Geometric deep learning for computational mechanics Part II: Graph embedding for interpretable multiscale plasticity

Nikolaos N. Vlassis, WaiChing Sun

TL;DR

This paper introduces a geometric deep learning framework that replaces traditional, "hand-picked" internal variables in plasticity models with interpretable graph embeddings. By treating finite element meshes as graphs and using a Graph Autoencoder (GAE), the authors can compress complex microstructural deformation patterns into a low-dimensional space, predict their evolution, and decode them back into visualizable patterns.

What Problem Does This Paper Address?

In computational mechanics, predicting how materials permanently deform (plasticity) is notoriously difficult. Traditional models rely on internal variables—mathematical placeholders like "plastic strain" or "hardening parameters"—to track a material's history.

However, these variables face three major criticisms:

  1. Lack of Interpretability: It is often unclear what a specific scalar value actually represents in the physical microstructure.
  2. Measurement Gap: These variables are often impossible to measure directly in a lab, making validation a headache.
  3. Phenomenological Bias: The laws governing these variables are often "hand-picked" by humans based on intuition, which may miss complex spatial patterns or topological changes in the material.

The authors aim to bridge this gap by using Geometric Deep Learning to automatically discover these internal variables from raw data while keeping them physically interpretable.

Methods

The authors propose a multi-step pipeline that integrates graph theory with classical mechanics:

  • Graph Representation: They represent a Finite Element (FE) mesh as an undirected graph. The nodes are integration points, and the edges represent the connectivity of the elements.
  • Graph Autoencoder (GAE): They use a Graph Isomorphism Network (GIN) architecture.
  • The Encoder ($\mathcal{L}_{enc}$) compresses the high-dimensional plastic strain field of a mesh into a low-dimensional "feature vector" ($\zeta$).
  • The Decoder ($\mathcal{L}_{dec}$) performs the reverse, mapping that vector back onto the graph.
  • Hybrid Constitutive Model: Instead of a pure "black-box" neural network, they embed these learned variables into a traditional Return Mapping Algorithm. They use:
  • A Recurrent Neural Network (RNN/GRU) to act as a "kinetic law," predicting how the feature vector $\zeta$ evolves over time.
  • Feed-forward Networks to define the yield surface (when the material starts to deform) and the plastic flow direction.

Datasets

The study relies on synthetic data generated through high-fidelity Finite Element Method (FEM) simulations:

  • Microstructure A: A square domain with a single circular hole.
  • Microstructure B: A more complex square domain with multiple irregular voids.
  • Loading Paths: 10,000 training samples were generated for each microstructure by simulating 100 different combinations of axial and shear loading over 100 time steps.
  • Mesh Variations: The authors tested the model on different mesh densities ($N=100, 250, 576$ nodes) to check for mesh sensitivity.

Key Insights and Ideas

  • Interpretability via Decoding: The "aha moment" is the use of the decoder. Because the low-dimensional embedding can be mapped back to the original mesh, researchers can visualize exactly what the "hidden" variables represent—such as the specific spatial patterns of strain localization around a void.
  • Topology Matters: By using GIN layers, the model respects the topology of the material. It doesn't just look at values; it looks at how those values are connected across the mesh, which is critical for capturing how "force chains" or "slip systems" evolve.
  • Physics-ML Hybridization: The model isn't trying to learn physics from scratch. By using a return-mapping scheme, it enforces the distinction between elastic (reversible) and plastic (irreversible) behavior, which makes it far more robust than a standard RNN when faced with complex unloading and reloading cycles.

Key Results

  • High Compression, Low Loss: The GAE successfully compressed the state of hundreds of nodes into a vector of just 16 components ($D=16$) while maintaining high reconstruction accuracy.
  • Superior Generalization: When tested on "unseen" loading paths (like random unloading and reloading), the graph-enhanced model significantly outperformed standard black-box RNNs. The standard RNNs often "hallucinated" plastic deformation during elastic unloading, whereas this model correctly identified the elastic regime.
  • Mesh Robustness: The framework showed that increasing mesh density actually improves the autoencoder's performance, as the higher resolution provides a "denser" information field for the network to learn from.

Limitations and Future Work

  • Isotropy Restriction: The current implementation is tailored for isotropic materials. The authors note that extending this to anisotropic materials (where properties change with direction) is mathematically straightforward but will require training in much higher-dimensional parametric spaces.
  • Data Hunger: Like most deep learning approaches, the model requires a massive amount of high-fidelity simulation data to train the initial autoencoder.
  • Noise Sensitivity: While the model works well on clean simulation data, the authors suggest that future work should investigate how it handles "noisy" data from real-world sensors or experimental digital image correlation (DIC).

Summary updated

Abstract

The history-dependent behaviors of classical plasticity models are often driven by internal variables evolved according to phenomenological laws. The difficulty to interpret how these internal variables represent a history of deformation, the lack of direct measurement of these internal variables for calibration and validation, and the weak physical underpinning of those phenomenological laws have long been criticized as barriers to creating realistic models. In this work, geometric machine learning on graph data (e.g. finite element solutions) is used as a means to establish a connection between nonlinear dimensional reduction techniques and plasticity models. Geometric learning-based encoding on graphs allows the embedding of rich time-history data onto a low-dimensional Euclidean space such that the evolution of plastic deformation can be predicted in the embedded feature space. A corresponding decoder can then convert these low-dimensional internal variables back into a weighted graph such that the dominating topological features of plastic deformation can be observed and analyzed.
BibTeX
@inproceedings{vlassis2022geometric,
  author = {Nikolaos N. Vlassis and WaiChing Sun},
  title = {{Geometric deep learning for computational mechanics Part II: Graph embedding for interpretable multiscale plasticity}},
  booktitle = {Computer},
  year = {2022},
  doi = {10.1016/j.cma.2022.115768},
}
Figure