Source-linked AI summary
Geometric deep learning for computational mechanics Part I: Anisotropic Hyperelasticity
Nikolaos Vlassis, Ran Ma, WaiChing Sun
TL;DR
Complex microstructures contain topological information that homogenized descriptors do not capture well for anisotropic hyperelastic modeling. The paper combines weighted connectivity graphs, a hybrid graph-convolutional/regression network, and first-order Sobolev training to predict smooth stored-energy functionals; the graph input adds value in forward predictions and requires less data. The approach is demonstrated on anisotropic polycrystal fracture responses, within stated graph-structure and convexity limitations.
Problem
Homogenized Euclidean descriptors do not capture the topological microstructural information needed to model anisotropic hyperelastic responses across complex polycrystals.
Method
The paper jointly learns graph representations and stored elastic energy using a hybrid graph convolutional network and multilayer perceptron, with first-order Sobolev training for energy and stress.
Results
The graph-based approach requires less data, and its added value for forward predictions is demonstrated in k-fold validation; generated energy functionals reproduce anisotropic polycrystal fracture responses.
Takeaways & Limitations
Weighted graphs provide a compact route to incorporate relational microstructural information into learned anisotropic constitutive models and study fracture responses of unseen polycrystals.
Takeaways & Limitations
The GCN is not inductive, so generalization is restricted to unseen graph structures statistically similar to those used in training; convexity constraints remain future work.
Abstract
from arXiv · showhide
This paper is the first attempt to use geometric deep learning and Sobolev training to incorporate non-Euclidean microstructural data such that anisotropic hyperelastic material machine learning models can be trained in the finite deformation range. While traditional hyperelasticity models often incorporate homogenized measures of microstructural attributes, such as porosity averaged orientation of constitutes, these measures cannot reflect the topological structures of the attributes. We fill this knowledge gap by introducing the concept of weighted graph as a new mean to store topological information, such as the connectivity of anisotropic grains in assembles. Then, by leveraging a graph convolutional deep neural network architecture in the spectral domain, we introduce a mechanism to incorporate these non-Euclidean weighted graph data directly as input for training and for predicting the elastic responses of materials with complex microstructures. To ensure smoothness and prevent non-convexity of the trained stored energy functional, we introduce a Sobolev training technique for neural networks such that stress measure is obtained implicitly from taking directional derivatives of the trained energy functional. By optimizing the neural network to approximate both the energy functional output and the stress measure, we introduce a training procedure the improves efficiency and generalize the learned energy functional for different microstructures. The trained hybrid neural network model is then used to generate new stored energy functional for unseen microstructures in a parametric study to predict the influence of elastic anisotropy on the nucleation and propagation of fracture in the brittle regime.
1 Introduction
Traditional constitutive models rely on human-selected geometric descriptors, which may be insufficient for complex microstructures. This work motivates graph-based learning to encode topological information and jointly learn physically useful representations with energy predictions.
- Motivation: Human-selected descriptors such as porosity, dislocation density, and grain orientation may inadequately capture complex microstructural geometry and topology.The paper argues that limited interpretation of microstructural state variables can reduce prediction quality.
- Contribution: The proposed approach uses a weighted crystal connectivity graph as the sole additional anisotropy input for approximating stored elastic energy across polycrystals.The graph stores non-Euclidean information such as crystal or grain connectivity.
- Limitations of Euclidean inputs: Directly using scanned images requires training pools with sufficient quantity and resolution, while filtering high-resolution 3D voxel data can be computationally expensive.The paper characterizes image-based feature discovery as dependent on empirical trial and error.
- Motivation: Graph representations can preserve relational microstructural information while remaining more interpretable, flexible, economical, and efficient than high-resolution voxel feature spaces.Domain expertise can be used to design node connectivity, node weights, and edge weights tailored to mechanics predictions.
- Related work: Graph learning methods include unsupervised representation learning and graph neural networks that operate directly on graph structures.The paper distinguishes converting graph data into machine-readable features from processing graph inputs directly.
- Methodological motivation: Separately learned graph representations may not be physically meaningful or useful for energy prediction, motivating simultaneous representation learning and regression.The hybrid architecture addresses this concern through a common back-propagation process and loss function.
2 Graphs as non-Euclidean descriptors for micro-structures
The paper represents polycrystal microstructures as weighted graphs to preserve connectivity and node features as non-Euclidean descriptors. It defines graph matrices, Laplacians, and feature representations used to encode these structures.
- Graph theory terminologies: A graph consists of vertices and edges, which may connect unordered or ordered vertex pairs in undirected or directed graphs.The paper also distinguishes loops, multiple edges, simple graphs, and adjacent vertices.
- Graph descriptors: Weighted graphs store microstructural information such as crystal connectivity as non-Euclidean input for anisotropic stored-energy approximation.The paper focuses on graphs whose node weights quantify microstructural features.
- Graph matrices: The adjacency matrix records connectivity, the degree matrix records neighbor counts, and the graph Laplacian is defined as L = D − A for binary graphs.The normalized symmetric Laplacian Lsym is also used to represent the graph structure.
- Polycrystal graph construction: For polycrystals, each crystal becomes a node and an edge connects two nodes when the crystals share a surface.The resulting connectivity graph is undirected, with binary adjacency entries indicating whether crystal pairs are connected.
- Node weights: Node features such as crystal volume or orientation are organized in a feature matrix X, with rows representing nodes and columns representing features.The graph can therefore combine connectivity with multiple attributes for each crystal.
- Scope of graph descriptors: Connectivity graphs are suitable for polycrystals but may not apply to evolving granular contacts or porous-material flow networks.Other applications may require graph descriptors that evolve with time or represent pore space and flow networks.
3 Deep learning on graphs
The paper combines graph convolution with multilayer perceptron regression so microstructural graphs and deformation jointly inform stored-energy predictions. The architecture learns neighborhood feature correlations while recognizing that graph generalization is limited to statistically similar structures.
- Hybrid architecture: The graph branch accepts the normalized Laplacian Lsym and node feature matrix X, while the regression branch also receives the right Cauchy–Green tensor C.The predicted output is the stored energy functional.
- Graph convolution: Graph convolution aggregates neighboring crystal features, allowing the model to learn correlations between connected nodes rather than treating crystals independently.This spatial filtering is presented as useful for constitutive modeling of polycrystals.
- Graph inputs: In this work, four crystal features—volume and three Euler angles—form the node weights, giving X dimensions N × 4.Unweighted graphs can instead use the identity matrix X = I.
- Hybrid architecture: The hybrid architecture combines a graph convolutional encoder for polycrystal graphs with an MLP regression branch for energy-functional prediction.The encoder produces a lower-dimensional feature vector that is combined with deformation information.
- Joint learning: The hybrid network performs unsupervised graph classification and energy-functional regression simultaneously under a common objective function.Its two branches are updated together through back-propagation.
- Limitations and regularization: The GCN is not inductive, so the model is intended for unseen graphs with statistically similar crystal-number distributions.Dropout and L2 regularization are used to address overfitting on the graph encoder branch.
4 Sobolev training for hyperelastic energy functional predictions
The paper trains hyperelastic energy-functional approximators using Sobolev objectives that constrain both energy and stress responses. The surrogate takes C as input, predicts energy, and obtains stress from its gradient while emphasizing smoothness and derivative accuracy.
- Sobolev objective: Sobolev training incorporates both stored-energy values and their derivatives into the loss rather than minimizing energy prediction errors alone.The derivative with respect to C corresponds to one half of the second Piola–Kirchhoff stress tensor.
- Smoothness: Sobolev training is intended to reduce spurious oscillations and produce smooth energy and stress predictions by regularizing the learned function's derivatives.Traditional L2 training does not regularize gradients, while the schematic identifies the gradient with respect to C as the predicted stress tensor.
- Model formulation: The surrogate approximates the stored energy as ˆψ(C|W, b), using the right Cauchy–Green deformation tensor C as input.The model parameters are the network weights W and biases b.
- Sobolev objective: The H1-style objective constrains both the predicted energy and its first derivative with respect to C.The stress information is used as the higher-order training constraint, enforcing component-wise stress accuracy.
- Scope: The first-order Sobolev scheme is used because ReLU-based networks make direct control of second derivatives difficult.An H2 objective could constrain energy, stress, and stiffness, but the paper mainly focuses on first-order constraints.
5 Verification exercises for checking compatibility with physical constraints
The verification exercises assess whether the trained constitutive models satisfy objectivity, anisotropy, and convexity-related conditions. The models use C to automatically satisfy frame indifference, while convexity is checked separately and constraint penalties are proposed for future enforcement.
- Objectivity: Modeling the response as a function of C automatically satisfies the required invariance under rigid-body rotations.The paper therefore trains all models as functions of the right Cauchy–Green deformation tensor.
- Anisotropy: The predicted polycrystal behavior is anisotropic rather than right-rotationally invariant.The paper reports this behavior as a verification result for the hybrid architecture.
- Convexity: Convexity is required for thermodynamical consistency but is difficult to verify directly for the high-dimensional black-box energy functional.The paper notes that black-box convexity estimation exists but is outside the work's scope.
- Convexity: The convexity check tests whether the first-order Taylor expansion at Cβ remains a global under-estimator for every pair of deformation tensors in the domain.This condition is applied to the approximated energy functional for a specific connectivity graph G.
- Constraint enforcement: Additional weighted penalty terms can be added to the loss function when objectivity, anisotropy, or convexity conditions require enforcement during optimization.The paper presents this as a mechanism for future constraint enforcement rather than a necessary component of the current training.
6 FFT offline database generation
The offline database is generated through FFT-based mesoscale homogenization simulations using anisotropic Fung elasticity. Random grain structures, orientations, and strains provide graph structures and material responses for training the macroscale graph-based model.
- Database generation: The homogenized response database supplies graph structures and material responses for geometric learning.The workflow uses mesoscale simulations to generate training data for the macroscale model.
- FFT homogenization: An FFT-based spectral method solves the mesoscale periodic homogenization problem through the Lippmann–Schwinger equation.The convolution with Green's operator is performed efficiently in the Fourier domain.
- Anisotropic Fung elasticity: The generalized Fung model defines the mesoscale strain-energy density through a scalar material constant, Green strain tensor, and fourth-order stiffness tensor.Anisotropy is represented through the stiffness tensor, which depends on spatial orientation and material symmetry.
- Simulation variables: The elastic constants remain fixed across mesoscale simulations, while grain structure and initial RVE orientation vary.This isolates microstructural and orientation variation in the generated database.
- RVE generation: Random equiaxed RVEs contain 40 to 50 grains and are discretized on a 49 × 49 × 49 grid.Grain structures are generated with NEPER, and initial orientations with MTEX using randomly generated orientation distributions.
- Loading protocol: Each RVE is simulated for 200 randomly generated average strains within a strain-component range of 0.0 to 0.1.Because the constitutive relation is hyperelastic, the simulation results are path independent.
7 Numerical Examples
The numerical examples evaluate whether graph-dependent hybrid models generalize anisotropic constitutive predictions across polycrystals, capture anisotropy, and support fracture simulations. Results indicate improved stress and energy prediction with Sobolev training, distinct responses for different graphs and rotations, and remaining concerns about overfitting and convexity verification.
- 7 Numerical Examples: The database contains anisotropic hyperelastic behavior from simulations of 150 polycrystal RVEs.The RVEs are used to test predictive capabilities and generalization across different microstructures.
- 7.1 Training constitutive models for polycrystals with non-Euclidean data: The hybrid model generalizes predictions across RVEs, whereas a single-RVE surrogate cannot retain accuracy when applied to other RVEs.The additional RVE input enables one learned constitutive law to represent a family of microstructures.
- 7.1 Training constitutive models for polycrystals with non-Euclidean data: A weighted connectivity graph is added as input by representing each polycrystal grain as a node and each grain contact as an edge.This graph supplies microstructural information beyond the deformation tensor.
- 7.1 Training constitutive models for polycrystals with non-Euclidean data: Sobolev training greatly reduces blind prediction errors for energy and stress compared with the alternative model combinations.The results are presented using scaled MSE versus eCDF curves for energy and stress predictions.
- 7.1 Training constitutive models for polycrystals with non-Euclidean data: The hybrid model distinguishes different RVE behaviors, producing different energy surfaces and stress responses from graph-dependent encoded features.An encoded feature vector of dimension 9 was used in the illustrated experiment, while fewer than three features performed less well.
- 7.3 Parametric study: Anisotropic responses of polycrystals in phase-field fracture: Rotating the RVE produces different energy and stress responses, confirming anisotropic behavior in the learned material model.The responses are evaluated for 0°, 30°, and 60° rotations, with fracture branching becoming more prevalent at doubled impact velocity.
- 7.2 Model assessment: The hybrid model outperforms the simple MLP but is prone to overfitting, with substantially lower training error than blind prediction error.Regularization reduces the gap between training and testing eCDF curves.
- 7.2.2 Convexity: Convexity checks across all RVEs satisfy the tested inequality, but these checks are necessary rather than sufficient for establishing convexity.Approximately 265,000 pairwise checks are performed per RVE.
8 Conclusion
The paper develops a hybrid neural network using weighted graphs and Sobolev training to model smooth elastic stored-energy functionals from rich microstructural information. Graph inputs improve forward predictions and produce polycrystal-dependent crack-branching behavior without additional anisotropy measures.
- 8 Conclusion: The hybrid neural network predicts elastic stored-energy functionals for Green-elastic materials using weighted graphs to encode microstructural information beyond porosity and density.The graph representation provides non-Euclidean descriptors for geometric learning.
- 8 Conclusion: Sobolev training reduces spurious oscillations in the derivative of the learned hyperelastic energy functional across the parametric space.The technique addresses oscillations associated with insufficient regularization or overfitting.
- 8 Conclusion: The graph-based approach requires less data than the image-based hybrid architecture, while its added value in forward predictions is shown by k-fold validation.