Fully Convolutional Spatiotemporal Learning for Microstructure Evolution Prediction
Authors: Michael Trimboli, Mohammed Alsubaie, Sirani M. Perera, Ke-Gang Wang, Xianqi Li
Organizations: Department of Mathematics and Systems Engineering, Florida Institute of Technology, Melbourne, FL, 32901, USA · Department of Mathematics, Embry-Riddle Aeronautical University, Daytona Beach, FL, 32114, USA · Department of Mechanical and Civil Engineering, Florida Institute of Technology, Melbourne, FL, 32901, USA
Understanding and predicting microstructure evolution is fundamental to materials science, as it governs the resulting properties and performance of materials. Traditional simulation methods, such as phase-field models, offer high-fidelity results but are computationally expensive due to the need to solve complex partial differential equations at fine spatiotemporal resolutions. To address this challenge, we propose a deep learning-based framework that accelerates microstructure evolution predictions while maintaining high accuracy. Our approach utilizes a fully convolutional spatiotemporal model trained in a self-supervised manner using sequential images generated from simulations of microstructural processes, including grain growth and spinodal decomposition. The trained neural network effectively learns the underlying physical dynamics and can accurately capture both short-term local behaviors and long-term statistical properties of evolving microstructures, while also demonstrating generalization to unseen spatiotemporal domains and variations in configuration and material parameters. Compared to recurrent neural architectures, our model achieves state-of-the-art predictive performance with significantly reduced computational cost in both training and inference. This work establishes a robust baseline for spatiotemporal learning in materials science and offers a scalable, data-driven alternative for fast and reliable microstructure simulations.
Figures & tables
Fig. 1 : Flowchart of the proposed fully convolutional spatiotemporal framework for microstructure evolution prediction. An input sequence of length T is first processed by a convolutional spatial encoder to extract microstructural spatial correlations and project the data into a compact latent feature space. The latent representations are then passed through stacked Gated Spatiotemporal Attention (gSTA) modules, which serve as the temporal translator of the network and model the dynamic evolution patterns of the microstructure. Within this module, the encoded features are propagated and transformed to generate forecast frames corresponding to future time steps T+1 through T+T′ . Finally, a convolutional spatial decoder reconstructs the predicted microstructure fields by upsampling the latent representations back to the original spatial resolution, producing the output sequence.
Fig. 2 : Grain growth prediction (10 input frames, 90 output frames) using the proposed fully convolutional spatiotemporal model: Two representative test samples are shown (indexed numerically). (A) Predicted and ground-truth microstructure frames at t=11,25,50,75, and 100 ; (B) RMSE and SSIM evaluated every 5 frames over the prediction horizon; (C–D) Grain-size distributions (GSDs) for prediction and ground truth at t=25 and t=100 , respectively.
Fig. 3 : Accuracy metrics on grain growth prediction (10 input frames, 90 output frames): (A) Dataset-averaged RMSE and SSIM evaluated every 5 frames over the prediction horizon; (B) Temporal evolution of the mean grain area for predicted and ground-truth sequences (computed every 10 frames); (C) Corresponding linear regression fits highlighting agreement in coarsening trends between prediction and ground truth.
Fig. 4 : Grain Growth prediction (10 input frames, 190 output frames) using the proposed fully convolutional spatiotemporal model: Two representative test samples are shown (indexed numerically). (A) Predicted and ground-truth microstructure frames at t = 11, 50, 100, 150, 200; (B) RMSE and SSIM evaluated every 5 frames over the prediction horizon; (C–D) Grain size distributions (GSD) for prediction and ground truth at t=50 and t=200, respectively.
Fig. 5 : Accuracy metrics on grain growth prediction (10 input frames, 190 output frames): (A) Dataset-averaged RMSE and SSIM evaluated every 5 frames over the prediction horizon; (B) Temporal evolution of the mean grain area for predicted and ground-truth sequences (computed every 10 frames); (C) Corresponding linear regression fits highlighting agreement in coarsening trends between prediction and ground truth.
Fig. 6 : Grain growth prediction (5 input frames, 95 output frames) using the proposed fully convolu- tional spatiotemporal model: Two representative test samples are shown (indexed numerically). (A) Predicted and ground-truth microstructure frames at t = 25, 40, 60, 80, 100; (B) RMSE and SSIM computed between the predicted and ground-truth sequences; (C–D) Grain size distributions evaluated (GSDs) for prediction and ground truth at t=25 and t=100, respectively.
Fig. 7 : Accuracy metrics on grain growth prediction (5 input frames, 95 output frames): (A) Dataset-averaged RMSE and SSIM evaluated every 5 frames over the prediction horizon; (B) Temporal evolution of the mean grain area for predicted and ground-truth sequences (computed every 10 frames); (C) Corresponding linear regression fits highlighting agreement in coarsening trends between prediction and ground truth.
Fig. 8 : Spinodal decomposition prediction (10 input frames, 90 output frames) using the proposed fully convolutional spatiotemporal model: Two test samples are shown (indexed numerically). (A) Prediction sequence and the corresponding ground-truth sequence, shown at t = 11, 25, 50, 75, 100; (B) RMSE and SSIM computed between the predicted and ground-truth sequences at 5-frame intervals.
Fig. 9 : Accuracy metrics on spinodal decomposition prediction (10 input frames, 90 output frames): (A) Dataset-averaged RMSE and SSIM evaluated every 5 frames over the prediction horizon. (B–C) The growth/shrinking of a select particle, boxed by a red square. The ground truth (red) and predicted (blue) are displayed together, measured at every frame.
Fig. 10 : Spinodal decomposition prediction (10 input frames, 200 output frames) using the proposed fully convolutional spatiotemporal model: Two test samples are shown (indexed numerically). (A) Prediction sequence and the corresponding ground-truth sequence, shown at t = 11, 60, 110, 160, 210; (B) RMSE and SSIM computed between the predicted and ground-truth sequences at 5-frame intervals.
Fig. 11 : Accuracy metrics on spinodal decomposition prediction (10 input frames, 200 output frames): (A) Dataset-averaged RMSE and SSIM evaluated every 5 frames over the prediction horizon. (B–C) The growth/shrinking of a select particle, boxed by a red square. The ground truth (red) and predicted (blue) are displayed together, measured at every frame.
Models
Inference Time (hh/mm/ss)
FLOPs
The proposed model
00:00:30
9.327G
ConvLSTM
00:08:48
1.168T
PredRNN++
00:22:14
3.552T
Table 1: Inference efficiency comparison across deep spatiotemporal models on grain growth prediction . Wall-clock inference time and estimated FLOPs for grain growth and spinodal decomposition prediction. The proposed fully convolutional model achieves markedly lower inference time than recurrent baselines (ConvLSTM and PredRNN++), consistent with its substantially reduced FLOPs.