Physics-Informed GCN-LSTM Framework for Long-Term Forecasting of 2D and 3D Microstructure Evolution

2025-09-18

AI Summary

This work addresses the challenges of low accuracy, high computational cost, and poor generalizability in long-term evolution prediction of multi-component 2D/3D microstructures. We propose a physics-informed GCN-LSTM framework featuring a composition-aware latent graph modeling mechanism: a convolutional autoencoder compresses phase-field simulation data into a latent space where spatiotemporal evolution patterns—across dimensions and compositions—are jointly learned. Physics-based constraints—including mass conservation and thermodynamic principles—are embedded as regularization terms in the loss function to enhance model interpretability and extrapolation capability. Experiments demonstrate that our method significantly outperforms state-of-the-art approaches across multiple metrics: prediction error is reduced by over 30%, computational overhead is decreased by an order of magnitude, and stable long-term predictions spanning more than 100 time steps are achieved.

Technology Category

Machine Learning: Graph-based Machine LearningComputer Vision: Low Level & Physics-based VisionKnowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphsWeb Mining and Content Analysis: Models for Web evolutionSemantics and Knowledge: Methods, algorithms and applications for the development of semantic models, knowledge graphs and other forms of structured data models with machine-interpretable semantics

Abstract

This paper presents a physics-informed framework that integrates graph convolutional networks (GCN) with long short-term memory (LSTM) architecture to forecast microstructure evolution over long time horizons in both 2D and 3D with remarkable performance across varied metrics. The proposed framework is composition-aware, trained jointly on datasets with different compositions, and operates in latent graph space, which enables the model to capture compositions and morphological dynamics while remaining computationally efficient. Compressing and encoding phase-field simulation data with convolutional autoencoders and operating in Latent graph space facilitates efficient modeling of microstructural evolution across composition, dimensions, and long-term horizons. The framework captures the spatial and temporal patterns of evolving microstructures while enabling long-range forecasting at reduced computational cost after training.

Problem

• Forecasting 2D and 3D microstructure evolution long-term
• Capturing composition and morphological dynamics efficiently
• Enabling long-range forecasting at reduced computational cost

Innovation

• Physics-informed GCN-LSTM integration
• Latent graph space operation
• Convolutional autoencoders data compression