Open Access

ARTICLE

A Graph-Based Spatio-Temporal Attention Network for Stress–Strain Behavior Prediction of Copper-Based Composites

Chuhan Zhang¹, Jinguo You^1,*, Jialin Xu¹, Mingqian Li¹, Xiaofeng Chen², Jingmei Tao², Caiju Li², Jianhong Yi²

1 Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, China
2 Faculty of Materials Science and Engineering, Kunming University of Science and Technology, Kunming, China

* Corresponding Author: Jinguo You. Email:

(This article belongs to the Special Issue: Advances in Computational Materials Science: Focusing on Atomic-Scale Simulations and AI-Driven Innovations)

Computers, Materials & Continua 2026, 88(1), 15 https://doi.org/10.32604/cmc.2026.079587

Received 24 January 2026; Accepted 19 March 2026; Issue published 08 May 2026

Abstract

With the rapid development of artificial intelligence and data-driven modeling, deep learning has become an effective tool for analyzing scientific discovery such as predicting material behaviors. Graphene-reinforced copper-based composites, which exhibit excellent mechanical, electrical, and thermal properties, have attracted extensive attention in advanced engineering applications; however, accurate prediction of their stress–strain behavior still relies heavily on computationally expensive molecular dynamics simulations or experiments. In this work, we propose a Graph-based Spatio-Temporal Attention Network, termed GraphSTAN, for stress–strain behavior prediction of copper-based composites. Specifically, atomic-scale initial microstructures are encoded as graphs and integrated with static physical parameters. A topology-aware spatio-temporal feature interaction mechanism is introduced to effectively couple structural representations with stress–strain time-series dynamics, enabling accurate prediction of full stress–strain evolution. Moreover, a multi-features dataset is constructed based on LAMMPS molecular dynamics simulations, consisting of 596 independent simulation samples corresponding to distinct combinations of microstructural configurations, loading conditions, and stress–strain time series. Experimental results demonstrate that GraphSTAN effectively predicts full stress–strain curves and achieves the higher performance of R2, MAE and RMSE for yield strength and Young’s modulus, respectively, significantly outperforming baseline methods.

---

Keywords

---