Neural message passing for predicting abnormal grain growth in Monte Carlo simulations of microstructural evolution

Abstract

Abnormal grain growth can significantly alter the properties of materials during processing. This can cause significant variation in the properties and performance of in-spec feedstock components subjected to identical processing paths. Understanding and controlling abnormal grain growth has proved to be elusive due to the stochastic nature of this phenomenon. However, recent advances in deep learning provide a promising alternative to traditional experimental and physics-based methods for understanding this phenomenon. Neural message passing allows deep learning to be applied to irregular inputs including graph representations of grain structures in a material. In this study we generate a large database of Monte Carlo simulations of abnormal grain growth in an idealized system. We apply message passing neural networks to predict the occurrence of abnormal grain growth in these simulations using only the initial state of the system as input. A computer vision model is also trained for the same task for comparison. The preliminary results indicate that the message passing approach outperforms the computer vision method and achieved 75% prediction accuracy, significantly better than random guessing. Analysis of the uncertainty in the Monte Carlo simulations provides a road map for ongoing work on this project.

Recent Advances and Applications of Deep Learning Methods in Materials Science

“…Similarly, Cohn and Holm present preliminary work applying GNNs to predict the occurrence of abnormal grain growth (AGG) in Monte Carlo simulations of microstructure evolution [293]. AGG appears to be stochastic, making it notoriously difficult to predict, control, and even observe experimentally in some materials.…”

Automated, high-accuracy classification of textured microstructures using a convolutional neural network

“…Machine learning (ML) methods use simulated and/or experimental data to learn relationships that may be indeterminate via traditional analysis methods, and for scientific discovery tasks where fundamental understanding of the underlying physical process still remains elusive Mjolsness and DeCoste (2001). In recent literature, ML methods have been used for microstructure quantification tasks which are challenging to accomplish with traditional data processing methods, including microstructure classification Gola et al (2019), identifying morphological features of interest such as dendrites Chowdhury et al (2016), and abnormal grain growth prediction from simulated data Cohn and Holm (2021), among others. Many ML methods use a training dataset to learn a non-linear mapping between inputs and outputs.…”