Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GrainPaint: A multi-scale diffusion-based generative model for microstructure reconstruction of large-scale objects

Nathan Hoffman, Cashen Diniz, Dehao Liu, Theron Rodgers, Anh Tran, Mark Fuge

TL;DR

GrainPaint addresses the challenge of generating large-scale, realistic microstructures for CAD-scale objects by leveraging diffusion-based outpainting (RePaint) to overcome fixed-generation-size limitations of previous ML approaches. The method is trained on SPPARKS-generated 3D grain datasets and uses a 3D U-Net with a multi-stage parallelization to assemble arbitrary geometries. It demonstrates qualitative similarity to SPPARKS across CAD-referenced objects and quantitative agreement for isotropic microstructures, and reveals some gaps for anisotropic cases. The work suggests GrainPaint can accelerate scale-appropriate microstructure generation for computationally expensive processes and CAD-driven applications, with opportunities for conditioning and multi-scale extensions.

Abstract

Simulation-based approaches to microstructure generation can suffer from a variety of limitations, such as high memory usage, long computational times, and difficulties in generating complex geometries. Generative machine learning models present a way around these issues, but they have previously been limited by the fixed size of their generation area. We present a new microstructure generation methodology leveraging advances in inpainting using denoising diffusion models to overcome this generation area limitation. We show that microstructures generated with the presented methodology are statistically similar to grain structures generated with a kinetic Monte Carlo simulator, SPPARKS.

GrainPaint: A multi-scale diffusion-based generative model for microstructure reconstruction of large-scale objects

TL;DR

GrainPaint addresses the challenge of generating large-scale, realistic microstructures for CAD-scale objects by leveraging diffusion-based outpainting (RePaint) to overcome fixed-generation-size limitations of previous ML approaches. The method is trained on SPPARKS-generated 3D grain datasets and uses a 3D U-Net with a multi-stage parallelization to assemble arbitrary geometries. It demonstrates qualitative similarity to SPPARKS across CAD-referenced objects and quantitative agreement for isotropic microstructures, and reveals some gaps for anisotropic cases. The work suggests GrainPaint can accelerate scale-appropriate microstructure generation for computationally expensive processes and CAD-driven applications, with opportunities for conditioning and multi-scale extensions.

Abstract

Simulation-based approaches to microstructure generation can suffer from a variety of limitations, such as high memory usage, long computational times, and difficulties in generating complex geometries. Generative machine learning models present a way around these issues, but they have previously been limited by the fixed size of their generation area. We present a new microstructure generation methodology leveraging advances in inpainting using denoising diffusion models to overcome this generation area limitation. We show that microstructures generated with the presented methodology are statistically similar to grain structures generated with a kinetic Monte Carlo simulator, SPPARKS.

Paper Structure

This paper contains 18 sections, 16 equations, 15 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Methodology
  3. Microstructure generation with SPPARKS
  4. Diffusion models
  5. Outpainting
  6. Model Design
  7. Parallelization Approach
  8. Segmentation
  9. Results
  10. CAD-based Microstructure Comparison
  11. Isotropic Microstructure Generation and Evaluation
  12. Anisotropic Microstructure Generation and Evaluation
  13. Discussion
  14. Conclusion
  15. Acknowledgment
  16. ...and 3 more sections

Figures (15)

  • Figure 2: A representation of the three steps of our microstructure generation process: Planning, Inpainting, and Segmentation
  • Figure 3: 3D U-net deep learning architecture used in this work.
  • Figure 4: Comparison of 128$\times$128 slices of a microstructure cube generated with different numbers of resamplings.
  • Figure 5: Microstructure generation plan. Blocks added in the current step are shown in blue and blocks added in previous steps are shown in yellow.
  • Figure 6: Example of the Results of the Segmentation Process.
  • ...and 10 more figures