu, Ligang Liu, Xiaoming FuComputer 2025 · DOI

TL;DR

This paper introduces a novel method for generating large-range, boundary-identical microstructure datasets using guided diffusion and active learning. This approach significantly accelerates multiscale topology optimization, enabling complex designs like mechanical cloaks to be achieved in under a minute.

What Problem Does This Paper Address?

Designing complex multiscale structures, where small-scale microstructures dictate the macroscopic properties of a material, is computationally intensive and faces a critical challenge: ensuring connectivity between these microstructures. Traditional methods struggle to generate diverse microstructures with consistent boundaries and a wide range of material properties, limiting the design space and efficiency of optimization. This paper tackles the problem of efficiently generating comprehensive microstructure datasets that overcome these limitations for faster and more effective multiscale topology optimization.

Methods

The core of this work lies in a guided diffusion model adapted for microstructure generation. This model is trained to produce microstructures that adhere to two key conditions: a predefined boundary and a specified homogenized elastic tensor. To address the "chicken-and-egg" problem of needing a dataset to train the model that generates the dataset, the authors employ a strategic approach:

  1. Initial Dataset Generation: A preliminary dataset of cubic symmetric microstructures (without the boundary-identical constraint) is created using optimization and perturbation methods.
  2. Boundary Clustering: These initial microstructures are clustered based on their boundaries to identify a set of representative "seed" boundaries.
  3. Guided Diffusion and Active Learning: The diffusion model is then used with these seed boundaries to generate diverse microstructures. An active learning strategy iteratively retrains the diffusion model with newly generated microstructures, progressively expanding the dataset's coverage of elastic moduli. This iterative process ensures both boundary consistency and a broad range of material properties.

Datasets

The paper introduces and utilizes several datasets:

  • Initial Microstructure Dataset ($D_0$): A dataset of 36,585 cubic symmetric microstructures generated via optimization and perturbation, serving as the starting point. This dataset does not enforce identical boundaries.
  • Boundary-Identical Microstructure Datasets: The core contribution is the generation of 16 large-range microstructure datasets, each containing over 100,000 microstructures. These datasets are characterized by identical boundaries and a wide range of elastic moduli. These are referred to as datasets 1-16 throughout the paper.

Key Insights and Ideas

  • Boundary-Identical Microstructures are Key: The paper strongly emphasizes that ensuring identical boundaries between microstructures is crucial for achieving mechanical compatibility and aligning with homogenization theory predictions.
  • Guided Diffusion for Controlled Generation: Using a diffusion model guided by both boundary and elastic tensor conditions allows for precise control over microstructure generation, leading to the desired properties.
  • Active Learning for Efficient Expansion: The iterative active learning strategy is vital for efficiently expanding the elastic tensor coverage of the datasets, allowing for the exploration of a wider design space.
  • Speedup is Dramatic: The ability to generate microstructures rapidly from these datasets leads to an unprecedented speedup in multiscale topology optimization, reducing design times from hours to under a minute.

Key Results

  • 16 Large-Range Datasets: The authors successfully generated 16 distinct, large-range microstructure datasets, each featuring identical boundaries and a wide spectrum of elastic moduli.
  • High Accuracy in Generation: The generated microstructures exhibit high R²-scores (average > 0.95) for elastic tensors and high degree of match (DoM) values (average > 0.98) for boundaries, demonstrating the model's accuracy.
  • Significant Optimization Speedup: Multiscale topology optimization tasks, such as designing a mechanical cloak, were completed in under one minute, a dramatic improvement over existing methods.
  • Improved Performance in Applications: Designs utilizing the boundary-identical microstructures showed better alignment with theoretical predictions (closer to Hashin-Shtrikman bounds) and reduced displacement errors in applications like mechanical cloaks.

Limitations and Future Work

  • 2D Focus: The current work focuses on 2D microstructures. The authors acknowledge that extending this to 3D microstructures is a significant future direction.
  • Computational Cost of 3D: Generating and classifying boundaries for 3D microstructures will demand substantially more computational resources and present new challenges in boundary categorization.
  • Dataset Size and Memory: While the current datasets are large, further expansion for 3D might require more advanced memory management and computational strategies.

Summary updated

BibTeX
@inproceedings{feng2025constructing,
  author = {Jingxuan Feng and Lili Wang and Xiaoya Zhai and Kai Chen and Wenming Wu and Ligang Liu and Xiaoming Fu},
  title = {{Constructing boundary-identical microstructures via guided diffusion for fast multiscale topology optimization}},
  booktitle = {Computer},
  year = {2025},
  doi = {10.1016/j.cma.2025.117735},
}
Figure