A Hybrid GNN-FEM Framework for Phase-Field Fracture Simulation. Physics-Preserving Hybridization for Generalizable Surrogate Modeling

Technical Summary: A Hybrid GNN–FEM Framework for Phase-Field Fracture Simulation

Problem Statement
Scientific machine learning (SciML) offers potential for accelerating simulations of complex physical systems, yet achieving physically consistent and generalizable predictions for nonlinear, history-dependent problems remains a significant challenge. Specifically, phase-field fracture modeling, while robust in simulating complex crack evolution (initiation, propagation, and branching) without explicit tracking, suffers from high computational costs. This cost arises from the need to solve a coupled, nonlinear, and history-dependent system incrementally within a finite element method (FEM) framework.

Existing data-driven surrogate models often struggle with generalization across varying geometries, loading conditions, material properties, and discretizations. Purely data-driven approaches frequently fail to maintain physical consistency when extrapolating beyond training distributions, while physics-informed machine learning (PIML) often lacks robustness when problem settings (e.g., mesh resolution or boundary conditions) change. The core difficulty lies in approximating the full solution trajectory of a path-dependent system without violating the underlying physical structure or requiring prohibitive amounts of training data.

Methodology
The authors propose a Hybrid GNN–FEM framework that integrates a Graph Neural Network (GNN) surrogate into the conventional staggered solution scheme for phase-field fracture. Rather than replacing the entire simulation procedure, the framework selectively substitutes the computationally intensive phase-field update step while retaining the FEM-based displacement solver to enforce mechanical equilibrium and boundary conditions.

Key methodological components include:

1. Hybrid Staggered Scheme:

   • The displacement field ($u$) and history field ($H$) are solved using standard FEM at each load increment.
   • The phase-field variable ($d$) is predicted by the GNN surrogate based on the current history field and domain geometry.
   • Irreversibility ($d_n \geq d_{n-1}$) is enforced as a post-processing step.
   • This preserves the incremental solution structure, ensuring the surrogate does not need to learn the full solution trajectory but only the local evolution of the phase field.

2. Graph Representation and Dimensionless Features:

   • The finite element mesh is represented as a graph $G=(V, E)$, where nodes correspond to mesh nodes and edges to connectivity.
   • Dimensionless Node Features: The history field $H$ is normalized by $G_c \ell_c$ (critical energy release rate $\times$ regularization length) and projected to nodal values. A binary boundary indicator is also included.
   • Dimensionless Edge Features: Edge lengths are normalized by the regularization length scale $\ell_c$.
   • This normalization ensures invariance to material property scales and mesh resolution, facilitating generalization.

3. GNN Architecture:

   • The model employs an Encoder–Processor–Decoder architecture.
   • Encoder: Maps input features to a latent space using Multi-Layer Perceptrons (MLPs).
   • Processor: Performs $K$ iterations of message passing with residual connections, updating node and edge representations based on local interactions.
   • Decoder: Maps the final latent node representations to the predicted phase-field value.
   • The architecture is designed to be invariant under coordinate rotations, consistent with isotropic material behavior.

4. Physics-Informed Training Strategy:

   • The model is trained using a combined loss function: $L = L_{data} + \lambda_{phys} L_{phys}$.
   • Data-Driven Loss ($L_{data}$): Mean squared error between predicted and FEM-computed phase-field values, derived from a single high-fidelity simulation.
   • Physics-Informed Loss ($L_{phys}$): Mean squared residual of the discretized weak form of the phase-field governing equation. This is computed using randomly generated history fields that do not necessarily correspond to physical crack states, thereby broadening the training distribution and enforcing physical laws.

Key Contributions

1. GNN-Integrated Hybrid Framework: The study introduces a surrogate that replaces only the phase-field update within a staggered FEM procedure. This selective replacement maintains the incremental solution structure and physical consistency without requiring the surrogate to approximate the full displacement-history-phase-field trajectory.
2. Multi-Level Physics Integration: Physical knowledge is incorporated simultaneously at three levels:
   • Feature Level: Dimensionless representations of history fields and edge lengths.
   • Loss Level: A physics-based residual loss derived from the governing phase-field equation.
   • Structure Level: A graph-based architecture operating directly on unstructured meshes.
3. Systematic Generalization Study: The framework is rigorously evaluated across unseen variations in mesh resolution, geometric configurations (including holes), loading conditions (shear vs. tension), and material properties (heterogeneous inclusions), demonstrating robustness without retraining.

Results
Numerical experiments validate the framework's performance against conventional Full FEM solutions:

• Discretization Generalization: The hybrid model (trained on $h=0.2$) accurately predicts crack evolution and force-displacement responses on finer meshes ($h=0.1, 0.05$). In contrast, a purely data-driven (DD) model trained on the same data fails to generalize to finer resolutions, deviating significantly in the post-peak regime.
• Geometric and Loading Generalization: The model successfully predicts crack paths in geometries with circular holes and under shear loading conditions, capturing crack deflection and initiation at stress concentrators. The purely DD model showed discrepancies in energy dissipation timing and crack progression.
• Material Generalization: The framework handled heterogeneous materials with properties differing by orders of magnitude from the training set, accurately predicting crack paths around inclusions.
• Computational Efficiency: The hybrid approach reduced total computational time by approximately 20–23% for refined meshes, with a more significant reduction (up to 88.6%) in the phase-field solution component. The displacement solver, which dominates the cost, remains unchanged.

Significance and Claims
The paper claims that the proposed framework provides a viable strategy for generalizable surrogate modeling in nonlinear, history-dependent fracture problems. By embedding data-driven components within the governing solution procedure rather than replacing the entire physics solver, the approach mitigates error accumulation and ensures physical consistency.

The authors emphasize that the success of this method relies on identifying physically meaningful, scale-consistent, and incrementally structured learning targets rather than relying on brute-force data generation. The integration of dimensionless features and physics-informed losses allows the model to generalize across varying problem settings (geometry, discretization, loading, and material) without retraining.

The work concludes that while the current formulation is limited to isotropic, quasi-static fracture, it establishes a data-efficient pathway for reliable scientific machine learning. The authors note that future work could extend this to anisotropic materials and 3D problems, or potentially apply surrogate modeling to the displacement field, though such extensions would involve trade-offs between generalizability and computational speed.