Conditional diffusion model for inverse prediction of process parameters and dendritic microstructures from mechanical properties

TL;DR

This work tackles the inverse design problem in polymeric thermoplastic resins by predicting processing temperature and microstructure from target mechanical properties. It integrates phase-field crystal growth, XFEM-based homogenization to obtain the elasticity matrix $\mathbf{D}$, and a conditional denoising diffusion probabilistic model (DDPM) with a spatiotemporal UNet to map $\mathbf{D}$ to a temperature pattern and a dendritic microstructure. The training data comprise microstructures generated at multiple crystallization temperatures, compressed images for ML, and $\mathbf{D}$ derived from XFEM; the model demonstrates accurate temperature and microstructure generation, even for conditions outside the training set. Enabled by this inverse-design pipeline, the approach can reduce trial-and-error in materials development and is extensible to other materials and property targets, offering a practical route to tailor processing and microstructure for desired mechanical performance.

Abstract

In this study, we develop a conditional diffusion model that proposes the optimal process parameters and predicts the microstructure for the desired mechanical properties. In materials development, it is costly to try many samples with different parameters in experiments and numerical simulations. The use of data-driven inverse design method can reduce the cost of materials development. This study develops an inverse analysis model that predicts process parameters and microstructures. This method can be used for any material, but in this study it is applied to polymeric material, which is the matrix resin of carbon fiber reinforced thermoplastics as an example. Matrix resins contain a mixture of dendrites, which are crystalline phases, and amorphous phases even after crystal growth is complete, and it is important to consider the microstructures consisting of the crystalline structure and the remaining amorphous phase to achieve the desired mechanical properties. Typically, the temperature during forming affects the microstructures, which in turn affect the macroscopic mechanical properties. The trained diffusion model can propose not only the processing temperature but also the microstructure when Young's modulus and Poisson's ratio are given. The capability of our conditional diffusion model to represent complex dendrites is also noteworthy.

Figures (22)

• Figure 1: Diagram of conditional diffusion model
• Figure 2: Analysis size of XFEM
• Figure 3:
• Figure 4: Diagram of (a) training and (b) generation of conditional diffusion model
• Figure 5: Structure of ST-UNet huang2023noise2music