Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting Grain Growth in Polycrystalline Materials Using Deep Learning Time Series Models

Eliane Younes, Elie Hachem, Marc Bernacki

TL;DR

This work addresses the predictive challenge of grain growth in polycrystalline materials by combining mean-field descriptors derived from validated TRM simulations with deep time-series forecasting models. It systematically compares RNN, LSTM, TCN, and Transformer architectures for recursive forecasting of evolving grain-size distributions, finding that the LSTM offers the best accuracy and long-horizon stability. The approach delivers substantial speed-ups over full-field simulations and maintains physical fidelity, making it well suited for real-time digital-twin and process-optimization workflows. The study highlights practical guidelines, such as a 30-bin histogram representation and an 80:15:5 data split, and points to future work including non-isothermal conditions and second-phase particle effects to extend realism.

Abstract

Grain Growth strongly influences the mechanical behavior of materials, making its prediction a key objective in microstructural engineering. In this study, several deep learning approaches were evaluated, including recurrent neural networks (RNN), long short-term memory (LSTM), temporal convolutional networks (TCN), and transformers, to forecast grain size distributions during grain growth. Unlike full-field simulations, which are computationally demanding, the present work relies on mean-field statistical descriptors extracted from high-fidelity simulations. A dataset of 120 grain growth sequences was processed into normalized grain size distributions as a function of time. The models were trained to predict future distributions from a short temporal history using a recursive forecasting strategy. Among the tested models, the LSTM network achieved the highest accuracy (above 90\%) and the most stable performance, maintaining physically consistent predictions over extended horizons while reducing computation time from about 20 minutes per sequence to only a few seconds, whereas the other architectures tended to diverge when forecasting further in time. These results highlight the potential of low-dimensional descriptors and LSTM-based forecasting for efficient and accurate microstructure prediction, with direct implications for digital twin development and process optimization.

Predicting Grain Growth in Polycrystalline Materials Using Deep Learning Time Series Models

TL;DR

This work addresses the predictive challenge of grain growth in polycrystalline materials by combining mean-field descriptors derived from validated TRM simulations with deep time-series forecasting models. It systematically compares RNN, LSTM, TCN, and Transformer architectures for recursive forecasting of evolving grain-size distributions, finding that the LSTM offers the best accuracy and long-horizon stability. The approach delivers substantial speed-ups over full-field simulations and maintains physical fidelity, making it well suited for real-time digital-twin and process-optimization workflows. The study highlights practical guidelines, such as a 30-bin histogram representation and an 80:15:5 data split, and points to future work including non-isothermal conditions and second-phase particle effects to extend realism.

Abstract

Grain Growth strongly influences the mechanical behavior of materials, making its prediction a key objective in microstructural engineering. In this study, several deep learning approaches were evaluated, including recurrent neural networks (RNN), long short-term memory (LSTM), temporal convolutional networks (TCN), and transformers, to forecast grain size distributions during grain growth. Unlike full-field simulations, which are computationally demanding, the present work relies on mean-field statistical descriptors extracted from high-fidelity simulations. A dataset of 120 grain growth sequences was processed into normalized grain size distributions as a function of time. The models were trained to predict future distributions from a short temporal history using a recursive forecasting strategy. Among the tested models, the LSTM network achieved the highest accuracy (above 90\%) and the most stable performance, maintaining physically consistent predictions over extended horizons while reducing computation time from about 20 minutes per sequence to only a few seconds, whereas the other architectures tended to diverge when forecasting further in time. These results highlight the potential of low-dimensional descriptors and LSTM-based forecasting for efficient and accurate microstructure prediction, with direct implications for digital twin development and process optimization.

Paper Structure

This paper contains 15 sections, 13 equations, 13 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset Description
  4. Data Preprocessing
  5. Time Series Forecasting Strategy
  6. Deep Learning Frameworks
  7. Recurrent Neural Networks (RNN)
  8. Long Short-Term Memory (LSTM)
  9. Transformer Networks
  10. Temporal Convolutional Network (TCN)
  11. Experimental Setup and Training
  12. Hyperparameters
  13. Evaluation Metrics
  14. Results and Discussion
  15. Conclusion

Figures (13)

  • Figure 1: Pipeline for generating microstructure evolution sequences using Laguerre-Voronoï tessellation (Lavogen) and TRM simulations.
  • Figure 2: Recursive Forecasting Using a Fixed-Length Sliding Window
  • Figure 3: Structure of the RNN
  • Figure 4: Computation process describing the RNN
  • Figure 5: Computation process involved in an LSTM
  • ...and 8 more figures