🤖 machine learning

A mesh-free multiresolution deep energy method with phase-field modeling of brittle fracture

This paper proposes a mesh-free multiresolution deep energy method that utilizes a single neural network with $C^1$ quadratic B-spline encoding and dynamic stratified Monte Carlo integration to efficiently model brittle fracture via phase-field theory, achieving high accuracy in load-displacement predictions and superior zero-shot crack classification compared to existing baselines.

Original authors: Han Zhang, Mehrisadat Makki Alamdari, Babak Shahbodagh, Mohammad Vahab, Cosmin Anitescu, Timon Rabczuk, Elena Atroshchenko

Published 2026-08-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Han Zhang, Mehrisadat Makki Alamdari, Babak Shahbodagh, Mohammad Vahab, Cosmin Anitescu, Timon Rabczuk, Elena Atroshchenko

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Predicting how a solid object breaks is one of the hardest puzzles in engineering. When a piece of glass shatters or a metal beam snaps, the damage does not happen in a neat, predictable line; it follows a jagged path that depends on tiny flaws inside the material and the exact forces pushing on it. For decades, engineers have tried to simulate this process on computers. The traditional approach treats a crack as a sharp, distinct line that must be tracked as it grows. This requires the computer to constantly redraw its internal map of the object, adding more detail where the crack is moving and removing it where the material is safe. It is a laborious process that struggles when cracks branch, merge, or change direction unexpectedly. A newer idea, called phase-field modeling, avoids this tracking altogether. Instead of drawing a line, it treats the crack as a blurry zone where the material is gradually weakening, like a fog that thickens until the material is gone. This removes the need to chase the crack, but it creates a new problem: the computer must be incredibly detailed everywhere, just in case the crack appears in a place no one expected.

A team of researchers has now combined this blurry-crack idea with a different kind of computer tool: neural networks. These are the same types of programs used to recognize faces or translate languages, but here they are used to describe the shape and movement of a breaking object. The researchers found that while neural networks are powerful, they tend to produce overly smooth results, which makes it hard for them to capture the sharp, localized damage of a breaking crack. To fix this, they built a special system where the network receives its information through a multi-layered grid, similar to a map that has both a broad overview and a zoomed-in view. This allows the computer to represent the fine details of a crack without needing to train for hours to learn them. Furthermore, to ensure the computer does not find a solution that looks good only at specific points, the team changed the points the computer checks at every single step of the calculation. It is like asking a surveyor to check a field for a hidden path, but instead of checking the same ten spots every time, the surveyor picks a completely new set of spots for every new question. This forces the computer to find a solution that is true everywhere, not just at a few lucky locations.

The researchers tested this new method on a series of challenging scenarios, starting with simple bars that had a small notch cut into them. When they pulled these bars apart, the computer predicted the force required to break them with an accuracy of within about one percent compared to the best existing methods. They then tested the system on more complex situations where cracks had to curve, split into two, or merge with other cracks. In every case, the new method reproduced the correct patterns of damage without ever needing to redraw the mesh or track the crack tip. One particularly difficult test involved a plate with a hole in the middle; the computer correctly predicted that a crack would start at the edge of the hole where the stress was highest, even though no crack was there at the beginning. The method also handled curved shapes, like a thick ring, by mapping the geometry exactly without needing to approximate it with straight lines.

The most rigorous test came from a public dataset containing one hundred random arrangements of multiple cracks inside a plate. In this challenge, the researchers ran their method twenty times on different random configurations without any prior training on those specific patterns. The system correctly identified whether each of the hundreds of cracks would grow or stay dormant in ninety percent of the cases. This was a significant improvement over previous attempts using similar neural network tools, which failed to produce consistent results on the same dataset. The researchers noted that their method is not yet faster than the traditional computer programs used in industry; in fact, it takes longer to run a single simulation. However, the value lies in its robustness and its ability to handle complex, changing crack networks without human intervention or manual adjustments. By proving that a neural network can solve these problems when paired with the right grid system and a constantly changing set of check-points, the work opens a path toward simulating fracture in ways that were previously too difficult or unstable for computers to handle.

Technical Summary: A Mesh-Free Multiresolution Deep Energy Method for Phase-Field Modeling of Brittle Fracture

Problem Statement
The prediction of crack nucleation and growth in brittle materials remains a central challenge in computational solid mechanics. Traditional sharp-crack methods require explicit tracking of crack geometry, remeshing, or enrichment strategies, which become complex when cracks branch, merge, or interact. Phase-field modeling addresses this by recasting fracture as the minimization of an energy functional, replacing the explicit crack set with a scalar phase field that localizes in a diffuse band of width proportional to a regularization length $l$ . However, this approach imposes strict resolution requirements: the discretization must resolve the localization band, and the path is unknown a priori.

Existing neural network-based solvers (Deep Ritz/Deep Energy methods) face two specific failure modes when applied to phase-field fracture:

Smoothness Bias: Standard neural networks with smooth activations are biased toward smooth functions, making it difficult to represent the sharp localization band without slow training or global oscillatory features that induce spurious Gibbs-like oscillations.
Fixed Point Set Exploitation: When integration points are fixed, the optimizer can find "zero-energy modes"—configurations that minimize the estimated energy at the sampled points while the fields deteriorate between them. This leads to non-physical crack propagation or failure to propagate.

Previous remedies often re-introduced mesh-like structures (e.g., adaptive quadrature grids, finite element interpolations for gradients) or relied on weight regularization that could suppress the necessary rapid variations near the crack.

Methodology
The authors propose a fully mesh-free deep energy method that addresses both failure modes within the discretization itself, without reverting to mesh-based structures. The core components are:

Multiresolution Feature Encoding:
Instead of relying solely on the network's capacity, the input coordinates are augmented with a multiresolution feature encoding built from $C^1$ quadratic B-spline grids.
- Resolution Cap: The finest grid spacing $h$ is set as a fixed fraction of the regularization length $l$ (e.g., $h = 0.26l$ ). This establishes a hard limit on the finest scale the representation can express, ensuring the crack band is representable from the first iteration.
- $C^1$ Continuity: The use of quadratic B-splines ensures global $C^1$ continuity. This is critical for the fourth-order fracture energy density (which involves the Laplacian $\Delta \phi$ ), allowing the model to run on a single discretization without requiring higher-order trial spaces or mixed formulations.
- Hierarchy: Coarse grids capture the smooth large-scale response, while the finest grid captures the localized residual, mimicking multigrid logic to accelerate training.
Resampled Monte Carlo Integration:
The energy functional is estimated via stratified Monte Carlo integration.
- Dynamic Sampling: Integration points are redrawn at every optimizer iteration (or every few iterations). This eliminates any fixed point set, preventing the optimizer from exploiting specific configurations to lower the estimated energy artificially.
- Stratification: The sampling density is a mixture of three strata: a uniform bulk stratum, a crack stratum (concentrated on existing and seeded cracks), and a process-zone stratum (concentrated on the damage band and ahead of the crack tip). This concentrates resolution where it is needed without a fixed mesh.
Field Representation and Boundary Conditions:
- A single neural network outputs raw displacement and phase field values.
- Exact Boundary Conditions: Essential boundary conditions are imposed exactly via "lifts" (multiplying network outputs by envelope functions that vanish on boundaries and adding prescribed data), avoiding penalty weights.
- Curved Domains: Curved geometries are handled via an exact NURBS map. Interior holes are managed via a domain mask (indicator function) rather than boundary-fitted meshes, with the mask applied directly to the energy density during integration.
- Irreversibility: Crack irreversibility is enforced via a one-sided quadratic penalty on the phase field relative to the previous load step, rather than history fields, to maintain compatibility with the resampling scheme.

Key Contributions

Resolution via Encoding, Not Training: The ability to resolve the crack band is determined by the choice of the feature grid resolution, not by the network's ability to learn sharp features through slow training.
Stability via Resampling: The combination of a resolution cap and per-iteration point resampling is shown to be necessary; removing either ingredient causes the crack to fail to propagate or produce non-physical results.
Unified Discretization for Higher-Order Models: The $C^1$ continuity of the encoding allows the fourth-order fracture energy density to be solved on the exact same discretization as the second-order model, requiring only an additional automatic differentiation pass.
Zero-Shot Generalization: The method is applied to a public benchmark dataset of random multi-crack configurations without any training data or per-configuration tuning, outperforming the dataset authors' Deep Ritz baseline.

Results
The method was assessed on six numerical examples:

Single-Edge-Notched (SEN) Tension and Shear: The computed load-displacement curves matched staggered finite element references within ~1% for peak loads in tension and ~3% in shear. The crack paths and phase fields agreed closely with references.
Crack Branching: The method successfully reproduced symmetric Y-shaped branching patterns without explicit tracking or kinking criteria, matching finite element results for the undecomposed energy functional.
Coalescence of En-Echelon Cracks: Three pre-existing cracks merged into a single staircase pattern, with peak loads within 6% of the reference.
Plate with a Circular Hole: The elastic stage was verified against the Kirsch solution. Nucleation occurred naturally at the points of maximum hoop stress without pre-existing cracks.
Thick-Walled Ring: A curved domain was solved on a single NURBS patch, reproducing the crack path reported in isogeometric literature.
Random Multi-Crack Benchmark: On a dataset of 10 random configurations (11–17 cracks each), the method correctly classified the active/dormant state of 90% of seeded cracks in zero-shot runs. It achieved mean Dice scores of 0.739 (tension) and 0.824 (shear), outperforming the dataset authors' trained surrogates (0.680 and 0.733) and the Deep Ritz baseline (which failed to produce consistent patterns).

Significance and Claims
The paper claims that the proposed discretization resolves the fundamental tension between the need for high expressiveness (to represent the crack band) and the need for stable integration (to avoid zero-energy modes) in mesh-free energy minimization.

Robustness: The method is robust to network initialization and sampling sequences, a significant improvement over previous neural network approaches that produced different crack patterns for different random seeds.
Flexibility: It handles complex topologies (branching, merging), curved geometries, and higher-order models without remeshing, tracking, or multi-patch coupling.
Limitations: The authors explicitly state that the method is not a faster route to a single quasi-static solution; the computational cost per load increment is roughly an order of magnitude higher than finite element methods. The value proposition lies in the elimination of mesh management, the ability to handle complex crack topologies naturally, and the robustness of the solution.
Scope: The study is limited to 2D quasi-static problems with a fixed regularization length. The hybrid strain energy decomposition used is noted as not variationally consistent, though it is accepted as a practical necessity for mode II loading.

The work demonstrates that a mesh-free deep energy method can achieve accuracy comparable to finite element references while offering the geometric flexibility and topological freedom inherent to phase-field modeling, provided the discretization is carefully designed to prevent the specific failure modes of neural network solvers.

Technical Summary: A Mesh-Free Multiresolution Deep Energy Method for Phase-Field Modeling of Brittle Fracture

More like this