A mesh-free deep energy method for phase-field modeling of thermally driven brittle fracture

Han Zhang, Mehrisadat Makki Alamdari, Babak Shahbodagh, Mohammad Vahab, Cosmin Anitescu, Timon Rabczuk, Elena AtroshchenkoarXiv 2026 · arXiv

Simulating how materials break when suddenly cooled—like a hot glass dish shattering in cold water—is a notoriously difficult problem in computational mechanics. Traditional methods rely on "meshes" (grids of triangles or squares) to solve the underlying physics. However, as cracks grow and branch, these meshes often need to be painstakingly updated or refined, a process that is both computationally expensive and prone to numerical bias.

A new paper, "A mesh-free deep energy method for phase-field modeling of thermally driven brittle fracture," proposes a radical shift. By leveraging neural networks and a "mesh-free" philosophy, the authors demonstrate a way to predict complex fracture patterns without ever defining a grid or pre-prescribing where a crack should start.

TL;DR

The authors present a Deep Energy Method (DEM) that uses multiresolution neural networks to simulate brittle fracture driven by thermal shock. Unlike previous neural-based approaches, this method successfully predicts spontaneous crack nucleation and complex branching patterns without requiring pre-existing defects or crack-adapted meshes.

What Problem Does This Paper Address?

In the world of fracture mechanics, the "phase-field" approach is a popular way to model cracks. Instead of treating a crack as a sharp line, it treats it as a narrow band where the material's "health" (the phase field) drops from 1 (intact) to 0 (broken).

The challenge arises when temperature enters the mix. Thermal shock creates thin "boundary layers" of intense stress near the surface of a material. To capture these layers and the resulting cracks, standard Finite Element Methods (FEM) require incredibly fine meshes. Furthermore, neural network approaches have historically struggled to "nucleate" cracks spontaneously; they usually need a researcher to place a small notch in the model to get the crack started. This paper aims to remove those crutches, allowing cracks to emerge naturally from the physics of the energy functional.

Methods

The researchers built a Deep Energy Method (DEM) framework that replaces traditional elements with neural networks. Key components include:

  • Dual Networks: One network ($N_M$) represents the displacement and the phase field (the crack), while a second network ($N_T$) handles the transient temperature field.
  • Multiresolution Feature Grids: To handle the sharp gradients of a crack band and the thinness of thermal boundary layers, the networks use a multiresolution encoding. This allows the model to "zoom in" on areas of interest without a global mesh.
  • Stochastic Integration: Instead of calculating energy at fixed points, the method uses a stratified Monte Carlo approach. Integration points are redrawn at every iteration. This "shuffling" prevents the network from over-fitting to specific locations and is crucial for allowing cracks to advance smoothly.
  • Irreversibility Constraints: Cracks shouldn't "heal." The authors implemented a penalty term that prevents the phase field from decreasing, but crucially restricted this penalty to the "crack set" to avoid creating a fuzzy layer of damage across the entire cooled surface.

Datasets

The method was evaluated against five classic numerical benchmarks in fracture mechanics:

  1. Single-edge-notched plate (Tension): A standard verification of the thermo-mechanical solver.
  2. Functionally graded plate (Shear): Tests the model's ability to handle varying material properties.
  3. Directionally quenched glass strip: A famous problem where a crack can be straight, wavy (oscillatory), or branched depending on the temperature.
  4. Quenched ceramic slab: A test of spontaneous nucleation where a whole array of parallel cracks forms simultaneously.
  5. Quenched annulus: Extends the problem to curved geometries using NURBS (Non-Uniform Rational B-Splines).

Results were compared against independent FEM references and experimental data from literature.

Key Insights and Ideas

  • The Power of Redrawing: A major "aha" moment in the paper is that cracks only advance effectively when integration points are redrawn every iteration. If points are fixed, the optimizer finds a local minimum that "traps" the crack tip.
  • Mesh-Free Spontaneity: By minimizing the total energy functional directly, the model allows cracks to nucleate wherever the energy balance dictates. This removes the "human bias" often found in simulations where researchers must guess where a crack might start.
  • Handling the "Damaged Film" Artifact: In thermal shock, the entire surface wants to crack. Previous models often resulted in a thin layer of diffuse damage everywhere. By refining how irreversibility is enforced, the authors ensured that distinct, sharp cracks form instead of a global "bruise."

Key Results

  • High Accuracy: On notched plates, the predicted peak loads were within 1.08% to 1.76% of the FEM reference.
  • Capturing Complexity: In the glass strip example, the model successfully reproduced the transition from straight to oscillatory to branched crack paths as the temperature jump increased. The oscillation threshold was remarkably close to theoretical linear stability predictions.
  • Denser Crack Arrays: In the ceramic slab and annulus tests, the DEM actually predicted denser crack arrays than the FEM reference. This suggests the neural approach may be more sensitive to the "bifurcation" points where a uniform stress state breaks into a periodic crack pattern.

Limitations and Future Work

While promising, the method has some hurdles:

  • Computational Efficiency: For simple problems, the neural network approach currently takes longer to converge than optimized FEM solvers.
  • 2D Constraint: The current work is limited to 2D quasi-static fracture. Real-world thermal shock is often a 3D, dynamic event.
  • Simplified Physics: The model does not yet account for the heat released by the act of fracturing itself, nor does it model the contact between crack faces if they press together.

The authors suggest that the next steps involve extending the method to 3D and exploring "inverse problems"—using observed crack patterns to work backward and identify unknown material properties.

Summary updated

Abstract

Phase-field simulation of thermally driven fracture requires resolving localized crack bands whose paths are unknown in advance and thermal boundary layers whose thickness evolves during the transient. Neural energy-based computations of thermal fracture have so far started from defects placed in advance. This paper proposes a mesh-free deep energy method based on multiresolution neural representations that predicts spontaneous crack nucleation, crack spacing, and selective crack propagation and arrest under transient thermal loading. Two networks, each with its own multiresolution encoding, represent the displacement and phase fields and the transient temperature, and the mechanical energy and the incremental functional of the backward Euler heat equation are estimated on integration points redrawn at every iteration. Restricting irreversibility to the crack set, carrying the optimizer state across load steps and decaying the learning rates within each step help prevent spurious damage accumulation. On two notched plates the initiation peak loads agree closely with an independent finite element reference. On a quenched glass strip under an analytical temperature field the crack path changes from straight to oscillatory and then to branched, with an oscillation threshold close to the linear stability prediction. Under a prescribed boundary temperature a quenched ceramic slab forms a crack array of the type observed in experiment, with a spacing close to that of a published benchmark. On the slab and a quenched annulus the arrays are denser than those of a finite element computation on the same material field. No crack paths, nucleation sites or crack-adapted meshes are prescribed.
BibTeX
@misc{zhang2026mesh,
  author = {Han Zhang and Mehrisadat Makki Alamdari and Babak Shahbodagh and Mohammad Vahab and Cosmin Anitescu and Timon Rabczuk and Elena Atroshchenko},
  title = {{A mesh-free deep energy method for phase-field modeling of thermally driven brittle fracture}},
  year = {2026},
  eprint = {2611.17630},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17630},
}
Figure