betaFeedback

WorksPhysics and Astronomy

Fracture-Bench: A Controlled Benchmark Suite for Machine Learning Surrogates in Phase-Field Fracture Mechanics

Abstract

Machine-learning surrogates are increasingly proposed to replace phase-field fracture simulations in many-query tasks such as uncertainty quantification, inverse analysis and design. Little controlled evidence shows which architectures are reliable, for which fracture problems and how far beyond their training data. We introduce Fracture-Bench, an open benchmark of two phase-field datasets with fixed splits, common test metrics and a description of each dataset's physical regime. Dataset~1 contains 1,000 AT2 simulations of multiple cracks growing from a notched hole in a hyperelastic plate under random non-uniform biaxial boundary displacements (44 load steps, 128×128128\times128 grid). Dataset~2 contains 4,000 PF-CZM simulations of a pre-cracked plate with two inclusions, each ten times stiffer or more compliant than the matrix (128×256128\times256 grid). We train a Fourier neural operator (FNO), DeepONet, Latent DeepONet, a UNet and a conditional diffusion model on identical splits and compare one test metric per dataset, training time, inference time and model size. On Dataset~1, FNO and Latent DeepONet attain mean test errors of order 10−610^{-6}, about ten times lower than UNet and DeepONet. On Dataset~2, the diffusion model is most accurate, with about one sixth of the UNet error, but requires \SI{706}{s} per prediction versus \SI{6}{ms} for UNet. For all architectures, most of the error occurs in the final load steps, after crack localization. Dataset~2 contains initial cracks less than one grid cell apart that produce qualitatively different paths. Architecture rankings therefore depend on the task, and errors averaged over densely sampled inputs reveal little about near-bifurcation cases. Data, codes, splits, regime descriptors and generalization splits are public, with a protocol for adding under-represented regimes such as mixed-mode, three-dimensional and experimentally paired fracture.

Highlights

  • Open benchmark of 5,000 phase-field fracture simulations with fixed splits

  • Five surrogates compared on accuracy, training cost, inference cost and size

  • No architecture is best on both datasets; rankings depend on the task

  • Errors concentrate where cracks localize and near crack-path bifurcations

  • Regime descriptors and generalization splits show where surrogates are tested

Introduction

Fracture often sets the safe life of engineering components such as aircraft engines, pipelines, reactor pressure vessels and additively manufactured parts. It often occurs abruptly, propagates rapidly and can lead to catastrophic failure. Because fracture cannot yet be predicted with confidence, design and certification still rely on extensive testing and conservative margins. A national study for the US National Bureau of Standards estimated the annual cost of fracture in the United States at $88 billion in 1978 dollars, about 4% of the gross national product, of which roughly a third could have been avoided with the technology then available [1]. Fully qualifying a single additively manufactured material by empirical testing often requires many thousands of individual tests, costing millions of dollars and taking 5 to 15 years [2]. The NASA-sponsored Vision 2040 roadmap envisions product certification that relies heavily on simulation, with physical testing only where necessary [3]. Reaching that goal for fracture requires predictions that are both fast and trustworthy.

Griffith [4] and Irwin [5] founded fracture mechanics on the energy release rate and the stress intensity factor. These criteria can be evaluated analytically only for simple geometries and loadings, and they do not describe the nucleation, branching or coalescence of cracks, which require solving the balance of momentum coupled with an evolving crack. Heterogeneous materials add difficulty, because spatial variations of stiffness and toughness make the crack path sensitive to the material layout [6]. Sharp-crack methods struggle with these features. The extended finite element method (X-FEM) [7] decouples the crack from the mesh through enrichment [8, 9, 10], but it needs crack tracking, which becomes cumbersome when cracks branch or merge, and its predictions depend on the propagation criterion and the mesh [11, 12]. Cohesive zone models confine the crack to element boundaries. Phase-field models [13, 14, 15, 16, 17] and their cohesive extensions [18, 19] instead smear the crack over a band whose width is set by a length scale ℓo\ell_o, so that nucleation, branching and merging follow from energy minimization without tracking. The price is a fine mesh in the band and an iterative coupled solve at every load step. Adaptive sub-stepping [20], robust monolithic solvers [21] and adaptive mesh refinement [22, 23, 24, 25, 26, 27, 28] reduce this cost. Even so, a specimen-scale simulation can take hours, which puts Monte Carlo uncertainty propagation, Bayesian inference and design optimization out of reach. Surrogate models address this bottleneck by learning an inexpensive approximation from precomputed solutions.

Neural networks were first applied to fracture by Theocaris and Panagiotopoulos [29]. Later surrogates estimated failure paths and damage evolution from high-fidelity simulations [30], but they predicted low-dimensional quantities and did not enforce the governing equations. Physics-informed neural networks (PINNs) build the governing equations or the energy functional into the loss, as Goswami et al. [31] did for phase-field crack paths. A PINN, however, must be retrained for each new geometry, loading or material. Neural operators instead learn a mapping between function spaces, so one model covers a family of inputs and, in principle, any discretization. The most widely used are DeepONet [32] and the Fourier neural operator (FNO) [33]. Latent DeepONet [34] learns the operator in the latent space of an autoencoder, which helps for high-dimensional outputs. Physics-informed variants [35] include a DeepONet trained with a variational loss to predict crack paths in quasi-brittle materials [36]. Convolutional and generative models treat prediction as an image-to-image task on a fixed grid. The UNet [37], an encoder–decoder with skip connections, has been applied to elasto-plastic fields in two-phase composites [38] and to spall failure in ductile polycrystals [39]. Denoising diffusion models [40] generate the field by iterative denoising. They preserve sharp features that regression tends to smooth and return samples from a distribution, but they need many network evaluations per prediction. Trained surrogates can be orders of magnitude faster than the solvers they replace [41, 42], including for phase-field simulations [43, 44].

Experience in other fields shows that such methods mature once an open, curated dataset exists. AlphaFold [46] was trained on the Protein Data Bank, which began in 1971 with seven structures [47]. Weather models trained on the ERA5 reanalysis [48] now have surpassed physics-based forecasts [49, 50], and have run operationally at the European Centre for Medium-Range Weather Forecasts since 2025 [51, 52]. Fracture has no comparable resource. Curated databases are licensed, industrial test data are proprietary, and open simulation data occupy a narrow regime. The largest collections are all two-dimensional and tensile: 400,000 dynamic phase-field simulations under uniaxial or biaxial tension [53], and 70,000 and 6,000 quasi-static brittle simulations in Mechanical MNIST Crack Path [54] and PFM-Fracture [44]. In our audit of eleven public sources, more than 99.7% of roughly 700,000 simulations describe two-dimensional brittle specimens under quasi-static tension, and none covers mixed-mode loading, physically grown three-dimensional cracks or a matching experiment. Experiments share this bias: standard toughness tests measure the opening mode [55], whereas many service failures involve mixed-mode cracks [56]. Coverage matters because purely data-driven neural operators are usually reliable only within their training data and can fail under extrapolation [57, 58]. Fracture adds sensitivity to inputs. In the first Sandia Fracture Challenge, thirteen teams made blind predictions of ductile tearing [59]. Nominally identical specimens failed

along tensile- or shear-dominated paths, a difference later traced to variation in the location of a hole, with out-of-tolerance specimens favoring the shear path [60]. No team had accounted for such tolerances [61]. A fracture benchmark should therefore report not only an average error but also where in the input space the predictions hold.

Each new architecture, proposed by a research group, is usually demonstrated on its authors’ own problem, the literature does not tell a practitioner which surrogate to choose for a given fracture problem. Differences in datasets, splits, preprocessing, losses, training budgets and metrics often outweigh the differences between the models themselves. Controlled comparisons exist for smooth problems, for example of DeepONet and FNO on common FAIR data [62], but general scientific-ML benchmarks such as PDEBench [63] and The Well [64] do not include fracture. For fracture, Hamdi and Lejeune [44] introduced PFM-Fracture: 6,000 two-dimensional quasi-static brittle simulations spanning three energy decompositions, two loading conditions and 1,000 random initial cracks. They evaluated a PINN, an FNO and a UNet, with ensembling. Fracture-Bench complements that effort in four ways. First, it covers different physics: Dataset 1 combines finite-deformation hyperelasticity, a vector-valued phase field and random non-uniform boundary loading, and Dataset 2 treats cohesive fracture in a plate with inclusions ten times stiffer or ten times more compliant than the matrix. Second, it compares five architectures from three families (neural operators, convolutional networks and diffusion models) on accuracy, training and inference cost, and model size, with the datasets, splits and metrics fixed for all. Third, each dataset carries a regime descriptor of the governing variables it covers, the effective dimension of its inputs and what it lacks, and the benchmark defines splits that test generalization beyond random held-out samples. Fourth, the datasets, simulation codes, training code and an interactive dashboard are public, and § 3.5 describes how to add architectures and fracture problems.

Both datasets lie within the dominant regime of open fracture data: they are two-dimensional and quasi-static. We do not claim that they fill the coverage gap described above. Their purpose is to provide a controlled starting point on which architectures can be compared on equal terms, and a format, including the regime descriptor and the contribution protocol, through which datasets from under-represented regimes can be added and compared on the same footing. Our contributions are:

  • Two high-fidelity simulation datasets generated with published phase-field solvers [65, 6]:

    • Dataset 1: 1,000 simulations of brittle fracture in a perforated hyperelastic plate under boundary loading drawn from a periodic Gaussian random field (GRF) (128×128128 \times128 grid, 44 load steps, AT2 formulation) [45]; https://doi.org/10.7281/T1XFF19D.

    • Dataset 2: 4,000 simulations of cohesive fracture and crack–inclusion interaction in heterogeneous plates containing two circular inclusions, generated with adaptive mesh refinement (128×256128 \times256 grid, 31 load steps, PF-CZM formulation) [66]; https://doi.org/10.7281/T1RZFI3I.

  • A regime descriptor for each dataset (Table 1), which records the governing variables covered, the effective dimension of the sampled inputs, the resolution of the crack band on the benchmark grid and the information that the dataset does not contain.

  • A reference study of five surrogate architectures, FNO, DeepONet, Latent DeepONet, UNet and a conditional denoising diffusion probabilistic model (cDDPM), trained on identical data splits and scored with common test metrics, together with training and inference wall-clock time, parameter count and temporal interpolation at load steps withheld from training.

  • An open, extensible platform. The datasets, the simulation codes, the training and evaluation code (https://github.com/Centrum-IntelliPhysics/fracture_bench) and an interactive dashboard (https://fracture-bench.github.io) are public, and the benchmark is designed so that the community can add both new architectures and new fracture problems (§ 3.5).

Dataset 1Dataset 2
Dimension, kinematics2D, finite deformation2D plane strain, small strain
Domain2.2 × 2.2 (non-dimensional)4 × 8 mm
Bulk modelNeo-HookeanIsotropic linear elastic
Material parametersμ=1\mu= 1, λ=0\lambda= 0 (ν=0\nu= 0), Gc=1.721×10−5G_c = 1.721 \times10^{-5}, ℓo=0.015\ell_o = 0.015Matrix: E=377E = 377 GPa, ν=0.3\nu= 0.3, Gc=10G_c = 10 N/m, ℓo=0.04\ell_o = 0.04 mm; inclusions: EE and GcG_c multiplied by 10 (stiffer) or 0.1 (less stiff)
Fracture modelAT2, vector phase field d∈R2\mathbf{d} \in\mathbb{R}^2, crack-face contactPF-CZM, scalar phase field dd, linear softening
LoadingDisplacement control, non-uniform biaxial tensionDisplacement control, uniaxial tension (Mode I)
Mode mixityNominally opening; local mixity at notches not quantifiedNominally Mode I; local mixity at inclusions
RateQuasi-static, 44 incrementsQuasi-static, 31 increments
HeterogeneityHomogeneous; central hole with eight notchesTwo circular inclusions of fixed size and position
Sampled inputsGRF boundary profile ffCrack-tip location (xc,yc)(x_c, y_c); inclusion materials
Effective input dimension≈4\approx4 (ξk,ηk\xi_k, \eta_k for k≤2k \le2)2 continuous + 2 binary
SymmetryMirror symmetry about x1=x2x_1 = x_2None
Records (train/test)1,000 (900/100)4,000 (3,600/400)
Grid, spacing hh1282128^2, h≈0.0173h \approx0.0173128×256128 \times256, h≈0.0315h \approx0.0315 mm
Crack band on gridℓo/h=0.87\ell_o/h = 0.87; d>0.5d > 0.5 band ≈1.2h\approx1.2hπℓo/h≈4.0\pi\ell_o/h \approx4.0; d>0.5d > 0.5 band ≈1.3h\approx1.3h
Stored fieldsu\mathbf{u}, dd at all 44 stepsu\mathbf{u}, dd, EE at all 31 steps
Known data issuesOne training sample (index 257) stops after step 42; steps 43–44 are zero34 samples have no initial crack in d(x,0)d(\mathbf{x}, 0)
Paired experimentsNoneNone
Not covered3D; dynamics; plasticity; cyclic loading; shear-dominated loading; geometry variation; experiments3D; dynamics; plasticity; cyclic loading; variation of inclusion size or position; experiments

Table 1. Regime descriptors of the two datasets. “Effective input dimension” counts the independent random variables that control the sampled inputs; for Dataset 1, it counts the GRF coefficients ξk\xi_k, ηk\eta_k of the modes that carry 99.9% of the variance of ff. hh is the spacing of the benchmark grid. Dataset 1 is non-dimensional: its parameters approximately correspond to PMMA, all quantities are normalized by the shear modulus μ\mu, and λ=0\lambda= 0 (equivalently ν=0\nu= 0) is chosen to remove the effect of Poisson’s ratio [45].

Section 2 describes the two datasets, their regime descriptors, the surrogate architectures, the training and evaluation protocols, and how the benchmark can be used and extended. Section 3 presents the benchmark results and discusses what they do and do not show. Section 4 summarizes the findings.

Methodology

This section describes the datasets, the surrogate architectures and the evaluation protocols. Dataset 1 captures crack pattern evolution in a perforated hyperelastic domain under spatially varying boundary displacements using the AT2 phase field formulation, while Dataset 2 examines crack–inclusion interactions in a heterogeneous plate solved with the PF-CZM phase field formulation. For each dataset we state the problem setup, the governing equations and the key simulation parameters. Five surrogate architectures are then described, and finally the training and evaluation protocols that are applied to all of them.

Phase-field fracture: governing equations

Consider a linear elastic body Ω⊂R2\Omega\subset\mathbb{R}^{2} with boundary ∂Ω\partial\Omega partitioned into a Dirichlet part ∂Ωu\partial\Omega_{u}, on which displacements u∗(x)\boldsymbol{u}^{*}(x) are prescribed, and a Neumann part ∂Ωt\partial\Omega_{t}, on which tractions are applied, such that ∂Ωu∪∂Ωt=∂Ω\partial\Omega_{u} \cup\partial\Omega_{t} = \partial\Omega and ∂Ωu∩∂Ωt=∅\partial\Omega_{u} \cap\partial\Omega_{t} = \varnothing. Let u(x)\boldsymbol{u}(x) denote the displacement field and d(x)∈[0,1]d(x) \in[0,1] the scalar damage (phase-field) variable at a point x∈Ωx \in\Omega, with d=0d = 0 representing intact material and d=1d = 1 a fully broken state. This scalar, small-strain formulation applies directly to Dataset 2 and is presented here as the common reference. Dataset 1 uses the vector-valued extension of Hakimzadeh et al. [65], in which the phase field is a vector d\boldsymbol{d} whose magnitude ∥d∥∈[0,1]\lVert\boldsymbol{d}\rVert\in[0,1] plays the role of the scalar damage dd and whose direction gives the orientation of the crack normal, and in which the bulk response is neo-Hookean at finite deformation; the choices specific to each dataset are given in § 3.2. Within the second-order phase-field framework [24], the system Lagrangian L(u,d)\mathcal{L}(\boldsymbol{u},d) is written as

L(u,d)=K−∫Ωψs(ϵ,d) dV−∫ΩGfcα[α(d)ℓo+ℓo(∇d⋅∇d)] dV+P(u),(1)\mathcal{L}(\boldsymbol{u},d) = \mathcal{K} - \int_{\Omega} \psi_{s}(\boldsymbol{\epsilon},d)\,\mathrm{d}V - \int_{\Omega} \frac{G_{\mathrm{f}}}{c_{\alpha}} \left[\frac{\alpha(d)}{\ell_{o}} + \ell_{o}(\nabla d \cdot\nabla d)\right]\,\mathrm{d}V + \mathcal{P}(\boldsymbol{u}), \tag*{(1)}

where ϵ=12(∇u+∇uT)\boldsymbol{\epsilon} = \frac{1}{2}(\nabla\boldsymbol{u} + \nabla\boldsymbol{u}^{\mathrm{T}}) is the infinitesimal strain tensor, ψs(ϵ,d)\psi_{s}(\boldsymbol{\epsilon},d) is the strain energy density of the damaged solid, GfG_{\mathrm{f}} is the material fracture energy (the energy dissipated per unit area of crack surface), and ℓo\ell_{o} is the phase-field regularization length scale that controls the width of the diffuse crack band. The geometric crack function α(d)\alpha(d) governs the spatial profile of the damage field across the crack band, and cα=4∫01α(β) dβc_{\alpha} = 4\int_{0}^{1}\sqrt{\alpha(\beta)}\,\mathrm{d}\beta is a normalization constant ensuring that the regularized surface energy converges to GfG_{\mathrm{f}} times the crack area in the limit ℓo→0\ell_{o} \to0. The kinetic energy K\mathcal{K} and the external work P\mathcal{P} are given by

K:=12∫Ωρ u˙⋅u˙ dV,P:=∫Ωb∗⋅u dV+∫∂Ωtt∗⋅u dA,(2)\mathcal{K} := \frac{1}{2}\int_{\Omega}\rho\,\dot{\boldsymbol{u}}\cdot\dot{\boldsymbol{u}}\,\mathrm{d}V,\qquad\mathcal{P} := \int_{\Omega}\boldsymbol{b}^{*}\cdot\boldsymbol{u}\,\mathrm{d}V + \int_{\partial\Omega_{t}}\boldsymbol{t}^{*}\cdot\boldsymbol{u}\,\mathrm{d}A, \tag*{(2)}

where ρ\rho is the mass density, u˙\dot{\boldsymbol{u}} is the velocity field, b∗\boldsymbol{b}^{*} is the prescribed body force per unit volume, t∗\boldsymbol{t}^{*} is the prescribed surface traction on ∂Ωt\partial\Omega_{t}, and dV\mathrm{d}V and dA\mathrm{d}A denote volume and surface measures, respectively.

To prevent crack growth under compression and interpenetration of crack faces, the strain energy density ψs(ϵ,d)\psi_{s}(\boldsymbol{\epsilon},d) in Equation (1) is decomposed into tensile and compressive contributions as

ψs(ϵ,d)=ψ0−(ϵ)+ω(d)ψ0+(ϵ).(3)\psi_{s}(\boldsymbol{\epsilon},d) = \psi_{0}^{-}(\boldsymbol{\epsilon}) + \omega(d)\psi_{0}^{+}(\boldsymbol{\epsilon}). \tag*{(3)}

where ψ0+\psi_0^{+} and ψ0−\psi_0^{-} are the tensile and compressive parts of the undamaged elastic strain energy density, obtained here via the principal (spectral) strain decomposition [16], and ω(d)\omega(d) is the energetic degradation function, a monotonically decreasing function with ω(0)=1\omega(0)=1 and ω(1)=0\omega(1)=0 that reduces the tensile stiffness as damage accumulates. Only the tensile part is degraded; the compressive part ψ0−\psi_0^{-} remains unaffected.

Applying Hamilton’s principle of least action and enforcing the irreversibility and boundedness of the damage field2, the strong-form governing equations for the cracking solid read

{∇⋅σ+b∗=ρu¨in Ω,ω′(d)Y‾+Gfcα(α′(d)ℓo−2ℓoΔd)≥0,d˙≥0in Ω,(4)\left\{ \begin{aligned} \nabla\cdot\boldsymbol{\sigma}+\boldsymbol{b}^{*} &= \rho\ddot{\boldsymbol{u}} && \text{in } \Omega,\\ \omega'(d)\overline{Y}+\frac{G_{\mathrm{f}}}{c_{\alpha}}\left(\frac{\alpha'(d)}{\ell_{o}}-2\ell_{o}\Delta d\right) &\ge0,\qquad\dot{d}\ge0 && \text{in } \Omega, \end{aligned} \right. \tag*{(4)}

with natural boundary conditions

{σ⋅n=t∗on ∂Ωt,∇d⋅n=0on ∂Ω,(5)\left\{ \begin{aligned} \boldsymbol{\sigma}\cdot\boldsymbol{n} &= \boldsymbol{t}^{*} && \text{on } \partial\Omega_{t},\\ \nabla d\cdot\boldsymbol{n} &= 0 && \text{on } \partial\Omega, \end{aligned} \right. \tag*{(5)}

where σ:=∂ψ0−/∂ϵ+ω(d) ∂ψ0+/∂ϵ\boldsymbol{\sigma}:=\partial\psi_0^{-}/\partial\boldsymbol{\epsilon}+\omega(d)\,\partial\psi_0^{+}/\partial\boldsymbol{\epsilon} is the Cauchy stress tensor, u¨\ddot{\boldsymbol{u}} is the acceleration field, ω′(d)\omega'(d) and α′(d)\alpha'(d) denote derivatives with respect to dd, Y‾=ψ0+\overline{Y}=\psi_0^{+} is the effective crack driving energy, Δd:=∑i∂2d/∂xi2\Delta d:=\sum_i \partial^2d/\partial x_i^2 is the Laplacian of the damage field, and n\boldsymbol{n} is the outward unit normal to the boundary. The inequality form of the damage evolution equation, together with the rate constraint d˙≥0\dot{d}\ge0 and the bounds 0≤d≤10\le d\le1, expresses the Karush–Kuhn–Tucker conditions enforcing crack irreversibility. The problem is completed by the Dirichlet conditions u=u∗(x)\boldsymbol{u}=\boldsymbol{u}^{*}(\boldsymbol{x}) on ∂Ωu\partial\Omega_u and the initial conditions u(x,0)=u0(x)\boldsymbol{u}(\boldsymbol{x},0)=\boldsymbol{u}_0(\boldsymbol{x}) and u˙(x,0)=v0(x)\dot{\boldsymbol{u}}(\boldsymbol{x},0)=\boldsymbol{v}_0(\boldsymbol{x}).

The choice of the pair {α(d),ω(d)}\{\alpha(d),\omega(d)\} distinguishes the principal PFM variants. The AT2 model [14, 16], the most widely used in engineering practice, employs α(d)=d2\alpha(d)=d^2 and ω(d)=(1−d)2\omega(d)=(1-d)^2; however, it lacks an elastic domain, meaning damage initiates immediately upon any loading no matter how small. The PF-CZM [18, 68] generalizes this through a rational degradation function,

ω(d)=(1−d)2(1−d)2+a1d⋅P(d),a1=4π⋅ℓchℓo,(6)\omega(d)=\frac{(1-d)^2}{(1-d)^2+a_1d\cdot P(d)},\qquad a_1=\frac{4}{\pi}\cdot\frac{\ell_{\mathrm{ch}}}{\ell_o}, \tag*{(6)}

with α(d)=2d−d2\alpha(d)=2d-d^2 and P(d)=1+a2d+a2a3d2P(d)=1+a_2d+a_2a_3d^2, where the coefficients {a2,a3}\{a_2,a_3\} are calibrated to reproduce a target traction-separation law (TSL); see Equation (7). Here, E0E_0 denotes Young’s modulus, ftf_t is the material tensile strength, and ℓch=E0Gf/ft2\ell_{\mathrm{ch}}=E_0G_f/f_t^2 is the Irwin’s characteristic length, which measures the intrinsic size of the fracture process zone. The rational form of ω(d)\omega(d) introduces a well-defined elastic threshold and enables PF-CZM to recover accurate peak loads and post-peak softening responses independent of the length scale ℓo\ell_o, extending its applicability to both brittle and cohesive fracture. The two models are summarized in Table 1. The table also lists the damage profile and the damage bandwidth of each model. The damage profile is the one-dimensional solution d(x)d(x) of the damage equation across a fully developed crack located at x=0x=0, obtained by minimizing the fracture energy term of Equation (1) subject to d(0)=1d(0)=1; it describes how the damage decays with the distance ∣x∣|x| from the crack. The damage bandwidth is the width of the zone in which d>0d>0. For AT2 the profile decays exponentially and never reaches zero, so the bandwidth is infinite, whereas for PF-CZM the damage vanishes at a distance πℓo/2\pi\ell_o/2 on either side of the crack, giving a finite bandwidth πℓ0\pi\ell_0. Neither quantity appears in the governing equations; both follow from the choice of α(d)\alpha(d) and set the mesh resolution required around the crack.

In this study, two softening laws are considered for PF-CZM.

{Linear softening:a2=−12,a3=0,Concrete softening:a2=1.3868,a3=0.9106.(7)\begin{cases} \text{Linear softening:} & a_2=-\frac{1}{2},\quad a_3=0,\\ \text{Concrete softening:} & a_2=1.3868,\quad a_3=0.9106. \tag*{(7)} \end{cases}

Mathematical problem statement

The central task is to learn an operator surrogate model that maps parametrized problem inputs to the spatio-temporal evolution of the phase-field damage variable over the spatial domain Ω\Omega. Two distinct phase-field formulations are employed across the two datasets, differing in both the mathematical structure of the phase-field variable and the governing energy functional.

Dataset 1 adopts the vector-valued phase-field formulation of Hakimzadeh et al. [65], in which the damage state is described by a vector field d(x,t)∈R2\mathbf{d}(\mathbf{x},t) \in\mathbb{R}^{2} under large-deformation contact mechanics, governed by the Lagrangian of Equation (1) with the AT2 choices α(d)=d2\alpha(d)=d^{2} and ω(d)=(1−d)2\omega(d)=(1-d)^{2} (see Table 1), where d≡∥d∥d \equiv\lVert\mathbf{d}\rVert is the scalar damage measure (Euclidean norm of the vector phase field). The surrogate predicts the full two-component vector field d(x,t)∈R2\mathbf{d}(\mathbf{x},t) \in\mathbb{R}^{2} across multiple load steps, given the initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0):

G1:u(x,0)↦d(x,t),d∈R2,t∈{t1,…,tNt}.(8)\mathcal{G}_1:\mathbf{u}(\mathbf{x},0)\mapsto\mathbf{d}(\mathbf{x},t),\quad\mathbf{d}\in\mathbb{R}^{2},\quad t\in\{t_1,\ldots,t_{N_t}\}. \tag*{(8)}

The scalar damage indicator ∥d(x,t)∥∈[0,1]\lVert\mathbf{d}(\mathbf{x},t)\rVert\in[0,1] is used for evaluation and visualization throughout.

Dataset 2 adopts the scalar PF-CZM formulation [18, 68], in which the damage state is described by a scalar phase-field d(x,t)∈[0,1]d(\mathbf{x},t)\in[0,1], where d=0d=0 denotes intact material and d=1d=1 denotes fully fractured material, governed by the rational degradation function of Equation (6) and solved with adaptive mesh refinement [6]. The surrogate predicts the scalar phase-field at the final load step, given the initial phase-field configuration, d(x,t=0)d(\mathbf{x},t=0), and the spatially varying Young’s modulus field E(x)E(\mathbf{x}):

G2:(d(x,0),E(x))↦d(x,tf),d∈[0,1],(9)\mathcal{G}_2:\bigl(d(\mathbf{x},0),E(\mathbf{x})\bigr)\mapsto d(\mathbf{x},t_f),\quad d\in[0,1], \tag*{(9)}

where tft_f denotes the final load step. In both datasets, the surrogate replaces a coupled nonlinear finite element solve at each load increment with a single forward pass through the trained network. The relevant mechanical and material properties associated with each dataset are summarized in Table 1.

Dataset 1: Square plate with a central hole

Dataset 1 contains 1,000 high-fidelity quasi-static phase-field fracture simulations on a square domain Ω\Omega with a central circular hole [45]. The simulations were performed with a FEniCS implementation of the vector phase-field model of Hakimzadeh et al. [65], using linear finite elements and a staggered (alternate minimization) solution scheme. The plate is a square of side length L=2.2L=2.2 (normalized units) with a central circular hole of radius rhole=0.5r_{\mathrm{hole}}=0.5. Throughout, x=(x1,x2)\mathbf{x}=(x_1,x_2) denotes coordinates measured from the center of the plate along fixed horizontal (x1x_1) and vertical (x2x_2) axes, so that the plate occupies [−L/2,L/2]2[-L/2,L/2]^2 and its outer boundary ∂Ωu\partial\Omega_u consists of the four edges x1=±L/2x_1=\pm L/2 and x2=±L/2x_2=\pm L/2. The hole boundary is traction-free. Eight radial pre-notches of length 0.05, equally spaced at 45∘45^\circ intervals around the hole rim, are seeded through the initial phase-field and act as crack nucleation sites.

Displacement is prescribed on the entire outer boundary as a spatially modulated biaxial stretch. At load step nn,

u∗(x,tn)=εˉnα(x)x,α(x)={f(x1)on the bottom and top edges, x2=∓L/2,f(x2)on the left and right edges, x1=∓L/2,(10)\mathbf{u}^{*}(\mathbf{x},t_n)=\bar{\varepsilon}_n\alpha(\mathbf{x})\mathbf{x},\qquad\alpha(\mathbf{x})=\begin{cases} f(x_1) & \text{on the bottom and top edges},\ x_2=\mp L/2,\\ f(x_2) & \text{on the left and right edges},\ x_1=\mp L/2, \end{cases} \tag*{(10)}

where εˉn=nΔεˉ\bar{\varepsilon}_n=n\Delta\bar{\varepsilon}, with Δεˉ=10−4\Delta\bar{\varepsilon}=10^{-4} and n=1,…,44n=1,\ldots,44, is the nominal applied strain, and f(⋅)f(\cdot) is a realization of a one-dimensional periodic Gaussian random field (GRF) evaluated at the coordinate that runs along the edge. One realization is drawn per simulation and used on all four edges. Substituting the edge coordinates into Equation (10) gives the explicit conditions

x2=∓L/2:u1∗=εˉnf(x1)x1,u2∗=∓εˉnL2f(x1),x1=∓L/2:u1∗=∓εˉnL2f(x2),u2∗=εˉnf(x2)x2.(11)\begin{aligned} x_2=\mp L/2:\qquad&u_1^{*}=\bar{\varepsilon}_n f(x_1)x_1,\qquad u_2^{*}=\mp\bar{\varepsilon}_n\frac{L}{2}f(x_1),\\ x_1=\mp L/2:\qquad&u_1^{*}=\mp\bar{\varepsilon}_n\frac{L}{2}f(x_2),\qquad u_2^{*}=\bar{\varepsilon}_n f(x_2)x_2. \tag*{(11)} \end{aligned}

which are illustrated in Figure 1.

Figure 1. Boundary conditions for Dataset 1. The plate occupies [−L/2,L/2]2[-L/2,L/2]^2 with L=2.2L=2.2, and all panels use the same plate-centered coordinates (x1,x2)(x_1,x_2). A single LL-periodic GRF realization f(⋅)f(\cdot) is evaluated at the coordinate running along each edge, f(x1)f(x_1) on the bottom and top edges and f(x2)f(x_2) on the left and right edges, and the prescribed displacement u∗=εˉα(x)x\mathbf{u}^{*}=\bar{\varepsilon}\alpha(\mathbf{x})\mathbf{x} follows Equation (10); the expressions next to each edge are those of Equation (11). Arrows show the prescribed boundary displacement and the dashed curve the displaced boundary, both magnified for visibility. Because f(−L/2)=f(L/2)f(-L/2)=f(L/2) (red markers), the displaced boundary is continuous at the corners. The central hole, with its eight radial pre-notches, is traction-free.

The GRF has unit mean and period LL,

f(s)=1+∑k=1K[ξkcos⁡(2πksL)+ηksin⁡(2πksL)],ξk,ηk∼N(0,sk2),sk=6252[(2πk)2+25]−2,(12)f(s)=1+\sum_{k=1}^{K}\left[\xi_k\cos\left(\frac{2\pi ks}{L}\right)+\eta_k\sin\left(\frac{2\pi ks}{L}\right)\right],\quad\xi_k,\eta_k\sim\mathcal{N}(0,s_k^2),\quad s_k=625\sqrt{2}\left[(2\pi k)^2+25\right]^{-2}, \tag*{(12)}

with K=100K=100 modes and mutually independent coefficients, and the 1,000 simulations correspond to 1,000 independent realizations of ff. Because ff is LL-periodic, f(−L/2)=f(L/2)f(-L/2)=f(L/2), so the prescribed displacement is single-valued at the four corners. At the bottom-right corner (L/2,−L/2)(L/2,-L/2), for instance, the bottom-edge and right-edge expressions of Equation (11) give εˉnf(L/2)(L/2,−L/2)\bar{\varepsilon}_n f(L/2)(L/2,-L/2) and εˉnf(−L/2)(L/2,−L/2)\bar{\varepsilon}_n f(-L/2)(L/2,-L/2), respectively, which coincide. With unit mean and a standard deviation of approximately 0.21, ff produces a predominantly tensile biaxial loading whose amplitude varies along the boundary. Each simulation stores the displacement field u(x,tn)\mathbf{u}(\mathbf{x},t_n) and the vector phase-field d(x,tn)\mathbf{d}(\mathbf{x},t_n) at all 44 load steps, interpolated from the finite element mesh onto a uniform 128×128128\times128 grid covering [−L/2,L/2]2[-L/2,L/2]^2, with grid points inside the hole set to zero. Representative snapshots of the GRF boundary profiles, displacement magnitude field ∥u(x,t)∥\lVert\mathbf{u}(\mathbf{x},t)\rVert, and the corresponding phase-field d(x,t)\mathbf{d}(\mathbf{x},t) at selected load steps are shown in Figure 2, demonstrating how the damage field evolves from the initial notched state to the final fracture across the loading sequence.

GRF boundary profiles and spatiotemporal evolution of displacement and phase-field for three representative samples

Figure 2. GRF boundary profiles and spatiotemporal evolution of displacement and phase-field for three representative samples. Top row: GRF realizations f(x)f(x) prescribed along the boundary coordinates for each sample, illustrating the diversity of loading scenarios across the dataset. Remaining rows: Displacement magnitude ∥u(x,t)∥\lVert\mathbf{u}(x,t)\rVert (left column of each sample) and phase-field magnitude ∥d(x,t)∥\lVert d(x,t)\rVert (right column) at selected load steps t∈{1,32,36,40,44}t \in\{1,32,36,40,44\}, showing crack nucleation and propagation from the hole boundary under GRF-driven boundary loading.

To assess surrogate performance under a varying number of timesteps, three cases are evaluated: Case 1 uses 11 time steps (every 4th step from the simulated data), Case 2 uses 22 time steps (every 2nd step from the simulated data), and Case 3 uses all 44 time steps.

The dataset does not prescribe a training objective, and users are free to choose their own. For the reference results in this paper, the test metric, and the training loss of every architecture except Latent DeepONet and the cDDPM (Table 2), is a masked mean squared error (MSE) that excludes the hole Ωhole\Omega_{\mathrm{hole}}, since grid points inside the hole carry no physical information. Specifically, a binary spatial mask M(x)∈{0,1}\mathcal{M}(\mathbf{x})\in\{0,1\} is defined as M(x)=1[x12+x22≥rhole2]\mathcal{M}(\mathbf{x})=\mathbf{1}[x_1^2+x_2^2\ge r_{\mathrm{hole}}^2], where rhole=0.5r_{\mathrm{hole}}=0.5 is the hole radius; both the training fields and the squared residuals are multiplied element-wise by M(x)\mathcal{M}(\mathbf{x}), so that hole-interior grid points contribute zero:

ArchitectureDataset 1Dataset 2
FNOMasked MSEWeighted MAE
DeepONetMasked MSEWeighted MAE
Latent DeepONetMSE on latent codesMAE on latent codes
UNetMasked MSEMSE
cDDPMℓ1\ell_1 noise-matching lossℓ1\ell_1 noise-matching loss
Test metric (all)Masked MSE, Equation (13)MSE, Equation (14)

Table 2. Training loss used for each architecture. All architectures are scored on the test set with the same metric, regardless of the loss used for training. “Weighted MAE” refers to Equation (20). Latent DeepONet is trained on the latent codes of a pretrained autoencoder, which is itself trained with an MSE reconstruction loss.

LMSE=1NΩ∑x∈ΩM(x)∥dpred(x,t)−dtrue(x,t)∥2,(13)\mathcal{L}_{\mathrm{MSE}}=\frac{1}{N_{\Omega}}\sum_{\mathbf{x}\in\Omega}\mathcal{M}(\mathbf{x})\left\lVert\mathbf{d}_{\mathrm{pred}}(\mathbf{x},t)-\mathbf{d}_{\mathrm{true}}(\mathbf{x},t)\right\rVert^2, \tag*{(13)}

where NΩ=128×128N_{\Omega}=128\times128 is the total number of grid points, and ∥⋅∥\lVert\cdot\rVert denotes the Euclidean norm over the two phase-field components.

Dataset 2: Heterogeneous plate with two circular inclusions

Dataset 2 contains 4,000 high-fidelity PF-CZM simulations of crack propagation in a heterogeneous plate containing two circular inclusions [66]. The growth of the defects in the domain was simulated following the formulation and adaptive mesh refinement strategy of Krishnan et al. [6]. The domain is a 4 mm×8 mm4\,\mathrm{mm}\times8\,\mathrm{mm} rectangular plate, Ω=[0,W]×[0,H]\Omega=[0,W]\times[0,H] with W=4 mmW=4\,\mathrm{mm} and H=8 mmH=8\,\mathrm{mm} and coordinates (x1,x2)(x_1,x_2) measured from the bottom-left corner. The two inclusions have fixed diameters of 2 mm (upper, centered at (1.5,6)(1.5,6) mm) and 3 mm (lower, centered at (2,2)(2,2) mm). The material properties for the base matrix and both inclusion types are listed in Table 1.

The plate is loaded in Mode I under displacement control. The bottom edge is fully clamped, u1=u2=0u_1=u_2=0 on x2=0x_2=0. On the top edge only the vertical displacement is prescribed, u2=uˉnu_2=\bar{u}_n on x2=Hx_2=H, where uˉn=nΔuˉ\bar{u}_n=n\Delta\bar{u} with Δuˉ=10−4 mm\Delta\bar{u}=10^{-4}\,\mathrm{mm} and n=1,…,31n=1,\ldots,31; the horizontal displacement on the top edge is left free (zero shear traction). The left and right edges, x1=0x_1=0 and x1=Wx_1=W, are traction-free.

Two inputs vary across the simulations: the inclusion configuration and the initial crack. Each inclusion is assigned, independently, a material that is either stiffer or less stiff than the base matrix, yielding four distinct inclusion configurations illustrated in Figure 3(a-d). Stiffer inclusions (Einc=10EmatE_{\mathrm{inc}} = 10 E_{\mathrm{mat}}) deflect the propagating crack around their boundary, while less stiff inclusions (Einc=0.1EmatE_{\mathrm{inc}} = 0.1 E_{\mathrm{mat}}) attract and guide the crack through them, where EmatE_{\mathrm{mat}} denotes the Young’s modulus of the base matrix. The material properties for the base matrix and both inclusion types are listed in Table 1.

Figure 3. Inclusion configurations and initial cracks in Dataset 2. The plate measures 4 mm × 8 mm. Blue circles denote less stiff inclusions (Einc=0.1EmatE_{\mathrm{inc}}=0.1E_{\mathrm{mat}}), which tend to attract and channel the propagating crack; red circles denote stiffer inclusions (Einc=10EmatE_{\mathrm{inc}}=10E_{\mathrm{mat}}), which act to deflect the crack around their boundary (see Table 1). The four inclusion configurations, labeled as smaller inclusion, larger inclusion, are (a) less stiff, stiffer; (b) stiffer, stiffer; (c) less stiff, less stiff; and (d) stiffer, less stiff. (e) The 1,000 crack-tip seeds (black dots), which lie in the left half of the plate, away from both inclusions and from the top and bottom edges. (f) One representative seed (red circle in (e) and (f)) and its horizontal pre-crack (thick black line), which runs from the left edge to the crack tip (xc,yc)(x_c,y_c) and constitutes the initial condition d(x,0)d(\mathbf{x},0) of the simulation. In (e) and (f) the inclusions are drawn in gray because the seed locations are generated independently of the inclusion configuration.

A set of 1,000 candidate crack-tip locations is generated randomly within the domain, constrained such that the xx-coordinate lies within the left half of the domain and the crack-tip is sufficiently far from both inclusions and the top and bottom boundaries, ensuring each crack can propagate freely toward the inclusions under the applied loading. The generation procedure is illustrated in Figure 3: panel (e) shows all 1,000 admissible crack-tip seeds distributed across the left half of the domain, and panel (f) shows a single representative seed together with its initialized horizontal pre-crack d(x,0)d(x,0), which serves as the initial condition for the corresponding simulation as well as the surrogate models. For each of the four inclusion configurations, the same 1,000 initial cracks are used, yielding a total of 4,000 simulations. The crack path can be highly sensitive to the initial crack location, even when the inclusions are unchanged. Figure 4 shows four pairs of simulations, one pair per inclusion configuration. Within each pair the inclusions are identical and the initial crack tips are about 0.02 mm apart, which is less than one cell of the 128 × 256 grid, yet the final crack patterns differ qualitatively: the pre-existing crack crosses the plate between the inclusions, is drawn toward an inclusion, or arrests while a new crack nucleates elsewhere. Such divergence is not the rule. For most pairs of neighboring initial cracks the final paths nearly coincide, and only a few percent of the pairs whose tips lie within 0.1 mm of each other end in clearly different crack patterns. These near-bifurcation cases are nevertheless what makes the prediction task difficult, because the surrogate must resolve differences in the input that are smaller than one grid cell.

Figure 4. Sensitivity of the crack path to the initial crack location in Dataset 2. Each panel shows two simulations with the same inclusion configuration, labeled as smaller inclusion, larger inclusion: (a) less stiff, stiffer; (b) stiffer, stiffer; (c) less stiff, less stiff; and (d) stiffer, less stiff. Within each pair the initial crack tips (coordinates in mm above each column) are about 0.02 mm apart. Top row: initial phase-field d(x,0)d(\mathbf{x}, 0). Bottom row: phase-field d(x,tf)d(\mathbf{x}, t_f) at the final load step tf=31t_f = 31. Blue and red circles mark less stiff and stiffer inclusions, respectively. Despite nearly identical inputs, the two simulations of each pair end in qualitatively different crack patterns.

Each simulation stores the displacement field and the scalar phase-field at all 31 load steps, together with the Young’s modulus field E(x)E(\mathbf{x}), interpolated from the adaptively refined finite element mesh onto a uniform 128 × 256 grid.

The benchmark task maps the initial phase-field and the modulus field to the phase-field at the final load step, Equation (9); the intermediate load steps and the displacement fields remain available for other tasks (§ 3.5). The dataset is split into 3,600 training and 400 test samples (see Table 1), that is, 900 and 100 per inclusion configuration. As for Dataset 1, the training objective is the user’s choice, and the losses used for the reference results are listed in Table 2. All architectures are scored on the test set with the MSE over the full domain:

LMSE=1NΩ∑x∈Ω(dpred(x,tf)−dtrue(x,tf))2,(14)\mathcal{L}_{\mathrm{MSE}} = \frac{1}{N_{\Omega}} \sum_{\mathbf{x}\in\Omega} \left(d_{\mathrm{pred}}(\mathbf{x},t_f)-d_{\mathrm{true}}(\mathbf{x},t_f)\right)^2, \tag*{(14)}

where NΩ=128×256N_{\Omega}=128 \times256 is the total number of grid points and tf=31t_f=31 denotes the final load step.

Surrogate model architectures

Five surrogate architectures are benchmarked, spanning two broad categories: (i) neural operators, which learn mappings between infinite-dimensional function spaces: Fourier neural operator (FNO), deep operator network (DeepONet), and Latent DeepONet, and (ii) deep generative and convolutional models, which treat the prediction as a structured image-to-image regression or conditional generation task: UNet and a Conditional Denoising Diffusion Probabilistic Model (cDDPM).

Each architecture receives the same input fields, is trained on the same data split within the same computational budget, and is evaluated on the same test set with the same metric, so that the reported errors are directly comparable. The training losses are not identical across architectures; they are listed in Table 2 and explained in § 3.4. These choices define the reference protocol of this study and are not a constraint of the benchmark: users may adopt other losses, preprocessing, or training budgets, provided the results are reported with the common test metrics of § 3.4.

Fourier neural operator (FNO)

The Fourier neural operator [33] learns a mapping between function spaces by parameterizing the integral kernel in Fourier space, enabling discretization-invariant predictions on functions defined over continuous domains. Each FNO layer applies a global convolution in Fourier space combined with a local linear transformation:

v(ℓo+1)(x)=σ(W(ℓo)v(ℓo)(x)+F−1(Rθ(ℓo)⋅F(v(ℓo)))(x))(15)\mathbf{v}^{(\ell_o+1)}(\mathbf{x}) = \sigma\left(W^{(\ell_o)}\mathbf{v}^{(\ell_o)}(\mathbf{x}) + \mathcal{F}^{-1}\left(R_{\theta}^{(\ell_o)} \cdot\mathcal{F}\left(\mathbf{v}^{(\ell_o)}\right)\right)(\mathbf{x})\right) \tag*{(15)}

where v(ℓo)\mathbf{v}^{(\ell_o)} is the feature map at layer ℓo\ell_o, W(ℓo)W^{(\ell_o)} is a pointwise linear transformation, F\mathcal{F} and F−1\mathcal{F}^{-1} denote the forward and inverse Fourier transforms, respectively, Rθ(ℓo)R_{\theta}^{(\ell_o)} is a learnable weight tensor truncated to the lowest Fourier modes, and σ\sigma is the activation function. The input is first lifted to a high-dimensional feature space by a shallow MLP, passed through LL successive FNO layers, and then projected to the output space by a second shallow MLP.

For Dataset 1, the problem requires predicting the spatiotemporal evolution of the vector phase-field d(x,t)\mathbf{d}(\mathbf{x},t) over NtN_t load steps given the initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0). The input tensor is constructed by broadcasting u(x,0)\mathbf{u}(\mathbf{x},0) across all time steps and appending the spatial coordinates (x1,x2)(x_1,x_2) and the normalized time coordinate tt as additional channels, yielding a five-channel spatiotemporal input of shape (Nt×128×128×5)(N_t \times128 \times128 \times5). A three-dimensional FNO3D is used, with Fourier convolutions applied jointly across the two spatial dimensions and the temporal dimension, and ReLU activations. The output is the two-channel vector phase-field d(x,t)\mathbf{d}(\mathbf{x},t) of shape (Nt×128×128×2)(N_t \times128 \times128 \times2). The masked MSE loss of Equation (13) is used during training, excluding the hole region Ωhole\Omega_{\mathrm{hole}}. The Dataset 1 FNO architecture is illustrated in Figure 5(a).

Figure 5. FNO architecture (a) Dataset 1: The initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0) is broadcast across NtN_t time steps and concatenated with spatial coordinates (x,y)(x,y) and normalized time tt to form a five-channel spatiotemporal input. After lifting to a high-dimensional feature space, LL Fourier layers apply global convolutions jointly across the spatial and temporal dimensions via Equation (15). The output is the two-channel vector phase-field d(x,t)∈R2\mathbf{d}(\mathbf{x},t) \in\mathbb{R}^2 over all NtN_t load steps; (b) Dataset 2: The initial scalar phase-field d(x,0)\mathbf{d}(\mathbf{x},0) and Young’s modulus field E(x)E(\mathbf{x}) are concatenated with spatial coordinates (x,y)(x,y) to form a four-channel spatial input. After lifting, LL Fourier layers apply global convolutions across the two spatial dimensions only. The output is the single-channel scalar phase-field d(x,tf)d(\mathbf{x},t_f) at the final load step.

For Dataset 2, the problem reduces to a static state-to-state mapping: given the initial scalar phase-field d(x,0)d(\mathbf{x},0) and the spatially-varying Young’s modulus E(x)E(\mathbf{x}), predict the scalar phase-field d(x,tf)d(\mathbf{x},t_f) at the final load step, tft_f. The two input fields are concatenated as a four-channel spatial tensor of shape (128×256×4)(128 \times256 \times4), where the additional two channels are the spatial coordinates (x1,x2)(x_1,x_2). Since the temporal dimension is absent, we employ FNO2D, with Fourier convolutions restricted to the spatial dimensions and GELU activations. The output is the single-channel scalar phase-field d(x,tf)d(\mathbf{x},t_f) of shape (128×256×1)(128 \times256 \times1). The weighted MAE of Equation (20) is used as the training loss. The Dataset 2 FNO architecture is illustrated in Figure 5(b).

Deep operator network (DeepONet)

DeepONet [32] is grounded in the universal approximation theorem for operators [69] and decomposes the target operator into two subnetworks: a branch network that encodes the input function evaluated at a fixed set of sensor points, xs\mathbf{x}_s, and a trunk network that encodes the continuous query coordinates at which the output is evaluated. The predicted output at query location (x,t)(\mathbf{x},t) is given by:

G(u)(x,t)=∑k=1pbk(u(xs,0))⋅τk(x,t)+b0,(16)\mathcal{G}(\mathbf{u})(\mathbf{x},t) = \sum_{k=1}^{p} b_k(\mathbf{u}(\mathbf{x}_s,0)) \cdot\tau_k(\mathbf{x},t) + b_0, \tag*{(16)}

where bkb_k are the pp embeddings of the branch network for the input function, τk\tau_k are the pp embeddings of the trunk network encoding the query location (x,t)(\mathbf{x},t), and b0b_0 is a learnable bias. The branch and trunk outputs are combined via a dot product, making the architecture highly flexible for operator learning across arbitrary query locations.

For Dataset 1, the branch network processes the initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0) concatenated with the spatial coordinates (x1,x2)(x_1,x_2), yielding a four-channel input of shape (128×128×4)(128 \times128 \times4). The branch network is a CNN with max pooling followed by an MLP, encoding the spatial structure of the input into a latent vector of dimension pp. The trunk network is an MLP that takes the spatiotemporal query coordinates (x,t)∈R3(\mathbf{x},t) \in\mathbb{R}^{3} as input. The branch and trunk outputs are combined via the dot product of Equation (16) followed by a tanh activation, and the output is the two-channel vector phase field d(x,t)∈R2\mathbf{d}(\mathbf{x},t) \in\mathbb{R}^{2} evaluated at the queried spatiotemporal locations. The masked MSE loss of Equation (13) is applied during training. The DeepONet architecture corresponding to Dataset 1 is illustrated in Figure 6(a).

Figure 6. DeepONet architecture (a) Dataset 1: The branch network encodes the initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0) concatenated with spatial coordinates via a CNN–MLP, producing pp latent coefficients {bk}\{b_k\}. The trunk network encodes the spatiotemporal query (x,t)∈R3(\mathbf{x},t) \in\mathbb{R}^3 via an MLP, producing pp basis functions {τk}\{\tau_k\}. Their dot product gives the two-channel vector phase field d(x,t)∈R2\mathbf{d}(\mathbf{x},t) \in\mathbb{R}^2 at the queried locations; (b) Dataset 2: The branch network encodes the concatenated initial scalar phase field d(x,0)d(\mathbf{x},0) and Young’s modulus field E(x)E(\mathbf{x}) via a ResNet-style CNN, producing pp latent coefficients {bk}\{b_k\}. The trunk network encodes purely spatial query coordinates x∈R2\mathbf{x} \in\mathbb{R}^2 via an MLP, producing pp basis functions {τk}\{\tau_k\}. Their dot product gives the single-channel scalar phase-field d(x,tf)d(\mathbf{x},t_f) at the final load step.

For Dataset 2, the temporal dimension is removed entirely. The branch network encodes the concatenated initial scalar phase-field d(x,0)d(\mathbf{x},0) and Young’s modulus field E(x)E(\mathbf{x}) as a two-channel input of shape (128×256×2)(128 \times256 \times2), using a ResNet-style CNN, swish activations, and average pooling to better capture the multi-scale spatial structure of the inclusion configurations. The trunk network takes only spatial coordinates (x)∈R2(\mathbf{x}) \in\mathbb{R}^{2} as input. Since our desired phase-field corresponds to a single time step (final), the temporal coordinates are not queried. The output is passed directly without a final nonlinear activation, and the loss is a weighted mean absolute error (MAE) that penalizes errors in the damaged region (d>0.5d > 0.5) more heavily to account for the class imbalance between intact and fractured material. The output is the single-channel scalar phase-field d(x,tf)d(\mathbf{x},t_f) at the final load step. The DeepONet architecture for Dataset 2 is illustrated in Figure 6(b).

Latent DeepONet

Latent DeepONet [34] extends the standard DeepONet framework by operating in a compressed latent space, reducing the dimensionality of the prediction target before learning the neural operator. A convolutional autoencoder is first trained to learn a low-dimensional latent representation of the high-dimensional output field. DeepONet is then trained to predict directly in this latent space, with the decoder subsequently mapping latent predictions back to the physical domain. The encoder and decoder are pretrained and the trainable parameters are frozen before DeepONet training begins.

For Dataset 1, let the two-channel vector phase-field solution be denoted by d(x,t):Ω×[0,T]→R2\mathbf{d}(\mathbf{x},t) : \Omega\times[0,T] \to\mathbb{R}^2, where x=(x1,x2)∈Ω⊂R2\mathbf{x} = (x_1,x_2) \in\Omega\subset\mathbb{R}^2. A convolutional autoencoder is first trained to learn an encoder E\mathcal{E} and decoder D\mathcal{D}, where the encoder maps the phase-field solution to a latent representation z(t)=E(d(⋅,t))∈R256\mathbf{z}(t) = \mathcal{E}(\mathbf{d}(\cdot,t)) \in\mathbb{R}^{256} and the decoder reconstructs the physical solution as d(⋅,t)≈D(z(t))\mathbf{d}(\cdot,t) \approx\mathcal{D}(\mathbf{z}(t)). The branch network receives as input the initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0) concatenated with the spatial coordinates (x1,x2)(x_1,x_2), forming the four-channel input b(x)=[u1(x,0),u2(x,0),x1,x2]\mathbf{b}(\mathbf{x}) = [u_1(\mathbf{x},0),u_2(\mathbf{x},0),x_1,x_2]. A CNN–MLP architecture with SiLU activations maps this input to a branch feature tensor B=B(b)∈Rr×256\mathbf{B} = \mathcal{B}(\mathbf{b}) \in\mathbb{R}^{r \times256}. The trunk network takes only the normalized temporal coordinate t∈Rt \in\mathbb{R} as input and produces the feature vector T(t)=T(t)∈Rr\mathbf{T}(t) = \mathcal{T}(t) \in\mathbb{R}^{r}, as the spatial dependence has already been encoded through the latent representation. The latent prediction is obtained through the DeepONet bilinear pairing z^(t)=∑k=1rbk,:τk(t)∈R256\hat{\mathbf{z}}(t)=\sum_{k=1}^{r}\mathbf{b}_{k,:}\tau_k(t)\in\mathbb{R}^{256}, which is implemented using an Einstein summation (einsum) operation over the shared feature dimension. The predicted latent state is subsequently decoded using the pre-trained decoder to obtain the phase-field prediction in the physical domain, d^(x,t)=D(z^(t))\hat{\mathbf{d}}(\mathbf{x},t)=\mathcal{D}(\hat{\mathbf{z}}(t)). The operator is trained by minimizing the MSE between the predicted and the encoded latent states, z^(t)\hat{\mathbf{z}}(t) and z(t)\mathbf{z}(t). The decoder is applied only afterwards, and the test error is evaluated in physical space with the masked MSE of Equation (13). The complete Dataset 1 Latent DeepONet architecture is illustrated in Figure 7(a). For an output field d(x,t)∈RDx×Dy\mathbf{d}(\mathbf{x},t)\in\mathbb{R}^{D_x\times D_y}, the three-stage mapping is:

Figure 7. Latent DeepONet architecture (a) Dataset 1: A pretrained convolutional autoencoder compresses the two-channel vector phase-field d(x,t)\mathbf{d}(\mathbf{x},t) into a latent sequence z(t)∈R256\mathbf{z}(t) \in\mathbb{R}^{256}. The branch network encodes u(x,0)\mathbf{u}(\mathbf{x},0) and spatial coordinates via a CNN–MLP; the trunk network encodes only the normalized time tt. Their combined output predicts z^(t)\hat{\mathbf{z}}(t), which is decoded to recover d^(x,t)\hat{\mathbf{d}}(\mathbf{x},t); (b) Dataset 2: A pretrained convolutional autoencoder compresses the scalar phase-field d(x,tf)d(\mathbf{x},t_f) into a latent vector z∈Rdz\mathbf{z} \in\mathbb{R}^{d_z}. The trunk network is bypassed entirely since there is no temporal query coordinate. The branch network encodes the concatenated inputs d(x,0)\mathbf{d}(\mathbf{x},0) and E(x)E(\mathbf{x}) via a ResNet-style CNN, producing a predicted latent vector z^\hat{\mathbf{z}} that is decoded to recover the scalar phase-field d^(x,tf)\hat{d}(\mathbf{x},t_f) at the final load step.

Eθenc:d(x,t)↦z(t)∈Rdz,Gθ:u(x,0)↦z(t)∈Rdz,Dθdec:z(t)↦d(x,t),(17)\begin{aligned} \mathcal{E}_{\theta_{\mathrm{enc}}}:\mathbf{d}(\mathbf{x},t)&\mapsto\mathbf{z}(t)\in\mathbb{R}^{d_z},\\ \mathcal{G}_{\theta}:\mathbf{u}(\mathbf{x},0)&\mapsto\mathbf{z}(t)\in\mathbb{R}^{d_z},\\ \mathcal{D}_{\theta_{\mathrm{dec}}}:\mathbf{z}(t)&\mapsto\mathbf{d}(\mathbf{x},t), \tag*{(17)} \end{aligned}

where Eθenc\mathcal{E}_{\theta_{\mathrm{enc}}} is the encoder, Gθ\mathcal{G}_{\theta} is the DeepONet operator trained in latent space, Dθdec\mathcal{D}_{\theta_{\mathrm{dec}}} is the decoder, z(t)\mathbf{z}(t) is the latent vector, and dzd_z is the latent dimension, where dz≪Dx×Dyd_z\ll D_x\times D_y.

For Dataset 2, the Latent DeepONet reduces to a purely spatial mapping with no temporal component, since the prediction target is a single spatial field at the final load step rather than a sequence. A convolutional autoencoder is trained to compress the scalar phase-field d(x,tf)∈R128×256\mathbf{d}(\mathbf{x},t_f)\in\mathbb{R}^{128\times256} into a low-dimensional latent vector z∈Rdz\mathbf{z}\in\mathbb{R}^{d_z}, where dz≪128×256d_z\ll128\times256. The trunk network, having no continuous temporal query coordinates in a static mapping, is bypassed entirely. The branch network alone maps the concatenated inputs (d(x,0),E(x))(\mathbf{d}(\mathbf{x},0),E(\mathbf{x})), encoded as a two-channel spatial tensor of shape (128×256×2)(128\times256\times2), directly to a predicted latent vector z^∈Rdz\hat{\mathbf{z}}\in\mathbb{R}^{d_z} via a ResNet-style CNN with residual shortcuts and swish activations. The decoder Dθdec\mathcal{D}_{\theta_{\mathrm{dec}}} subsequently maps z^\hat{\mathbf{z}} back to the physical domain, recovering the scalar phase-field d^(x,tf)\hat{\mathbf{d}}(\mathbf{x},t_f) at the final load step. The network is trained by minimizing the mean absolute error between the predicted and the encoded latent vectors, and the test error is evaluated on the decoded field with the MSE of Equation (14). The Dataset 2 Latent DeepONet architecture is illustrated in Figure 7(b).

UNet

UNet [37] is employed as a fully convolutional approximation of the nonlinear solution operator between discretized field variables. Let UℓU_\ell and DℓD_\ell denote the encoder and decoder operators at resolution level ℓ\ell, respectively. The network constructs a hierarchy of latent feature representations through successive restriction operations in the encoder and reconstructs the output through corresponding prolongation operations in the decoder. Skip connections introduce identity mappings between encoder and decoder feature spaces at matching resolutions, allowing the reconstructed solution to simultaneously incorporate coarse-scale contextual information and fine-scale spatial features.

Dataset 1. For Dataset 1, the objective is to learn the mapping G1:d(x,0)↦d(x,t)\mathcal{G}_1 : \mathbf{d}(\mathbf{x},0) \mapsto\mathbf{d}(\mathbf{x},t), for all discrete loading steps t∈{t1,…,tNt}t \in\{t_1,\ldots,t_{N_t}\}, where d(x,t)∈R2\mathbf{d}(\mathbf{x},t) \in\mathbb{R}^{2} denotes the vector phase-field. Rather than predicting each load step independently, the entire space–time solution trajectory is represented as a tensor D∈R2×Nt×128×128\mathbf{D} \in\mathbb{R}^{2 \times N_t \times128 \times128}, with channel, temporal, and spatial dimensions corresponding to the two phase-field components, the loading sequence, and the spatial discretization, respectively. A three-dimensional UNet [70] is therefore used to approximate the operator G1:R2×Nt×128×128→R2×Nt×128×128G_{1} : \mathbb{R}^{2 \times N_{t} \times128 \times128} \to\mathbb{R}^{2 \times N_{t} \times128 \times128}, treating the temporal coordinate as an additional dimension of the computational domain. The architecture consists of three encoder-decoder levels with convolutional kernels of size 5×5×55 \times5 \times5, feature widths [64,128,256,512][64, 128, 256, 512], group normalization with 32 groups, and GELU nonlinearities. The final layer applies a pointwise tanh activation so that the predicted phase-field remains bounded in [−1,1][-1, 1]. For the shortest loading sequence (Nt=11N_{t} = 11), the encoder depth is restricted to three levels because additional downsampling in the temporal direction would require excessive padding, introducing artificial boundary effects in the latent representation. Consequently, increasing NtN_{t} provides a larger effective temporal domain for hierarchical feature extraction, which contributes to the observed improvement in predictive accuracy. The network parameters are optimized by minimizing the masked mean-squared error defined in Equation (13). The complete Dataset 1 architecture is shown in Figure 8.

Figure 8. UNet architecture (a) Dataset 1: Three encoding levels with 5×5×55 \times5 \times5 convolutions, group normalization, and GELU activations process the joint spatiotemporal input. Skip connections concatenate encoder feature maps into the decoder at each resolution level to preserve spatial detail. Input and output shape: 2×Nt×128×1282 \times N_t \times128 \times128, where the two channels correspond to the components of the vector phase-field d(x,t)\mathbf{d}(\mathbf{x},t) and Nt∈{11,22,44}N_t \in\{11,22,44\}; (b) Dataset 2: Five encoding levels with 3×33 \times3 convolutions and channel widths [64,128,256,512,1024,2048][64,128,256,512,1024,2048] process the static spatial input. A sigmoid activation at the output layer constrains predictions to [0,1][0,1]. Input shape: 2×256×1282 \times256 \times128 (initial phase-field d(x,0)\mathbf{d}(\mathbf{x},0) and E(x)E(\mathbf{x})); output shape: 1×256×1281 \times256 \times128 (scalar phase-field d(x,tf)\mathbf{d}(\mathbf{x},t_f)).

Dataset 2. For Dataset 2, the task is to learn the static operator G2:(d(x,0),E(x))↦d(x,tf)G_{2} : (d(\mathbf{x}, 0), E(\mathbf{x})) \mapsto d(\mathbf{x}, t_{f}), where d(x,0)∈[0,1]d(\mathbf{x}, 0) \in[0, 1] is the initial scalar phase-field, E(x)E(\mathbf{x}) is the spatially varying Young’s modulus field, and d(x,tf)∈[0,1]d(\mathbf{x}, t_{f}) \in[0, 1] denotes the final damage state. Since no temporal dimension is present, the operator acts solely on spatial fields defined over the two-dimensional domain Ω\Omega.

The network input is represented by the tensor X∈R2×256×128\mathbf{X} \in\mathbb{R}^{2 \times256 \times128}, formed by concatenating d(x,0)d(\mathbf{x}, 0) and E(x)E(\mathbf{x}) along the channel dimension, while the output is Y^∈R1×256×128\hat{\mathbf{Y}} \in\mathbb{R}^{1 \times256 \times128}, corresponding to the predicted final phase-field. A two-dimensional UNet is used to approximate the mapping G2:R2×256×128→R1×256×128G_{2} : \mathbb{R}^{2 \times256 \times128} \to\mathbb{R}^{1 \times256 \times128}. The architecture employs five encoder-decoder levels with 3×33 \times3 convolutions and feature widths [64,128,256,512,1024,2048][64, 128, 256, 512, 1024, 2048]. A pointwise sigmoid activation is applied at the output layer to enforce the physically admissible range [0,1][0, 1] for the phase-field variable. Training is performed using the mean-squared error loss defined in Equation (14). The Dataset 2 UNet architecture is illustrated in Figure 8.

Conditional denoising diffusion probabilistic model

The cDDPM [40] frames the surrogate prediction task as a conditional generative problem. A forward Markov chain of TT steps gradually corrupts the ground truth output field with isotropic Gaussian noise:

q(xt∣xt−1)=N(xt;1−βt xt−1,βtI)(18)q(\mathbf{x}_t \mid\mathbf{x}_{t-1}) = \mathcal{N}\left(\mathbf{x}_t; \sqrt{1-\beta_t}\,\mathbf{x}_{t-1}, \beta_t I\right) \tag*{(18)}

where xt\mathbf{x}_t is the noised field at diffusion step tt, xt−1\mathbf{x}_{t-1} is the field at the previous step, N(⋅;μ,Σ)\mathcal{N}(\cdot;\boldsymbol{\mu},\boldsymbol{\Sigma}) denotes a Gaussian distribution with mean μ\boldsymbol{\mu} and covariance Σ\boldsymbol{\Sigma}, βt∈(0,1)\beta_t \in(0,1) is the variance (noise) schedule controlling the amount of noise injected at each of the TT forward steps, and II is the identity matrix. The model learns to reverse this process conditioned on the problem inputs, recovering the clean output field from pure noise. A cosine noise schedule is adopted in place of a linear one, maintaining a higher signal-to-noise ratio for longer during the forward process and enabling the network to learn semantically meaningful features during the early reverse steps. The training objective minimizes the ℓ1\ell_1 loss between the true injected noise ϵ\boldsymbol{\epsilon} and the network’s noise estimate ϵθ\epsilon_\theta:

LDDPM=Et,x0,ϵ[∥ϵ−ϵθ(xt,t,c)∥1](19)\mathcal{L}_{\mathrm{DDPM}} = \mathbb{E}_{t,\mathbf{x}_0,\epsilon}\left[\left\lVert\epsilon-\epsilon_\theta(\mathbf{x}_t,t,\mathbf{c})\right\rVert_1\right] \tag*{(19)}

where c\mathbf{c} denotes the conditioning input and x0\mathbf{x}_0 is the clean target field. The noise prediction network is a UNet backbone with ResNet blocks equipped with weight-standardized convolutions [71], which normalize convolutional kernel weights to zero mean and unit variance. Combined with group normalization (8 groups), this stabilizes training without requiring large batch sizes. To manage computational cost at high spatial resolution, a hybrid attention strategy is employed: linear attention with O(N)O(N) complexity in the high-resolution encoder and decoder paths, and full O(N2)O(N^2) self-attention at the bottleneck where spatial dimensions are smallest.

For Dataset 1, the model predicts the vector phase field d(x,tf)\mathbf{d}(\mathbf{x},t_f) at the final load step, conditioned on the initial displacement field u(x,0)\mathbf{u}(\mathbf{x},0). Conditioning is applied at two scales. Locally, a convolutional conditioning module aligns the displacement feature map with the noisy input and concatenates it at the first layer, providing pixel-wise geometric guidance. Globally, a separate deep encoder with 7×77 \times7 and 5×55 \times5 convolutional stages produces a latent context vector that is injected into each ResNet block via sinusoidal time embeddings and MLP projections. Channel multipliers (1,2,2,4)(1,2,2,4) progressively increase feature depth while reducing spatial resolution. The forward chain uses T=500T=500 noise steps. The network is trained with the noise-matching loss of Equation (19), and the masked MSE of Equation (13) is evaluated on the final denoised prediction. The Dataset 1 cDDPM architecture is illustrated in Figure 9.

Figure 9. Schematic overview of the proposed Conditional Denoising Diffusion Probabilistic Model (cDDPM) framework. The top row illustrates the Forward Process, where the ground truth phase-field data is gradually corrupted into Gaussian noise over discrete timesteps. The bottom row depicts the Reverse Process, where the model iteratively denoises the latent representation to reconstruct the phase-field sequence d(x,tf)∈R2\mathbf{d}(\mathbf{x},t_f) \in\mathbb{R}^2, conditioned on the initial displacement field ux(t=0)\mathbf{u}_x(t=0). The inset highlights the U-Net architecture used for noise prediction, featuring a symmetric encoder-decoder structure with ResNet blocks, weight standardization, and skip connections that facilitate multi-scale feature learning.

For Dataset 2, the cDDPM is applied in the same way, predicting the scalar phase-field d(x,tf)d(\mathbf{x},t_f) at the final load step conditioned on the concatenated inputs d(x,0)d(\mathbf{x},0) and E(x)E(\mathbf{x}). The forward chain corrupts the target d(x,tf)d(\mathbf{x},t_f) over T=500T=500 steps, and the reverse chain denoises conditioned on the initial configuration and material property field. The same UNet backbone, cosine noise schedule, weight-standardized convolutions, and ℓ1\ell_1 noise-matching loss of Equation (19) are used for training, and the MSE of Equation (14) is evaluated on the final denoised prediction.

Training and evaluation protocols

To make the architectures comparable, the dataset partitions, the input fields, the test metric, and the hardware are the same for all of them, and all experiments are conducted on a single NVIDIA A100 GPU. The optimizer settings and the training loss follow each architecture’s formulation. For both datasets, input variables are normalized using min–max scaling computed exclusively from the training set. Given a field vv, the normalized variable is defined as v~=2(v−vmin⁡)/(vmax⁡−vmin⁡)−1\tilde{v}=2(v-v_{\min})/(v_{\max}-v_{\min})-1, where vmin⁡v_{\min} and vmax⁡v_{\max} are determined from the training data and subsequently applied unchanged to the test set. For Dataset 1, the displacement components u1(x,0)u_1(\mathbf{x},0) and u2(x,0)u_2(\mathbf{x},0) are normalized independently. For Dataset 2, only the Young’s modulus field E(x)E(\mathbf{x}) is normalized, while the initial phase-field d(x,0)∈[0,1]d(\mathbf{x},0) \in[0,1] is used directly. The training losses are summarized in Table 2. For Dataset 1, FNO, DeepONet, and UNet are trained with the masked MSE of Equation (13), in which the hole region Ωhole\Omega_{\mathrm{hole}} is excluded. For Dataset 2, UNet is trained with the MSE of Equation (14), whereas FNO and DeepONet are trained with a weighted mean absolute error (MAE) that compensates for the severe imbalance between damaged and undamaged regions,

LwMAE=1NΩ∑x∈Ωw(x)∣dpred(x,tf)−dtrue(x,tf)∣,w(x)={6if dtrue(x,tf)>0.5,1otherwise.(20)\mathcal{L}_{\mathrm{wMAE}}=\frac{1}{N_{\Omega}}\sum_{\mathbf{x}\in\Omega}w(\mathbf{x})\left|d_{\mathrm{pred}}(\mathbf{x},t_f)-d_{\mathrm{true}}(\mathbf{x},t_f)\right|,\qquad w(\mathbf{x})=\begin{cases} 6 & \text{if }d_{\mathrm{true}}(\mathbf{x},t_f)>0.5,\\ 1 & \text{otherwise.} \end{cases} \tag*{(20)}

Two architectures cannot use these field-space losses by construction. Latent DeepONet is trained on the latent codes produced by its pretrained autoencoder, with an MSE for Dataset 1 and an MAE for Dataset 2, and the cDDPM is trained with the noise-matching loss of Equation (19). Because the training losses differ, every comparison in § 4 uses a single test metric per dataset, evaluated on the predicted field in physical space: the masked MSE of Equation (13) for Dataset 1 and the MSE of Equation (14) for Dataset 2.

For Dataset 1, training is performed in two stages. First, each model is optimized until the test error satisfies MSE <6×10−5< 6 \times10^{-5}; the corresponding wall-clock time is recorded as a measure of training efficiency. Training then continues from the same checkpoint until either convergence is observed or a maximum training budget of 24 hours is reached, yielding the minimum attainable test error for the architecture. For Dataset 2, each model is trained for a fixed number of epochs, long enough for the test error to plateau (Figure 14), and the checkpoint with the lowest test error is retained; no separate validation set is used. Performance is quantified using four metrics: (i) prediction accuracy, measured by the masked MSE (Equation (13)) for Dataset 1 and the MSE (Equation (14)) for Dataset 2 on the held-out test set; (ii) training efficiency, measured by the time required to attain the threshold MSE <6×10−5< 6 \times10^{-5} for Dataset 1 and the total time to convergence for Dataset 2; (iii) inference cost, measured as the average wall-clock time per test sample following a warm-up pass; and (iv) model complexity, measured by the total number of trainable parameters. To assess temporal super-resolution on Dataset 1, an additional interpolation experiment is conducted. Each surrogate is trained using Case 2, which contains Nt=22N_t = 22 snapshots obtained by retaining every second state of the full 44-step trajectory, and is subsequently evaluated on the complete 44-step sequence. Let Tobs\mathcal{T}_{\mathrm{obs}} denote the set of training time steps and Tint\mathcal{T}_{\mathrm{int}} the complementary set of withheld intermediate steps. Prediction errors are reported over Tobs\mathcal{T}_{\mathrm{obs}} and over the full sequence Tobs∪Tint\mathcal{T}_{\mathrm{obs}} \cup\mathcal{T}_{\mathrm{int}}, so that the increase between the two measures captures the cost of interpolating to unseen temporal locations.

Figure 14. Training (blue) and test (orange) MSE versus training epoch for each surrogate architecture on Dataset 2 (heterogeneous plate with circular inclusions). All models are trained jointly on all four inclusion configurations. UNet converges fastest in approximately 400 epochs; FNO converges within 2,000 epochs; DeepONet and Latent DeepONet require up to 20,000 epochs; and cDDPM, while achieving the lowest final test MSE, requires the longest training time of 23 hours before loss plateaus.

Using and extending Fracture-Bench

The benchmark tasks of § 3.2 are one way of using the datasets, chosen so that all five architectures could be compared on equal terms. They are not the only way.

Other tasks on the same data. Both datasets store the displacement and phase-field solutions at every load step, so they support tasks beyond those benchmarked here. Examples include predicting the full 31-step crack evolution in Dataset 2 rather than only its final state, predicting the displacement field or both fields jointly, step-by-step (autoregressive) prediction of crack growth, and extrapolation tests in which one inclusion configuration is withheld from training. Because the governing equations are known, the data can also be combined with physics-informed training objectives. The loss function, the preprocessing, and the training budget are the user’s choice.

Adding an architecture.. To make a new surrogate comparable with the reference results of § 4, it should be trained on the training split and evaluated on the test split of Table 1, and the following should be reported: the test error in the metric of Equation (13) for Dataset 1 or Equation (14) for Dataset 2, the training wall-clock time and hardware, the inference time per sample, and the number of trainable parameters.

Adding a fracture problem.. A new dataset can be added to the benchmark if it provides (i) the solution fields on a uniform grid in the same array layout as the existing datasets (samples ×\times load steps ×\times grid points ×\times field components), (ii) the simulation code or input files needed to regenerate it, (iii) a fixed train–test split, and (iv) a statement of the learning task, namely its inputs, its outputs, and the metric used to score it. A baseline result from at least one existing architecture is encouraged. Contributions can be submitted as an issue or pull request in the project repository, https://github.com/Centrum-IntelliPhysics/fracture_bench, or by contacting the authors.

Interactive dashboard.. A companion website, https://fracture-bench.github.io, lets users explore both datasets and the trained surrogates without writing code. For each dataset, the user selects one of ten held-out test simulations and compares the predictions of FNO, DeepONet and Latent DeepONet with the finite element solution and with pixel-wise error maps. For Dataset 1, the predicted crack evolution can be stepped through at 11, 22 or 44 load steps.

Results and discussion

This section presents a systematic evaluation of the five surrogate architectures - FNO, DeepONet, Latent DeepONet, UNet, and diffusion models - on both datasets. For each dataset, results are organized across four studies: prediction accuracy, computational efficiency, temporal generalization, and qualitative assessment of predicted crack fields. A final subsection compares neural operator predictions against linear temporal interpolation as a non-learning baseline.

Dataset 1: Square plate with a central hole

The aggregate performance metrics are summarized in Table 3. Across all temporal resolutions, FNO consistently achieves the best combination of accuracy and training efficiency, reaching the prescribed error threshold substantially faster than the remaining architectures while also attaining the lowest final test error. This behavior is consistent with the spectral parameterization of the operator kernel, which efficiently captures the global spatial structure governing crack evolution.

MetricStepsFNODeepONetLatent DONUNetcDDPM
Training time (hr)110.1531.56.5
220.22032.1–
440.3–43.5–
Inference time (s)114×10−14 \times10^{-1}1×10−31 \times10^{-3}6×10−36 \times10^{-3}17.7×10−317.7 \times10^{-3}13.42
224×10−14 \times10^{-1}1×10−31 \times10^{-3}6×10−36 \times10^{-3}34.8×10−334.8 \times10^{-3}–
444.5×10−14.5 \times10^{-1}–10×10−310 \times10^{-3}69.8×10−369.8 \times10^{-3}–
Parameters (×106\times10^{6})113143489822.4
2231456198–
44314–8798–
Min. test MSE (×10−6\times10^{-6})111290.920–
2215039–
442–710–

Table 3. Consolidated computational profile across all temporal resolutions for Dataset 1. Training time is measured to reach a test MSE of 6×10−56 \times10^{-5}. Minimum test MSE is the lowest value achieved after extended training under the 24-hour wall-clock budget, averaged across all test samples and time steps.

Latent DeepONet provides the strongest performance among the remaining architectures. Relative to DeepONet, the latent formulation converges considerably faster and achieves substantially lower error, indicating that learning the temporal operator in a compressed latent manifold is more effective than operating directly in the high-dimensional physical space. Although its final accuracy remains slightly below that of FNO, the gap is small compared to the substantial reduction in model complexity.

UNet exhibits stable convergence across all temporal resolutions and achieves competitive accuracy. Its performance improves as the number of temporal snapshots increases, which is attributed to the three-dimensional encoder–decoder architecture. For short sequences, repeated downsampling in the temporal direction introduces padding artifacts that become progressively less influential as the temporal extent grows.

In contrast, DeepONet scales poorly with temporal resolution. Both optimization difficulty and final error increase markedly with sequence length, and the model fails to reach the target accuracy for the most demanding case. The cDDPM exhibits a similar trend. While successful for the coarsest temporal resolution, its performance deteriorates as the temporal dimension increases, suggesting that conditioning solely on the final crack state provides an insufficient training signal for learning the underlying spatio-temporal evolution.

The computational trade-offs are also evident from Table 3. FNO offers the lowest training cost but incurs relatively expensive inference due to repeated Fourier transforms. DeepONet occupies the opposite extreme, requiring the longest training time while providing the fastest inference. Latent DeepONet and UNet occupy an intermediate regime, balancing training efficiency, prediction accuracy, and inference cost. The cDDPM is the most expensive architecture at inference because each prediction requires execution of the full reverse diffusion chain.

The convergence histories shown in Figure 10 are consistent with the aggregate statistics of Table 3. FNO and UNet exhibit rapid and stable optimization, whereas DeepONet requires substantially more training iterations to reach comparable error levels. The cDDPM converges only for the coarsest temporal discretization and plateaus at higher resolutions. Collectively, these results indicate that architectures designed to exploit spatial structure directly through convolutional or spectral representations are considerably easier to optimize than operator-learning approaches acting in the full physical space.

Figure 10. Training (blue) and test (orange) MSE versus training epoch for each surrogate architecture and each temporal resolution (Nt∈{11,22,44}N_t \in\{11, 22, 44\} load steps) in Dataset 1. Each row corresponds to one architecture and each column to one temporal case. FNO and UNet converge within hundreds to a few thousand epochs; DeepONet requires up to 10510^5 epochs to approach the same MSE level and does not converge at Nt=44N_t = 44 steps within the 24-hour budget (row left blank); cDDPM converges at 11 steps but plateaus well above the threshold at 22, and 44 steps. The Latent DeepONet curves are evaluated in the encoded latent space and therefore cannot be directly compared with the other architectures on the same vertical axis, which are evaluated in physical space.

The temporal error evolution for the 11-step case is shown in Figure 11. All architectures exhibit increasing error with load step, reflecting the transition from smooth damage fields to highly localized crack patterns. Differences among models are modest during the early stages of crack growth, but become increasingly pronounced as localization develops. FNO and Latent DeepONet remain the most accurate throughout the loading history, whereas DeepONet, UNet, and particularly the cDDPM experience a more rapid degradation near the final stages of fracture. This behavior suggests that accurately resolving the strongly nonlinear crack-tip dynamics, rather than the early diffuse damage evolution, is the primary challenge separating the surrogate architectures.

Figure 11. Test MSE per time step for the 11-step case of Dataset 1. The error increases monotonically with time step for all architectures as the phase-field evolves toward sharp localized crack bands. Latent DeepONet achieves the lowest error at early steps; FNO and Latent DeepONet converge in error beyond step 8; cDDPM degrades the fastest beyond step 8.

Temporal generalization is examined in Figure 12, which reports the per-sample test error of models trained on the 22-step sequence (Case 2) when they are evaluated on those 22 steps and on the full 44-step sequence, half of which was withheld during training. For both FNO and Latent DeepONet the mean error over the full 44-step sequence is roughly twice the error over the 22 training steps, from 6.1×10−56.1 \times10^{-5} to 1.3×10−41.3 \times10^{-4} for FNO and from 1.1×10−61.1 \times10^{-6} to 2.1×10−62.1 \times10^{-6} for Latent DeepONet, and the spread across test samples remains similar. Both models therefore interpolate in time without a breakdown in accuracy, with Latent DeepONet retaining the lower error.

Figure 12. Per-sample test MSE for models trained on 22 load steps and evaluated at both 22 steps (in-distribution) and 44 steps (out-of-distribution interpolation). Each point corresponds to one test sample. The mean MSE (dashed red line) increases for all architectures when evaluated at unseen intermediate steps, with Latent DeepONet showing the smallest relative degradation, from 1.1×10−61.1 \times10^{-6} (22 steps) to 2.1×10−62.1 \times10^{-6} (44 steps), and FNO showing a larger increase from 6.1×10−56.1 \times10^{-5} to 1.3×10−41.3 \times10^{-4}.

Qualitative predictions for a representative test sample are shown in Figure 13, which compares the predicted phase-field at the last load step of each case with the finite element reference. Every architecture that reached the error threshold recovers the correct set of cracks growing from the hole. The differences are confined to the crack bands: the errors of FNO and Latent DeepONet are barely visible at all three temporal resolutions, UNet produces slightly wider crack bands at Nt=11N_t = 11, and DeepONet and the cDDPM show the largest errors, concentrated along the cracks. This is consistent with Figure 11, where the error of every model grows in the last load steps as the cracks localize.

Figure 13. Predicted phase-field magnitude ∥d(x,tf)∥\lVert\mathbf{d}(\mathbf{x},t_f)\rVert and absolute pointwise error at the final load step for representative test sample 1 of Dataset 1, across all five surrogate architectures and three temporal resolutions (Nt∈{11,22,44}N_t \in\{11,22,44\} load steps). Top row: GRF boundary profile f(x)f(x) and reference displacement magnitude and phase-field at the initial load step t=1t=1. Left block of each pair: Predicted phase field at the final load step tft_f. Right block of each pair: Absolute pointwise error ∣dpred−dtrue∣\lvert\mathbf{d}_{\mathrm{pred}}-\mathbf{d}_{\mathrm{true}}\rvert between prediction and ground truth. Black bars indicate cases where the model did not converge to the threshold MSE of 6×10−56 \times10^{-5} within the 24-hour training budget. FNO and Latent DeepONet reproduce crack paths accurately at all temporal resolutions; UNet produces wider crack bands at Nt=11N_t=11 due to spatial padding in the 3D encoder; cDDPM converges only at Nt=11N_t=11, and DeepONet at Nt=11N_t=11 and 22.

Dataset 2: Heterogeneous plate with two circular inclusions

The aggregate performance metrics for Dataset 2 are summarized in Table 4, while the convergence histories are shown in Figure 14. In contrast to Dataset 1, the ranking of the architectures changes substantially. The cDDPM achieves the lowest test error, followed by UNet and FNO, whereas DeepONet remains the least accurate model. This reversal highlights an important distinction between the two benchmarks. Dataset 1 requires learning a spatio-temporal evolution operator, whereas Dataset 2 involves only a single spatial prediction. Without the burden of modeling temporal dynamics, the diffusion model can devote its representational capacity to resolving the highly localized crack band and its interaction with material interfaces, which appears particularly well suited to iterative denoising.

MetricFNODeepONetLatent DONUNetcDDPM
Training time (hr)2.4818.6016.70.8723.02
Inference time (s)1.6×10−11.6 \times10^{-1}6.2×10−46.2 \times10^{-4}4.9×10−44.9 \times10^{-4}5.9×10−35.9 \times10^{-3}7.06×1027.06 \times10^{2}
Parameters (×106\times10^{6})37.806.1072.43124.4031.79
Mean test MSE (×10−4\times10^{-4})3.8517.29.57, 7.82, 3.58, 4.662.930.45

Table 4. Computational profile for Dataset 2. All models are trained jointly on all four inclusion configurations in a single run. Training time is the total wall-clock time to convergence on a single NVIDIA A100 GPU. Inference time is per sample, averaged over the 400-sample test set. Mean test MSE is averaged across all test samples on the full 128×256128 \times256 domain. Note: For Latent DeepONet, per-case MSE values are reported as (less stiff-stiffer, stiffer-stiffer, less stiff-less stiff, stiffer-less stiff).

UNet benefits from the same simplification. Unlike the three-dimensional architecture used in Dataset 1, the two-dimensional UNet operates on a spatial grid whose dimensions are compatible with the encoder–decoder hierarchy, eliminating the padding artifacts previously observed at coarse temporal resolutions. FNO remains competitive but is consistently less accurate than UNet and the cDDPM, suggesting that the global spectral representation is less effective at resolving localized crack-tip behavior and sharp inclusion-boundary interactions. DeepONet again exhibits the weakest performance, indicating that compressing the input fields into a fixed-dimensional latent representation is insufficient for preserving the fine spatial information required to accurately predict crack deflection around inclusions. Latent DeepONet alleviates this limitation but does not fully eliminate it.

The convergence curves in Figure 14 reinforce these observations. UNet converges most rapidly, while the cDDPM requires substantially longer optimization but ultimately reaches the lowest error. The operator-learning models converge more slowly and plateau at higher error levels, indicating that the dominant challenge in Dataset 2 is spatial localization rather than global field reconstruction.

The computational trade-offs differ from those observed in Dataset 1. UNet provides the most favorable balance between accuracy and computational cost, achieving near-best accuracy with the shortest training time and modest inference cost. FNO offers similar efficiency but at slightly lower accuracy. Although the cDDPM delivers the most accurate predictions, its iterative reverse-diffusion process makes it several orders of magnitude more expensive at inference, limiting its practicality in applications requiring large numbers of forward evaluations. DeepONet and Latent DeepONet retain extremely low inference costs but sacrifice predictive accuracy.

Qualitative predictions for representative test cases are shown in Figure 15. All architectures correctly reproduce the primary physical mechanism: cracks tend to deflect around stiffer inclusions and propagate through less stiff inclusions. The principal differences lie in the resolution of the crack-tip region and the inclusion-transition zone. The cDDPM and UNet generate the sharpest crack representations, with errors largely confined to the crack band itself. FNO produces comparable crack trajectories but exhibits slightly larger localization errors near material interfaces. DeepONet generates noticeably more diffuse damage fields, particularly in configurations exhibiting strong crack deflection, consistent with the limitations of its fixed-dimensional latent representation. Latent DeepONet partially mitigates this issue and generally captures the correct crack trajectory, although its accuracy degrades as the stiffness contrast increases and the crack path becomes more geometrically complex.

Figure 15. Ground truth phase-field d(x,tf)d(\mathbf{x},t_f) and absolute prediction errors from FNO, DeepONet, Latent DeepONet, UNet, and cDDPM for five Dataset 2 test samples sharing an identical initial crack location but differing in inclusion configuration (less stiff, stiffer), (stiffer, stiffer), (less stiff, less stiff), and (stiffer, less stiff). Top row: Initial crack condition and ground truth final phase field. Subsequent rows: Predicted phase field (left of each pair) and absolute error (right) for each architecture. The variation in crack path across the four configurations demonstrates the sensitivity of fracture trajectories to inclusion stiffness, and the variation in error patterns across architectures reveals the differing capabilities of each model in resolving the crack-tip region and the inclusion-boundary transition zone.

Overall, Dataset 2 highlights a different regime from Dataset 1. When the task is dominated by resolving a single highly localized spatial field, convolutional generative models outperform operator-learning approaches. The cDDPM provides the highest accuracy, UNet offers the best balance between accuracy and computational efficiency, and DeepONet-based models remain limited by the information bottleneck imposed by their latent representations.

Conclusions

We introduced Fracture-Bench, an open benchmark for machine-learning surrogates of phase-field fracture. Its main contribution is the resource itself: two datasets of 1,000 and 4,000 simulations that differ in fracture formulation, material heterogeneity and prediction task, released with the codes that generated them, fixed splits, common metrics, regime descriptors, training code and an interactive dashboard. The comparison of five architectures, FNO, DeepONet, Latent DeepONet, UNet and the cDDPM, is a first use of this resource. It provides reference results against which new methods can be measured, and it shows why a shared benchmark is needed.

The main findings are as follows.

  • No architecture is best on both datasets. On the trajectory task of Dataset 1, FNO and Latent DeepONet are the most accurate (test MSE of order 10−610^{-6}), and FNO trains 10 to 100 times faster than the others. On the final-state task of Dataset 2, the cDDPM is the most accurate (0.45×10−40.45 \times10^{-4}), and UNet offers the best balance of accuracy (second-best error), training time (0.870.87 h) and inference time (66 ms).

  • Accuracy and cost trade off strongly. The most accurate model on Dataset 2 is also the slowest by five orders of magnitude at inference, and the fastest models at inference (DeepONet and Latent DeepONet) are among the least accurate.

  • The error of every architecture is concentrated where and when the crack localizes, and on Dataset 2 near crack-path bifurcations, where inputs less than one grid cell apart lead to different crack paths. Domain-averaged errors understate these failures, and for near-bifurcation inputs a probabilistic prediction is more appropriate than a single field.

  • The random splits test interpolation in low-dimensional, densely sampled input spaces. Generalization to new loadings, configurations and crack locations must be measured separately, with dedicated splits, for example by withholding one inclusion configuration from training (§ 3.5).

The tasks studied here use only part of what the datasets contain. The stored displacement fields and intermediate load steps support other learning problems, and users are free to choose their own loss functions and training strategies. Fracture-Bench is also meant to grow. § 3.5 sets out what is needed to add a new architecture or a new fracture problem, including a regime descriptor that records what each dataset covers and what it lacks. We particularly invite datasets from the regimes that open data cover least: mixed-mode and shear loading, three-dimensional crack growth, dynamic, cyclic and ductile fracture, and simulations paired with experiments. Together, such contributions would show where surrogate predictions of fracture can be trusted and where more data are needed. Dataset 1 is available at https://doi.org/10.7281/T1XFF190 [45] and Dataset 2 at https://doi.org/10.7281/T1RZFI3I [66]. The training and evaluation code is available at https://github.com/Centrum-IntelliPhysics/fracture_bench, and the interactive dashboard at https://fracture-bench.github.io.

Data and code availability

Both datasets are openly available from the Johns Hopkins Research Data Repository: Dataset 1 at https://doi.org/10.7281/T1XFF190 [45] and Dataset 2 at https://doi.org/10.7281/T1RZFI3I [66]. The training and evaluation code is available at https://github.com/Centrum-IntelliPhysics/fracture_bench. An interactive dashboard that shows the trained surrogates on held-out test cases is available at https://fracture-bench.github.io.

References

  1. [1]R. P. Reed, J. H. Smith, B. W. Christ, The Economic Effects of Fracture in the United States. Part 1: A Synopsis of the September 30, 1982 Report to NBS by Battelle Columbus Laboratories, NBS Special Publication 647-1, National Bureau of Standards, Washington, DC, 1983. doi:10.6028/NBS.SP.647p1, 19 pp.
  2. [2]National Institute of Standards and Technology, Qualification for additive manufacturing materials, processes, and parts, NIST program/project web page, 2014. URL: https://www.nist.gov/programs-projects/qualification-additive-manufacturing-materials-processes-and-parts, created 27 March 2014; updated 26 March 2025; accessed 13 October 2026.
  3. [3]X. Liu, D. Furrer, J. Kosters, J. Holmes, Vision 2040: A Roadmap for Integrated, Multiscale Modeling and Simulation of Materials and Systems, NASA Contractor Report NASA/CR-2018-219771, NASA Glenn Research Center, Cleveland, OH, 2018. URL: https://ntrs.nasa.gov/citations/20180002010, nTRS Document ID 20180002010; Contract NNC15BA06B.
  4. [4]A. A. Griffith, VI. The phenomena of rupture and flow in solids, Philosophical Transactions of the Royal Society of London. Series A 221 (1921) 163–198. doi:10.1098/rsta.1921.0006.DOI
  5. [5]G. R. Irwin, Fracture, in: S. Flügge (Ed.), Elasticity and Plasticity / Elastizität und Plastizität, Handbuch der Physik / Encyclopedia of Physics, Springer, Berlin, Heidelberg, 1958, pp. 551–590. doi:10.1007/978-3-642-45887-3_5.DOI
  6. [6]U. M. Krishnan, A. Gupta, A. Kumar, R. Chowdhury, Adaptive PF-CZM for multiphysics fracture analysis in functionally graded materials, Engineering Fracture Mechanics 310 (2024) 110461. doi:10.1016/j.engfracmech.2024.110461.DOI
  7. [7]N. Moës, J. Dolbow, T. Belytschko, A finite element method for crack growth without remeshing, International Journal for Numerical Methods in Engineering 46 (1999) 131–150. doi:10.1002/(SICI)1097-0207(19990910)46:1<131::AID-NME726>3.0.CO;2-J.doi.org/10.1002/(sici)1097-0207(19990910)46:1<131::aid-nme726>3.0.co;2-j
  8. [8]A. R. Khoei, Extended Finite Element Method: Theory and Applications, John Wiley & Sons, 2014. doi:10.1002/9781118869673.DOI
  9. [9]P. Laborde, J. Pommier, Y. Renard, M. Salaün, High-order extended finite element method for cracked domains, International Journal for Numerical Methods in Engineering 64 (2005) 354–381. doi:10.1002/nme.1370.
  10. [10]O. A. González-Estrada, J. J. Ródenas, S. P. A. Bordas, M. Duflot, P. Kerfriden, E. Giner, On the role of enrichment and statical admissibility of recovered fields in a posteriori error estimation for enriched finite element methods, Engineering Computations 29 (2012) 814–841. doi:10.1108/02644401211271609.
  11. [11]J.-Y. Wu, J.-F. Qiu, V. P. Nguyen, T. K. Mandal, L.-J. Zhuang, Computational modeling of localized failure in solids: XFEM vs PF-CZM, Computer Methods in Applied Mechanics and Engineering 345 (2019) 618–643. doi:10.1016/j.cma.2018.10.044.DOI
  12. [12]Y. Jin, O. González-Estrada, O. Pierard, S. Bordas, Error-controlled adaptive extended finite element method for 3D linear elastic crack propagation, Computer Methods in Applied Mechanics and Engineering 318 (2017) 319–348. doi:10.1016/j.cma.2016.12.016.
  13. [13]G. A. Francfort, J.-J. Marigo, Revisiting brittle fracture as an energy minimization problem, Journal of the Mechanics and Physics of Solids 46 (1998) 1319–1342. doi:10.1016/S0022-5096(98)00034-9.DOI
  14. [14]B. Bourdin, G. A. Francfort, J.-J. Marigo, Numerical experiments in revisited brittle fracture, Journal of the Mechanics and Physics of Solids 48 (2000) 797–826. doi:10.1016/S0022-5096(99)00028-9.DOI
  15. [15]H. Amor, J.-J. Marigo, C. Maurini, Regularized formulation of the variational brittle fracture with unilateral contact: Numerical experiments, Journal of the Mechanics and Physics of Solids 57 (2009) 1209–1229. doi:10.1016/j.jmps.2009.04.011.DOI
  16. [16]C. Miehe, F. Welschinger, M. Hofacker, Thermodynamically consistent phase-field models of fracture: Variational principles and multi-field FE implementations, International Journal for Numerical Methods in Engineering 83 (2010) 1273–1311. doi:10.1002/nme.2861.DOI
  17. [17]M. Ambati, T. Gerasimov, L. De Lorenzis, A review on phase-field models of brittle fracture and a new fast hybrid formulation, Computational Mechanics 55 (2015) 383–405. doi:10.1007/s00466-014-1109-y.DOI
  18. [18]J.-Y. Wu, A unified phase-field theory for the mechanics of damage and quasi-brittle failure, Journal of the Mechanics and Physics of Solids 103 (2017) 72–99. doi:10.1016/j.jmps.2017.03.015.DOI
  19. [19]R. J. M. Geelen, Y. Liu, T. Hu, M. R. Tupek, J. E. Dolbow, A phase-field formulation for dynamic cohesive fracture, Computer Methods in Applied Mechanics and Engineering 348 (2019) 680–711. doi:10.1016/j.cma.2019.01.026.DOI
  20. [20]A. Gupta, U. M. Krishnan, R. Chowdhury, A. Chakrabarti, An auto-adaptive sub-stepping algorithm for phase-field modeling of brittle fracture, Theoretical and Applied Fracture Mechanics 108 (2020) 102622. doi:10.1016/j.tafmec.2020.102622.DOI
  21. [21]R. Bharali, S. Goswami, C. Anitescu, T. Rabczuk, A robust monolithic solver for phase-field fracture integrated with fracture energy based arc-length method and under-relaxation, Computer Methods in Applied Mechanics and Engineering 394 (2022) 114927. doi:10.1016/j.cma.2022.114927.DOI
  22. [22]T. Heister, M. F. Wheeler, T. Wick, A primal-dual active set method and predictor-corrector mesh adaptivity for computing fracture propagation using a phase-field approach, Computer Methods in Applied Mechanics and Engineering 290 (2015) 466–495. doi:10.1016/j.cma.2015.03.009.DOI
  23. [23]P. Areias, T. Rabczuk, M. A. Msekh, Phase-field analysis of finite-strain plates and shells including element subdivision, Computer Methods in Applied Mechanics and Engineering 312 (2016) 322–350. doi:10.1016/j.cma.2016.01.020.
  24. [24]S. Goswami, C. Anitescu, T. Rabczuk, Adaptive phase field analysis with dual hierarchical meshes for brittle fracture, Engineering Fracture Mechanics 218 (2019) 106608. doi:10.1016/j.engfracmech.2019.106608.DOI
  25. [25]S. Goswami, C. Anitescu, T. Rabczuk, Adaptive fourth-order phase field analysis for brittle fracture, Computer Methods in Applied Mechanics and Engineering 361 (2020) 112808. doi:10.1016/j.cma.2019.112808.DOI
  26. [26]H. Hirshikesh, A. Pramod, H. Waisman, S. Natarajan, Adaptive phase field method using novel physics based refinement criteria, Computer Methods in Applied Mechanics and Engineering 383 (2021) 113874. doi:10.1016/j.cma.2021.113874.DOI
  27. [27]U. M. Krishnan, A. Gupta, R. Chowdhury, Adaptive phase-field modeling of brittle fracture using a robust combination of error-estimator and markers, Engineering Fracture Mechanics 274 (2022) 108758. doi:10.1016/j.engfracmech.2022.108758.DOI
  28. [28]A. Gupta, U. M. Krishnan, T. K. Mandal, R. Chowdhury, V. P. Nguyen, An adaptive mesh refinement algorithm for phase-field fracture models: Application to brittle, cohesive, and dynamic fracture, Computer Methods in Applied Mechanics and Engineering 399 (2022) 115347. doi:10.1016/j.cma.2022.115347.DOI
  29. [29]P. S. Theocaris, P. D. Panagiotopoulos, Neural networks for computing in fracture mechanics. Methods and prospects of applications, Computer Methods in Applied Mechanics and Engineering 106 (1993) 213–228. doi:10.1016/0045-7825(93)90191-Y.DOI
  30. [30]M. K. Mudunuru, N. Panda, S. Karra, G. Srinivasan, V. T. Chau, E. Rougier, A. Hunter, H. S. Viswanathan, Surrogate models for estimating failure in brittle and quasi-brittle materials, Applied Sciences 9 (2019) 2706. doi:10.3390/app9132706.DOI
  31. [31]S. Goswami, C. Anitescu, S. Chakraborty, T. Rabczuk, Transfer learning enhanced physics informed neural network for phase-field modeling of fracture, Theoretical and Applied Fracture Mechanics 106 (2020) 102447. doi:10.1016/j.tafmec.2019.102447.DOI
  32. [32]L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nature Machine Intelligence 3 (2021) 218–229. doi:10.1038/s42256-021-00302-5.DOI
  33. [33]Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations, in: International Conference on Learning Representations (ICLR), 2021. ArXiv:2010.08895.arxiv.org/abs/2010.08895
  34. [34]K. Kontolati, S. Goswami, G. E. Karniadakis, M. D. Shields, Learning nonlinear operators in latent spaces for real-time predictions of complex dynamics in physical systems, Nature Communications 15 (2024) 5101. doi:10.1038/s41467-024-49411-w.DOI
  35. [35]S. Goswami, A. Bora, Y. Yu, G. E. Karniadakis, Physics-informed deep neural operator networks, in: T. Rabczuk, K.-J. Bathe (Eds.), Machine Learning in Modeling and Simulation: Methods and Applications, Computational Methods in Engineering & the Sciences, Springer, Cham, 2023, pp. 219–254. doi:10.1007/978-3-031-36644-4_6.DOI
  36. [36]S. Goswami, M. Yin, Y. Yu, G. E. Karniadakis, A physics-informed variational DeepONet for predicting crack path in quasi-brittle materials, Computer Methods in Applied Mechanics and Engineering 391 (2022) 114587. doi:10.1016/j.cma.2022.114587.DOI
  37. [37]O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional networks for biomedical image segmentation, in: N. Navab, J. Hornegger, W. M. Wells, A. F. Frangi (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, Springer, Cham, 2015, pp. 234–241. doi:10.1007/978-3-319-24574-4_28.DOI
  38. [38]I. Saha, A. Gupta, L. Graham-Brady, Prediction of local elasto-plastic stress and strain fields in a two-phase composite microstructure using a deep convolutional neural network, Computer Methods in Applied Mechanics and Engineering 421 (2024) 116816. doi:10.1016/j.cma.2024.116816.DOI
  39. [39]I. Saha, L. Graham-Brady, Numerical and data-driven modeling of spall failure in polycrystalline ductile materials, Computer Methods in Applied Mechanics and Engineering 448 (2026) 118493. doi:10.1016/j.cma.2025.118493.DOI
  40. [40]J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, in: Advances in Neural Information Processing Systems, volume 33, 2020, pp. 6840–6851.arxiv.org/abs/2006.11239
  41. [41]M. M. Rashid, T. Pittie, S. Chakraborty, N. M. A. Krishnan, Learning the stress-strain fields in digital composites using Fourier neural operator, iScience 25 (2022) 105452. doi:10.1016/j.isci.2022.105452.DOI
  42. [42]M. M. Rashid, S. Chakraborty, N. M. A. Krishnan, Revealing the predictive power of neural operators for strain evolution in digital composites, Journal of the Mechanics and Physics of Solids 181 (2023) 105444. doi:10.1016/j.jmps.2023.105444.DOI
  43. [43]C. Bonneville, N. Bie­berdorf, A. Hegde, M. Asta, H. N. Najm, L. Capolungo, C. Safta, Accelerating phase field simulations through a hybrid adaptive Fourier neural operator with U-Net backbone, npj Computational Materials 11 (2025) 14. doi:10.1038/s41524-024-01488-z.DOI
  44. [44]E. Hamdi, E. Lejeune, Towards robust surrogate models: Benchmarking machine learning approaches to expediting phase field simulations of brittle fracture, Computer Methods in Applied Mechanics and Engineering 449 (2026) 118526. doi:10.1016/j.cma.2025.118526.DOI
  45. [45]M. Hakimzadeh, L. Graham-Brady, S. Goswami, Data and code associated with: Phase-field fracture simulation dataset: Hyperelastic multi-crack response under loading and unloading, Johns Hopkins Research Data Repository, 2025. URL: https://doi.org/10.7281/T1XFF190. doi:10.7281/T1XFF190, dataset.
  46. [46]J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Žídek, A. Potapenko, et al., Highly accurate protein structure prediction with AlphaFold, Nature 596 (2021) 583–589. doi:10.1038/s41586-021-03819-2.DOI
  47. [47]H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, P. E. Bourne, The Protein Data Bank, Nucleic Acids Research 28 (2000) 235–242. doi:10.1093/nar/28.1.235.DOI
  48. [48]H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Horányi, J. Muñoz-Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers, et al., The ERA5 global reanalysis, Quarterly Journal of the Royal Meteorological Society 146 (2020) 1999–2049. doi:10.1002/qj.3803.DOI
  49. [49]R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirnsberger, M. Fortunato, F. Alet, S. Ravuri, T. Ewalds, Z. Eaton-Rosen, W. Hu, A. Merose, S. Hoyer, G. Holland, O. Vinyals, J. Scott, A. Pritzel, S. Mohamed, P. Battaglia, Learning skillful medium-range global weather forecasting, Science 382 (2023) 1416–1421. doi:10.1126/science.adi2336.DOI
  50. [50]K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, Q. Tian, Accurate medium-range global weather forecasting with 3D neural networks, Nature 619 (2023) 533–538. doi:10.1038/s41586-023-06185-3.DOI
  51. [51]S. Lang, M. Alexe, M. Chantry, J. Dramsch, F. Pinault, B. Raoult, M. C. A. Clare, C. Lessig, M. Maier-Gerber, L. Magnusson, Z. Ben Bouallègue, A. Prieto Nemesio, P. D. Dueben, A. Brown, F. Pappenberger, F. Rabier, AIFS – ECMWF’s data-driven forecasting system, 2024. doi:10.48550/arXiv.2406.01465. arXiv:2406.01465.DOI
  52. [52]European Centre for Medium-Range Weather Forecasts, ECMWF’s AI forecasts become operational, ECMWF news item, 2025. URL: https://www.ecmwf.int/node/29308, published 25 February 2025; accessed 13 October 2026.
  53. [53]R. G. Hill, K. Gao, A. Pachalieva, A. Marcato, X. Wang, P. Grosset, E. Rougier, Z. Lei, J. E. Santos, V. Agrawal, Q. Kang, J. D. Hyman, A. Hunter, C. M. Sweeney, N. DeBardeleben, E. Lawrence, H. Viswanathan, D. O’Malley, Material fracturing and failure simulation datasets, Scientific Data 13 (2025) 99. doi:10.1038/s41597-025-06412-8, published online 13 December 2025.DOI
  54. [54]S. Mohammadzadeh, E. Lejeune, Predicting mechanically driven full-field quantities of interest with deep learning-based metamodels, Extreme Mechanics Letters 50 (2022) 101566. doi:10.1016/j.eml.2021.101566.DOI
  55. [55]ASTM International, Standard Test Method for Measurement of Fracture Toughness, Standard ASTM E1820-25e2, ASTM International, West Conshohocken, PA, 2025. doi:10.1520/E1820-25DE02.
  56. [56]J. Qian, A. Fatemi, Mixed mode fatigue crack growth: A literature survey, Engineering Fracture Mechanics 55 (1996) 969–990. doi:10.1016/S0013-7944(96)00071–9.
  57. [57]M. Zhu, H. Zhang, A. Jiao, G. E. Karniadakis, L. Lu, Reliable extrapolation of deep neural operators informed by physics or sparse observations, Computer Methods in Applied Mechanics and Engineering 412 (2023) 116064. doi:10.1016/j.cma.2023.116064.DOI
  58. [58]J. A. Lara Benitez, T. Furuya, F. Faucher, A. Kratsios, X. Tricoche, M. V. de Hoop, Out-of-distribution risk bounds for neural operators with applications to the Helmholtz equation, Journal of Computational Physics 513 (2024) 113168. doi:10.1016/j.jcp.2024.113168.DOI
  59. [59]B. L. Boyce, S. L. B. Kramer, H. E. Fang, T. E. Cordova, M. K. Neilsen, K. Dion, A. K. Kaczmarowski, E. Karasz, L. Xue, A. J. Gross, et al., The Sandia Fracture Challenge: blind round robin predictions of ductile tearing, International Journal of Fracture 186 (2014) 5–68. doi:10.1007/s10704-013-9904-6.DOI
  60. [60]S. Kramer, B. Boyce, A. Jones, J. Gearhart, B. Salzbr enner, The Sandia Fracture Challenge: How ductile failure predictions fare, in: Fracture, Fatigue, Failure and Damage Evolution, Volume 6, Conference Proceedings of the Society for Experimental Mechanics Series, Springer, Cham, 2019, pp. 25–29. doi:10.1007/978-3-319-95879-8_6, published online 19 September 2018.DOI
  61. [61]S. L. B. Kramer, Preface to the special volume on the third Sandia Fracture Challenge, International Journal of Fracture 218 (2019) 1–4. doi:10.1007/s10704-019-00370-0.DOI
  62. [62]L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, G. E. Karniadakis, A comprehensive and fair comparison of two neural operators (with practical extensions) based on FAIR data, Computer Methods in Applied Mechanics and Engineering 393 (2022) 114778. doi:10.1016/j.cma.2022.114778.DOI
  63. [63]M. Takamoto, T. Praditia, R. Leiteritz, D. MacKinlay, F. Alesiani, D. Pflüger, M. Niepert, PDEBench: An extensive benchmark for scientific machine learning, in: Advances in Neural Information Processing Systems 35 (NeurIPS 2022), Datasets and Benchmarks Track, 2022, pp. 1596–1611. doi:10.52202/068431-0117. arXiv:2210.07182.DOI
  64. [64]R. Ohana, M. McCabe, L. Meyer, R. Morel, F. Agocs, M. Benítez, M. Berger, B. Burkhart, S. Dalziel, D. Fielding, D. Fortunato, J. Goldberg, K. Hirashima, Y.-F. Jiang, R. Kerswell, S. Maddu, J. Miller, P. Mukhopadhyay, S. Nixon, J. Shen, R. Watteau, B. Blancard, F. Rozet, L. Parker, M. Crammer, S. Ho, The well: a large-scale collection of diverse physics simulations for machine learning, in: Advances in Neural Information Processing Systems 37 (NeurIPS 2024), Datasets and Benchmarks Track, 2024, pp. 44989–45037. doi:10.52202/079017-1430. arXiv:2412.00568.DOI
  65. [65]M. Hakimzadeh, V. Agrawal, K. Dayal, C. Mora-Corral, Phase-field finite deformation fracture with an effective energy for regularized crack face contact, Journal of the Mechanics and Physics of Solids 167 (2022) 104994. doi:10.1016/j.jmps.2022.104994.DOI
  66. [66]M. Hakimzadeh, U. M. Krishnan, L. Graham-Brady, S. Goswami, Data and code associated with the publication: Crack paths in functionally graded plates: A phase-field simulation dataset, Johns Hopkins Research Data Repository, Version 1, 2025. URL: https://doi.org/10.7281/T1RZF13I. doi:10.7281/T1RZF13I, dataset.
  67. [67]T. Li, Gradient-damage modeling of dynamic brittle fracture: Variational principles and numerical simulations, Ph.D. thesis, Université Paris-Saclay, 2016.DOI
  68. [68]J.-Y. Wu, V. P. Nguyen, A length scale insensitive phase-field damage model for brittle fracture, Journal of the Mechanics and Physics of Solids 119 (2018) 20–42. doi:10.1016/j.jmps.2018.06.006.DOI
  69. [69]T. Chen, H. Chen, Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems, IEEE Transactions on Neural Networks 6 (1995) 911–917. doi:10.1109/72.392253.DOI
  70. [70]Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, O. Ronneberger, 3D U-Net: Learning dense volumetric segmentation from sparse annotation, in: S. Ourselin, L. Joskowicz, M. R. Sabuncu, G. Unal, W. Wells (Eds.), Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016, Lecture Notes in Computer Science, Springer, Cham, 2016, pp. 424–432. doi:10.1007/978-3-319-46723-8_49.DOI
  71. [71]S. Qiao, H. Wang, C. Liu, W. Shen, A. Yuille, Micro-batch training with batch-channel normalization and weight standardization, arXiv preprint arXiv:1903.10520 (2019).arxiv.org/abs/1903.10520

Paper details

vlnc:W_w76yfyam
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents