# Accelerating discrete dislocation dynamics simulations with graph neural networks

**Authors:** Nicolas Bertin, Fei Zhou  
**arXiv:** [2208.03296](https://arxiv.org/abs/2208.03296)  
**Date:** 2022-08-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.03296v2/

## Key points

- Plasticity at the mesoscale involves studying the behavior of crystalline materials through dislocation lines movement
- Discrete Dislocation Dynamics (DDD) is a common computational method for studying plasticity, but its high computational cost limits its applicability
- DDD-GNN framework replaces time-integration process with a Graph Neural Network (GNN) model trained on DDD trajectories
- Success of DDD-GNN model attributed to partitioning stresses into short-range and long-range contributions, approximated using message-passing GNN
- DDD-GNN model represents dislocation configurations as graph structures, enabling efficient operation by GNN algorithm
- DDD-GNN model demonstrates stability and accuracy in predicting force fields and material properties without explicitly computing nodal forces or dislocation mobilities

## Summary

, , , , 

At the mesoscale, the study of plasticity involves examining the behavior of crystalline materials through the movement of dislocation lines. A commonly used computational method for this is Discrete Dislocation Dynamics (DDD), but its high computational cost has limited its applicability. To address this issue, a new framework called DDD-GNN has been introduced, which replaces the time-integration process with a Graph Neural Network (GNN) model trained on DDD trajectories. The success of this model is due in part to the partitioning of stresses between short-range and long-range contributions, with short-range interactions being characterized by local dislocation configurations and long-range interactions depending on the entire dislocation network. By utilizing a message-passing GNN, local dependence on neighboring dislocation configurations can be approximated. This approach has shown promise in predicting force fields and other material properties, making it suitable for simulating complex physics scenarios. A key component of the DDD-GNN framework is representing dislocation configurations as graph structures with nodes and edges, allowing for efficient operation by the GNN algorithm using node features such as type flags and uniform input stress values incorporated into both edge and vertex features. Overall, the DDD-GNN model has demonstrated stability and accuracy in reproducing unseen ground-truth DDD simulation responses across various straining rates and obstacle densities without explicitly computing nodal forces or dislocation mobilities during time-integration. This innovative approach not only accelerates DDD simulations but also opens up possibilities for incorporating more complex behaviors related to dislocation motion in materials research.