Graph representations of dislocation networks

Gaurav Lathiya, Markus StrickerarXiv 2026 · arXiv

TL;DR

This paper demonstrates that the complex "bird’s nest" of dislocations in a crystal can be effectively modeled using graph theory. By extracting a compact set of topological features from these graphs, the researchers can predict a material's yield stress with high accuracy ($R^2$ up to 0.94), proving that the connectivity of the dislocation network is a fundamental driver of mechanical strength.

What Problem Does This Paper Address?

Predicting how a metal will deform under load is a cornerstone of materials science, yet it remains notoriously difficult. The culprit is the dislocation network—a tangled web of line defects within the crystal lattice. While simple metrics like "dislocation density" (the total length of these lines per unit volume) provide a rough estimate of strength, they ignore the topology—how these lines are connected, branched, and knotted.

The authors address the fundamental gap in linking the microscopic, evolving state of these networks to macroscopic mechanical properties like yield stress. They ask: Can we use graph theory to distill the complexity of a dislocation network into a few key numbers that tell us how strong a material is?

Methods

The researchers combined high-fidelity simulations with machine learning and graph theory:

  1. Simulation: They generated a dataset of 627 dislocation trajectories using Discrete Dislocation Dynamics (DDD), varying box sizes, crystal orientations, initial densities, and strain rates.
  2. Graph Mapping: They converted the 3D line networks into five different graph representations ($R_0$ through $R_4$), ranging from literal segment-by-segment translations to abstract models where entire dislocation loops are nodes and junctions are edges.
  3. $R_0$ (Discrete Dislocation Graph): Direct mapping of simulation vertices to nodes.
  4. $R_1$ (Loop Connectivity): Loops are nodes; junctions are edges.
  5. Topological Characterization: For every graph, they calculated seven global properties: average degree, transitivity (clustering), assortativity (degree correlation), graph density, diameter, and the total count of nodes and edges.
  6. Machine Learning: They used a Random Forest regression model to predict the logarithm of yield stress, using the graph properties and the applied strain rate as inputs.
The $R_0$ representation directly maps the discretized dislocation segments from the simulation into a graph.

Datasets

The study relies on a newly generated dataset of 627 DDD simulation trajectories.

  • Material: FCC crystal.
  • Variations: Four box sizes (0.5 to 2.0 $\mu$m), four crystallographic orientations ([100], [110], [111], [234]), and seven strain rates ($5.0$ to $5.0 \times 10^6$ $s^{-1}$).
  • Validation: The simulated data was validated against the "universal scaling curve" established by Fan et al. (2021), showing excellent agreement across several orders of magnitude in strain rate and density.

Key Insights and Ideas

  • Topology is Robust: The model's predictive accuracy was remarkably consistent across all five graph representations. This suggests that the relationship between network topology and stress is a physical reality, not an artifact of how the graph is drawn.
  • Low Intrinsic Dimensionality: Although they calculated seven different graph properties, Principal Component Analysis (PCA) revealed that the "state" of the network actually lives in a much lower-dimensional space (roughly 3 dimensions).
  • Property Grouping: Graph features naturally fell into two main camps: Size/Extent (number of nodes, edges, and diameter) and Connectivity/Clustering (degree, transitivity, and assortativity).
  • Statistical Stability at Scale: The model performed significantly better on larger simulation volumes. In larger "bird's nests," finite-size fluctuations smooth out, making the global topological measures more representative of the material's true state.
Different graph abstractions, like $R_1$ (loops as nodes) and $R_2$ (junctions as nodes), provide complementary views of the same physical network.

Key Results

  • High Predictive Power: The Random Forest model achieved an $R^2$ of 0.86–0.90 for the full dataset.
  • Improved Accuracy for Large Systems: When restricted to larger simulation boxes (1.5–2.0 $\mu$m), the accuracy increased to $R^2 = 0.91–0.94$.
  • Beyond Density: The graph-based features improved upon a baseline model that used only dislocation density and strain rate, particularly for representations $R_0$ through $R_3$.

Limitations and Future Work

  • Yield Point Focus: The current study focuses specifically on the yield point. The authors note that the next logical step is to determine if these topological descriptors can predict flow stress along the entire strain path.
  • Global vs. Local: The model currently uses global graph properties. Incorporating local descriptors (e.g., node-level centrality or local clustering) might provide even deeper insights into where a network is likely to fail or strengthen.
  • Experimental Bridge: While DDD simulations provide perfect data, applying this to experimentally reconstructed 3D dislocation structures (which are often noisy or incomplete) remains a significant future challenge.

Summary updated

Abstract

The mechanical response of a crystal is governed by the evolving network of dislocations it contains, yet quantitatively linking the `state of that network' to its yield stress remains challenging. Here we represent the dislocation networks obtained from discrete dislocation dynamic simulation as graphs, using five distinct graph representations, characterize each network through a compact set of topological properties, and predict the yield stress from these graph properties and the applied strain rate using a standard random forest model evaluated with strain rate stratified cross-validation. The model predicts the yield stress with $R^2\approx0.86-0.90$ across all box sizes and $R^2\approx0.91-0.94$ for larger box sizes, and this accuracy is consistent across all five representations, indicating that the topology-stress relationship is a robust feature of the graph properties rather than an artefact of any single graph representation. We propose graph topology as a compact and robust route for connecting the microstructural state of a dislocation network to its mechanical response in addition to established variables.
BibTeX
@misc{lathiya2026graph,
  author = {Gaurav Lathiya and Markus Stricker},
  title = {{Graph representations of dislocation networks}},
  year = {2026},
  eprint = {2611.09534},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09534},
}
Figure