[Paper Review] Conditional diffusion-based microstructure reconstruction

Christian Düreth, Paul Seibert+4 more|arXiv (Cornell University)|Nov 24, 2022

Composite Material Mechanics26 citations

TL;DR

The paper investigates applying diffusion models to reconstruct real-world microstructure data, demonstrating that diffusion-based methods can capture diverse morphologies and work with small datasets.

ABSTRACT

Microstructure reconstruction, a major component of inverse computational materials engineering, is currently advancing at an unprecedented rate. While various training-based and training-free approaches are developed, the majority of contributions are based on generative adversarial networks. In contrast, diffusion models constitute a more stable alternative, which have recently become the new state of the art and currently attract much attention. The present work investigates the applicability of diffusion models to the reconstruction of real-world microstructure data. For this purpose, a highly diverse and morphologically complex data set is created by combining and processing databases from the literature, where the reconstruction of realistic micrographs for a given material class demonstrates the ability of the model to capture these features. Furthermore, a fiber composite data set is used to validate the applicability of diffusion models to small data set sizes that can realistically be created by a single lab. The quality and diversity of the reconstructed microstructures is quantified by means of descriptor-based error metrics as well as the Fréchet inception distance (FID) score. Although not present in the training data set, the generated samples are visually indistinguishable from real data to the untrained eye and various error metrics are computed. This demonstrates the utility of diffusion models in microstructure reconstruction and provides a basis for further extensions such as 2D-to-3D reconstruction or application to multiscale modeling and structure-property linkages.

Motivation & Objective

• Motivate microstructure reconstruction as a key task in inverse computational materials engineering.
• Assess the applicability of diffusion models to real-world microstructure data.
• Evaluate reconstruction quality across diverse and morphologically complex datasets.
• Test diffusion models on a small dataset from a fiber composite to reflect practical lab conditions.

Proposed method

• Create a highly diverse and morphologically complex microstructure dataset by combining and processing existing literature databases.
• Train diffusion models on the assembled dataset to learn microstructure priors.
• Validate reconstruction with descriptor-based error metrics and Fréchet Inception Distance (FID).
• Compare visual fidelity of generated samples to real data and assess generalization beyond training data.
• Discuss potential extensions such as 2D-to-3D reconstruction and multiscale modeling.

Experimental results

Research questions

• RQ1 Can diffusion models accurately reconstruct realistic, morphologically diverse microstructures from real-world data?
• RQ2 Do diffusion models perform well on small datasets typical of single-lab experiments (e.g., fiber composites)?
• RQ3 How do descriptor-based metrics and FID quantify reconstruction quality and diversity?
• RQ4 Are generated microstructures visually indistinguishable from real data to non-experts?
• RQ5 What are the prospects for extending diffusion-based reconstruction to 2D-to-3D and structure–property linkages?

Key findings

• Diffusion models can reconstruct realistic micrographs with high visual fidelity and capture diverse features not present in training data.
• The approach demonstrates applicability to small datasets typical of lab-scale data collection.
• Descriptor-based error metrics and FID are used to quantify quality and diversity of reconstructions.
• Generated samples are visually indistinguishable from real data to the untrained eye.
• The work provides a basis for extensions like 2D-to-3D reconstruction and multiscale modeling.