betaFeedback

WorksPhysics and Astronomy

A deep energy method formulation for phase-field modelling of ductile fracture

Abstract

This work extends the Deep Energy Method to phase-field modelling of ductile fracture in elasto-plastic solids. Its key contribution is an incremental DEM-based formulation that couples history-dependent J2 plasticity and plasticity-dependent fracture toughness with separate energy-optimised displacement and phase-field representations. The formulation combines the second-order Ambrosio-Tortorelli phase-field fracture model with small-strain J2 plasticity and linear isotropic hardening. A volumetric-deviatoric split defines the elastic crack-driving energy, while accumulated equivalent plastic strain governs the degradation of local fracture toughness. Plastic internal variables are updated locally through radial return. The reduced energy is minimised alternately with respect to the displacement and phase fields, represented by a ReZero residual neural network and spatially localised Gaussian radial basis functions with trainable centres, respectively. This split addresses the diffuse phase-field profiles observed in preliminary calculations with a shared global network. Alternating optimisation is combined with transfer learning to reuse trainable parameters and optimiser states across loading increments, while the converged constitutive history is carried forward separately at quadrature points. The numerical results demonstrate close agreement with finite-element reference solutions in predicting force-displacement responses, energy evolution, plastic deformation and crack propagation. The framework captures the main features of ductile fracture, although accuracy becomes more sensitive to loading increments near abrupt fracture events.

Introduction

Ductile fracture is a critical failure mechanism in metallic materials. [1, 2] It arises from the interaction between plastic deformation, damage accumulation and crack growth. At the macroscopic scale, ductile failure is accompanied by progressive losses of stiffness, strength and ductility before rupture. [3] At the microscale, damage may develop through void nucleation, growth and coalescence, as well as through micro-shear-band formation and micro-cracking. [2, 4] The resulting response depends strongly on the stress state, material hardening, geometry and loading path. [2, 5] Reliable numerical models are therefore essential for assessing structural integrity, predicting crack initiation and estimating the residual load-carrying capacity of engineering components.

Phase-field modelling provides a well-established variational framework for fracture. Cracks are represented by a continuous auxiliary field rather than explicitly tracked discontinuities. [6, 7, 8] This regularised description allows crack nucleation, propagation, branching and coalescence to be captured on a fixed computational mesh, without remeshing or separate crack-path criteria. The internal length scale controls the width of the diffuse crack representation, although adequate spatial resolution remains necessary near strongly localised regions. For elasto-plastic solids, the phase-field formulation can be coupled with plastic flow and the associated internal variables. Plastic deformation, material degradation and crack growth can then be described within the same continuum framework. [9, 10] This provides a systematic basis for representing the transition from distributed plastic deformation to macroscopic fracture.

Ductile phase-field models have progressed from early plasticity-damage couplings to formulations tailored to specific failure mechanisms. Ambati et al. [6] proposed a small-strain elasto-plastic model in which damage evolution is linked to accumulated plastic strain. This formulation captures the transition from plastic deformation to crack initiation and subsequent growth. Miehe et al. [9] developed a finite-strain variational theory that couples gradient plasticity and gradient damage through independent length scales. Borden et al. [7] subsequently introduced finite-deformation balance laws, plastic degradation and stress-triaxiality effects to improve the post-initiation response. Yin and Kaliske [10] followed a different approach, in which accumulated plastic strain reduces the local fracture toughness while the crack-driving force remains elastic. Han et al. [11] combined a plastic contribution to the crack-driving force with degrading fracture toughness, thereby linking stiffness degradation and the loss of fracture resistance. More recent developments have addressed dynamic fracture, coupled spall and adiabatic shear banding, and thermomechanically coupled finite-strain elasto-plasticity. [8, 12] These studies demonstrate the versatility of ductile phase-field modelling. Their numerical solution, however, remains predominantly based on conventional discretisation techniques.

Ductile phase-field fracture is computationally demanding because the displacement and phase fields are coupled to history-dependent plastic variables. Fine spatial resolution is often required near evolving cracks and plastic zones. The problem must also be solved incrementally, with constitutive variables updated throughout the loading history. These requirements motivate the development of alternative computational approaches based on artificial intelligence and machine learning. Physics-informed neural networks (PINNs) approximate the unknown fields using NNs trained through residuals of the governing equations, initial conditions and boundary conditions. [13] Variational neural methods, including the Deep Ritz Method and the DEM, instead employ NNs as trial functions and determine their parameters by minimising an energy functional. [14, 15, 16] Energy-based methods are particularly compatible with phase-field fracture because the governing equations follow from variational principles.

Neural approaches to fracture have developed from general physics-informed solvers towards formulations designed specifically for phase-field models. Early studies introduced variational energy-based PINNs for brittle fracture and used transfer learning to reduce the optimisation cost across successive loading increments. [17] Adaptive refinement was subsequently incorporated to improve the resolution of evolving crack paths. [18] Ghaffari Motlagh et al. [19] compared several physics-informed and energy-based formulations, with particular attention to boundary-condition enforcement and identification of the phase-field length scale. Manav et al. [20] demonstrated that the Deep Ritz Method can reproduce brittle crack nucleation, propagation, kinking, branching and coalescence. Their work also highlighted the optimisation difficulties associated with non-convex fracture energies. Further developments have addressed anisotropic and heterogeneous materials, transfer learning, and hybrid finite-element–neural-network strategies. [21, 22, 23, 24] Wang et al. [25] proposed an extended DEM that places discrete and continuous fracture descriptions within a common neural framework. In its continuous phase-field setting, separate approximations are employed for the displacement and phase fields, and RBFs are used to improve the representation of localised cracks. That formulation, however, considers elastic fracture and does not incorporate the constitutive evolution associated with plastic deformation. More recently, Zhang et al. [26] extended the DEM to thermo-mechanical brittle fracture, using separate NNs for displacement and temperature and an explicitly represented sharp crack.

Energy-based neural methods have also been applied independently to elasto-plasticity. He et al. [27] extended the DEM to classical J2J_2 plasticity and incorporated local constitutive updates through a radial-return algorithm. This established a practical treatment of plastic history dependence within an energy-based neural solver, but phase-field fracture was not included. Aldakheel et al. [28] investigated physics-based machine-learning models for brittle and ductile fracture in elastic-viscoplastic materials. Their constitutive surrogates were trained using finite-element-generated homogeneous one-dimensional responses and were subsequently embedded within boundary-value simulations. Despite these advances, a direct and label-free energy-based neural solver that combines spatial crack evolution, local elasto-plastic history and plasticity-dependent fracture resistance remains underdeveloped.

To address this gap, the present study develops a DEM-based variational formulation for ductile phase-field fracture in isotropic elasto-plastic solids. The constitutive framework follows the small-strain model of Yin and Kaliske. [10] In this model, accumulated plastic strain progressively reduces the local fracture toughness, whereas the crack-driving force is derived from the elastic strain energy. The formulation therefore distinguishes between the energetic force that drives fracture and the plasticity-induced reduction of material resistance. This setting permits the influence of plastic history on crack evolution to be examined without introducing additional thermo-mechanical fields.

The contribution of this work lies in integrating the ductile phase-field model and the DEM-based solution strategy within one incremental computational procedure. Three features are central to the formulation. Firstly, the displacement and phase fields are represented by trainable approximations, while the plastic strain and hardening variables are advanced locally using a radial-return algorithm. The path-dependent constitutive variables are therefore retained at the integration points and are not introduced as additional trainable fields. Secondly, distinct approximations are adopted for the two global fields. A ReZero residual neural network represents the displacement field, whereas spatially localised Gaussian RBFs with trainable centres resolve the localised phase-field profile. Thirdly, alternating optimisation is combined with transfer learning to reuse trainable parameters and optimiser states across loading increments, while the converged constitutive history is carried forward separately at the quadrature points. The resulting procedure couples J2J_2 plasticity, plasticity-dependent fracture resistance and irreversible crack evolution without requiring labelled training data.

Numerical benchmarks are used to evaluate force-displacement responses, energy evolution, accumulated equivalent plastic strain and crack propagation against finite-element reference solutions. The examples also examine the sensitivity of the response to loading-increment size, particularly near abrupt fracture events. The results generally agree closely with the reference solutions, with larger differences concentrated near abrupt rupture. The present work is restricted to small strains, linear isotropic hardening and isothermal conditions. More general ductile phase-field formulations, including finite-strain plasticity, thermomechanical coupling and thermo-viscoplastic behaviour, [12] require additional fields and constitutive couplings. Their incorporation into the proposed neural framework is left for future investigation.

The remainder of this manuscript is organised as follows. Section 2 introduces the adopted ductile phase-field formulation and elasto-plastic constitutive model. Section 3 presents the DEM implementation and incremental solution procedure. Section 4 reports the numerical examples and comparisons with the reference solutions. Finally, Section 5 summarises the main findings and outlines directions for future research.

Phase-field formulation for ductile fracture

This section defines the ductile phase-field model adopted in this work. The formulation follows the small-strain framework of Yin and Kaliske, [10] in which accumulated equivalent plastic strain reduces the local fracture toughness. The crack-driving force is derived from the elastic strain energy, while the plastic internal variables evolve through associative J2J_2 plasticity with linear isotropic hardening.

Variational form

Consider a deformable body occupying a domain Ω⊂Rd\Omega\subset\mathbb{R}^{d}, where dd denotes the spatial dimension and x∈Ω\mathbf{x} \in\Omega denotes the position vector. Its boundary is decomposed as ∂Ω=∂Ωuˉ∪∂Ωhˉ\partial\Omega= \partial\Omega_{\bar{u}} \cup\partial\Omega_{\bar{h}}, with ∂Ωuˉ∩∂Ωhˉ=∅\partial\Omega_{\bar{u}} \cap\partial\Omega_{\bar{h}} = \varnothing. The displacement satisfies u=uˉ\mathbf{u} = \bar{\mathbf{u}} on ∂Ωuˉ\partial\Omega_{\bar{u}}, while the prescribed traction is hˉ\bar{\mathbf{h}} on ∂Ωhˉ\partial\Omega_{\bar{h}}. A body force fb\mathbf{f}^{\mathrm{b}} may act in Ω\Omega.

The sharp crack set Γ\Gamma is regularised by a scalar phase field c:Ω→[0,1]c : \Omega\to[0,1]. The value c=1c = 1 denotes intact material and c=0c = 0 denotes fully fractured material. The formulation is written in general form, while the numerical examples considered in this work are restricted to two-dimensional plane-strain problems. The J2J_2 invariants and tensor contractions retain their standard three-dimensional definitions.

Under quasi-static loading, the incremental total potential energy is written as

Π(u,c,εp,α,ζ)=Wint(u,c,εp,α,ζ)−Wext(u)+Πirr(c;cn−1),(1)\Pi(\mathbf{u},c,\boldsymbol{\varepsilon}^{\mathrm{p}},\alpha,\zeta) = \mathcal{W}^{\mathrm{int}}(\mathbf{u},c,\boldsymbol{\varepsilon}^{\mathrm{p}},\alpha,\zeta) - \mathcal{W}^{\mathrm{ext}}(\mathbf{u}) + \Pi^{\mathrm{irr}}(c;c_{n-1}), \tag*{(1)}

where εp\boldsymbol{\varepsilon}^{\mathrm{p}} is the plastic strain tensor, α\alpha is the isotropic-hardening variable and ζ\zeta is the accumulated equivalent plastic strain. The previously converged phase field, cn−1c_{n-1}, enters the irreversibility penalty Πirr\Pi^{\mathrm{irr}} as defined in Section 2.6 below.

The external work is

Wext(u)=∫∂Ωhˉhˉ⋅u dS+∫Ωfb⋅u dΩ.(2)\mathcal{W}^{\mathrm{ext}}(\mathbf{u}) = \int_{\partial\Omega_{\bar{h}}} \bar{\mathbf{h}} \cdot\mathbf{u}\,\mathrm{d}S + \int_{\Omega} \mathbf{f}^{\mathrm{b}} \cdot\mathbf{u}\,\mathrm{d}\Omega. \tag*{(2)}

The internal contribution is decomposed as

Wint≡Ψe+Ψp+Ψf.(3)\mathcal{W}^{\mathrm{int}} \equiv\Psi^{\mathrm{e}} + \Psi^{\mathrm{p}} + \Psi^{\mathrm{f}}. \tag*{(3)}

The elastic contribution contains crack-driving and non-driving parts:

Ψe=∫Ω[g(c)ψe+(εe)+ψe−(εe)] dΩ.(4)\Psi^{\mathrm{e}} = \int_{\Omega} \left[g(c)\psi^{\mathrm{e}+}(\varepsilon^{\mathrm{e}}) + \psi^{\mathrm{e}-}(\varepsilon^{\mathrm{e}})\right]\,\mathrm{d}\Omega. \tag*{(4)}

The elastic strain follows the additive small-strain decomposition

εe=ε−εp,ε=12(∇u+(∇u)T).(5)\varepsilon^{\mathrm{e}} = \varepsilon- \varepsilon^{\mathrm{p}}, \qquad\varepsilon= \frac{1}{2}\left(\nabla\mathbf{u} + (\nabla\mathbf{u})^{\mathrm{T}}\right). \tag*{(5)}

Only the crack-driving contribution is degraded. With the present convention for cc, the quadratic degradation function is

g(c)=c2.(6)g(c) = c^{2}. \tag*{(6)}

The plastic contribution is

Ψp=∫Ωψp(α) dΩ,(7)\Psi^{\mathrm{p}} = \int_{\Omega} \psi^{\mathrm{p}}(\alpha)\,\mathrm{d}\Omega, \tag*{(7)}

where ψp\psi^{\mathrm{p}} is defined in Section 2.3.

The crack-surface contribution follows the AT2 regularisation:

Ψf=∫ΓGc(ζ) dΓ≈∫ΩGc(ζ)Zc(c,∇c) dΩ,(8)\Psi^{\mathrm{f}} = \int_{\Gamma} \mathcal{G}_{\mathrm{c}}(\zeta)\,\mathrm{d}\Gamma\approx\int_{\Omega} \mathcal{G}_{\mathrm{c}}(\zeta)\mathcal{Z}_{\mathrm{c}}(c,\nabla c)\,\mathrm{d}\Omega, \tag*{(8)}

with

Zc=(1−c)22l0+l02∣∇c∣2,(9)\mathcal{Z}_{\mathrm{c}} = \frac{(1-c)^{2}}{2l_{0}} + \frac{l_{0}}{2}\lvert\nabla c\rvert^{2}, \tag*{(9)}

where l0∈R+l_{0} \in\mathbb{R}^{+} is the phase-field length-scale parameter. The dependence of Gc\mathcal{G}_{\mathrm{c}} on ζ\zeta introduces the toughness-degradation mechanism used to represent ductile fracture. A schematic representation of the phase-field approximation for ductile fracture is presented in Figure 1.

Figure 1. (a) Solid body subjected to prescribed boundary conditions and capable of undergoing plastic deformation and developing a crack path Γ\Gamma, and (b) its phase-field approximation.

Elastic energy split

The crack-driving and non-driving elastic energy densities in Eq. (4) are defined using the volumetric-deviatoric split adopted by Yin and Kaliske. [10] This split is consistent with classical J2J_{2} plasticity and excludes the compressive volumetric contribution from the crack-driving energy.

The first invariant and deviatoric part of the elastic strain are

I1=tr⁡(εe),εdeve=εe−13I1I,(10)I_1 = \operatorname{tr}(\boldsymbol{\varepsilon}^{e}), \qquad\boldsymbol{\varepsilon}^{e}_{\mathrm{dev}} = \boldsymbol{\varepsilon}^{e} - \frac{1}{3}I_1\mathbf{I}, \tag*{(10)}

where I\mathbf{I} is the second-order identity tensor. The two energy densities are

ψe+(εe)=12κ⟨I1⟩+2+μ(εdeve:εdeve),ψe−(εe)=12κ⟨I1⟩−2,(11)\psi^{e+}(\boldsymbol{\varepsilon}^{e}) = \frac{1}{2}\kappa\langle I_1\rangle_{+}^{2} + \mu\left(\boldsymbol{\varepsilon}^{e}_{\mathrm{dev}}:\boldsymbol{\varepsilon}^{e}_{\mathrm{dev}}\right), \qquad\psi^{e-}(\boldsymbol{\varepsilon}^{e}) = \frac{1}{2}\kappa\langle I_1\rangle_{-}^{2}, \tag*{(11)}

where κ\kappa and μ\mu are the bulk and shear moduli. The Macaulay brackets are defined by

⟨x⟩+=12(x+∣x∣),⟨x⟩−=12(x−∣x∣).(12)\langle x\rangle_{+} = \frac{1}{2}(x+|x|), \qquad\langle x\rangle_{-} = \frac{1}{2}(x-|x|). \tag*{(12)}

The corresponding stress response is

σ=κ[g(c)⟨I1⟩++⟨I1⟩−]I+2μg(c)εdeve.(13)\boldsymbol{\sigma} = \kappa\left[g(c)\langle I_1\rangle_{+}+\langle I_1\rangle_{-}\right]\mathbf{I}+2\mu g(c)\boldsymbol{\varepsilon}^{e}_{\mathrm{dev}}. \tag*{(13)}

Consequently, the deviatoric stress entering the plasticity model is

σdev=2μg(c)εdeve.(14)\boldsymbol{\sigma}_{\mathrm{dev}}=2\mu g(c)\boldsymbol{\varepsilon}^{e}_{\mathrm{dev}}. \tag*{(14)}

Thus, the tensile volumetric contribution and the full deviatoric contribution enter the crack-driving energy, whereas the compressive volumetric contribution remains undegraded.

Plastic energy and hardening

The plastic response is described by associative J2J_2 plasticity with linear isotropic hardening. The plastic energy density is

ψp(α)=σyα+12Hα2,(15)\psi^{\mathrm{p}}(\alpha)=\sigma_y\alpha+\frac{1}{2}H\alpha^2, \tag*{(15)}

where σy\sigma_y is the initial yield stress and H≥0H\geq0 is the hardening modulus. The thermodynamic force conjugate to α\alpha is

tα=∂ψp∂α=σy+Hα.(16)t_\alpha=\frac{\partial\psi^{\mathrm{p}}}{\partial\alpha}=\sigma_y+H\alpha. \tag*{(16)}

The yield function is

F(σ,α)=∥σdev∥−23tα≤0,(17)\mathcal{F}(\boldsymbol{\sigma},\alpha)=\lVert\boldsymbol{\sigma}_{\mathrm{dev}}\rVert-\sqrt{\frac{2}{3}}t_\alpha\leq0, \tag*{(17)}

where ∥⋅∥\lVert\cdot\rVert denotes the Frobenius norm. For plastic loading, the associative flow rule and hardening evolution are

ε˙p=χ˙N,N=σdev∥σdev∥,α˙=23χ˙,(18)\dot{\boldsymbol{\varepsilon}}^{p}=\dot{\chi}\mathbf{N}, \qquad\mathbf{N}=\frac{\boldsymbol{\sigma}_{\mathrm{dev}}}{\lVert\boldsymbol{\sigma}_{\mathrm{dev}}\rVert}, \qquad\dot{\alpha}=\sqrt{\frac{2}{3}}\dot{\chi}, \tag*{(18)}

where χ˙\dot{\chi} is the plastic multiplier. Plastic incompressibility follows from tr⁡(ε˙p)=0\operatorname{tr}(\dot{\boldsymbol{\varepsilon}}^{p})=0. The admissible evolution satisfies the Kuhn–Tucker conditions

χ˙≥0,F≤0,Fχ˙=0.(19)\dot{\chi}\geq0, \qquad\mathcal{F}\leq0, \qquad\mathcal{F}\dot{\chi}=0. \tag*{(19)}

The accumulated equivalent plastic strain controlling the fracture toughness is

ζ(t)=23∫0tε˙p:ε˙p dτ.(20)\zeta(t)=\sqrt{\frac{2}{3}}\int_{0}^{t}\sqrt{\dot{\boldsymbol{\varepsilon}}^{p}:\dot{\boldsymbol{\varepsilon}}^{p}}\,\mathrm{d}\tau. \tag*{(20)}

Since N:N=1\mathbf{N}:\mathbf{N}=1, Eq. (18) gives ζ˙=2/3 χ˙\dot{\zeta}=\sqrt{2/3}\,\dot{\chi}. Therefore, α\alpha and ζ\zeta evolve identically when their initial values coincide. They are retained as separate symbols to distinguish their roles in isotropic hardening and fracture-toughness degradation.

Fracture-toughness degradation

Following Yin and Kaliske, [10] ductile effects are introduced by allowing the accumulated equivalent plastic strain to reduce the local fracture toughness:

Gc(ζ)=f(ζ)Gc0,(21)\mathcal{G}_{c}(\zeta)=f(\zeta)\mathcal{G}_{c}^{0}, \tag*{(21)}

where Gc0\mathcal{G}_{c}^{0} is the undegraded fracture toughness. It represents the critical energy release rate of the corresponding brittle response before the plasticity-induced reduction becomes active. The degradation function satisfies

0<f(ζ)≤1,f(ζ)=1 for ζ≤ζcr,dfdζ≤0 where defined for ζ>ζcr.(22)0 < f(\zeta) \le1,\qquad f(\zeta)=1\ \text{for }\zeta\le\zeta^{\mathrm{cr}},\qquad\frac{\mathrm{d}f}{\mathrm{d}\zeta}\le0\ \text{where defined for }\zeta>\zeta^{\mathrm{cr}}. \tag*{(22)}

The parameter ζcr\zeta^{\mathrm{cr}} denotes the accumulated equivalent plastic strain threshold at which toughness degradation begins. The function used in this work is

f(ζ)={1,ζ<ζcr,1−bξ2(ζ−ζcr−ξ)2+b,ζcr≤ζ<ξ+ζcr,b,ζ≥ξ+ζcr.(23)f(\zeta)= \begin{cases} 1, & \zeta<\zeta^{\mathrm{cr}},\\ \frac{1-b}{\xi^{2}}\left(\zeta-\zeta^{\mathrm{cr}}-\xi\right)^{2}+b, & \zeta^{\mathrm{cr}}\le\zeta<\xi+\zeta^{\mathrm{cr}},\\ b, & \zeta\ge\xi+\zeta^{\mathrm{cr}}. \end{cases} \tag*{(23)}

where ξ>0\xi>0 controls the strain interval over which the toughness decreases and 0<b≤10<b\le1 is its residual fraction. A strictly positive value of bb prevents the local fracture toughness from vanishing. Consistent with the assumption adopted by Yin and Kaliske,[10] f(ζ)f(\zeta) is treated as a known history-dependent coefficient during each variational subproblem. Its derivative with respect to the current plastic variables is therefore not included when differentiating the subproblem. This is an algorithmic assumption, rather than a general identity for Eq. (23). Because f(ζ)f(\zeta) remains inside the fracture-energy integral in Eq. (8), its spatial variation is retained. In an equivalent strong form, this variation generates a contribution involving ∇f(ζ)⋅∇c\nabla f(\zeta)\cdot\nabla c.

Radial-return algorithm

The plastic internal variables are updated locally using the radial-return algorithm of associative J2J_{2} plasticity.[10, 29] Consider the increment from tnt_n to tn+1t_{n+1}. The converged history (εnp,αn,ζn)(\boldsymbol{\varepsilon}^{p}_{n},\alpha_n,\zeta_n) is known. For a candidate strain εn+1\boldsymbol{\varepsilon}_{n+1} and a fixed phase-field value cn+1c_{n+1}, the trial elastic deviatoric strain and trial deviatoric stress are

εdev,n+1e,tr=εdev,n+1−εdev,np,σdev,n+1tr=2μg(cn+1)εdev,n+1e,tr.(24)\boldsymbol{\varepsilon}^{e,\mathrm{tr}}_{\mathrm{dev},n+1} =\boldsymbol{\varepsilon}_{\mathrm{dev},n+1}-\boldsymbol{\varepsilon}^{p}_{\mathrm{dev},n}, \qquad \boldsymbol{\sigma}^{\mathrm{tr}}_{\mathrm{dev},n+1} =2\mu g(c_{n+1})\boldsymbol{\varepsilon}^{e,\mathrm{tr}}_{\mathrm{dev},n+1}. \tag*{(24)}

The corresponding trial yield function is

Fn+1tr=∥σdev,n+1tr∥−23(σy+Hαn).(25)\mathcal{F}^{\mathrm{tr}}_{n+1} =\left\lVert\boldsymbol{\sigma}^{\mathrm{tr}}_{\mathrm{dev},n+1}\right\rVert -\sqrt{\frac{2}{3}}\left(\sigma_y+H\alpha_n\right). \tag*{(25)}

If Fn+1tr≤0\mathcal{F}^{\mathrm{tr}}_{n+1}\le0, the step is elastic and

εn+1p=εnp,αn+1=αn,ζn+1=ζn,σdev,n+1=σdev,n+1tr.(26)\boldsymbol{\varepsilon}^{p}_{n+1}=\boldsymbol{\varepsilon}^{p}_{n},\qquad \alpha_{n+1}=\alpha_n,\qquad \zeta_{n+1}=\zeta_n,\qquad \boldsymbol{\sigma}_{\mathrm{dev},n+1}=\boldsymbol{\sigma}^{\mathrm{tr}}_{\mathrm{dev},n+1}. \tag*{(26)}

If Fn+1tr>0\mathcal{F}^{\mathrm{tr}}_{n+1}>0, the trial state is returned to the yield surface along

Nn+1=σdev,n+1tr∥σdev,n+1tr∥.(27)\mathbf{N}_{n+1} =\frac{\boldsymbol{\sigma}^{\mathrm{tr}}_{\mathrm{dev},n+1}} {\left\lVert\boldsymbol{\sigma}^{\mathrm{tr}}_{\mathrm{dev},n+1}\right\rVert}. \tag*{(27)}

For linear isotropic hardening, the plastic-multiplier increment has the closed form

Δχ=Fn+1tr2μg(cn+1)+23H.(28)\Delta\chi =\frac{\mathcal{F}^{\mathrm{tr}}_{n+1}} {2\mu g(c_{n+1})+\frac{2}{3}H}. \tag*{(28)}

The corrected variables are then

εn+1p=εnp+ΔχNn+1,(29)\boldsymbol{\varepsilon}^{p}_{n+1} =\boldsymbol{\varepsilon}^{p}_{n}+\Delta\chi\mathbf{N}_{n+1}, \tag*{(29)}
αn+1=αn+23Δχ,\alpha_{n+1}=\alpha_n+\sqrt{\frac{2}{3}}\Delta\chi,
ζn+1=ζn+23Δχ,\zeta_{n+1}=\zeta_n+\sqrt{\frac{2}{3}}\Delta\chi,
σdev,n+1=σdev,n+1tr−2μg(cn+1)ΔχNn+1.\boldsymbol{\sigma}_{\mathrm{dev},n+1} =\boldsymbol{\sigma}^{\mathrm{tr}}_{\mathrm{dev},n+1} -2\mu g(c_{n+1})\Delta\chi\mathbf{N}_{n+1}.

The plastic correction enforces Fn+1=0\mathcal{F}_{n+1}=0 up to numerical precision. The complete stress is recovered as

σn+1=κ[g(cn+1)⟨I1,n+1⟩++⟨I1,n+1⟩−]I+σdev,n+1.(30)\boldsymbol{\sigma}_{n+1} =\kappa\left[g(c_{n+1})\langle I_{1,n+1}\rangle_{+} +\langle I_{1,n+1}\rangle_{-}\right]\mathbf{I} +\boldsymbol{\sigma}_{\mathrm{dev},n+1}. \tag*{(30)}

After the return mapping, f(ζn+1)f(\zeta_{n+1}) determines the local degraded fracture toughness used in the phase-field subproblem. During the global iterations at a given loading increment, every constitutive trial update starts from the same converged history at tnt_n. The internal variables at tn+1t_{n+1} are committed only after the displacement and phase fields have converged. This procedure preserves the constitutive history without introducing εp\varepsilon^p, α\alpha or ζ\zeta as trainable fields. A schematic representation of the return algorithm is presented in Figure 2.

Radial-return mapping for the degraded $J_2$ constitutive response

Figure 2. Radial-return mapping for the degraded J2J_2 constitutive response. The trial state is accepted when Ftr≤0\mathcal{F}^{\mathrm{tr}} \le0 and otherwise projected onto the current yield surface by the plastic corrector.

Irreversibility

Crack healing is prevented between successive loading increments. Since c=1c = 1 denotes intact material and c=0c = 0 denotes fully fractured material, irreversibility requires

cn≤cn−1in Ω.(31)c_n \le c_{n-1} \quad\text{in } \Omega. \tag*{(31)}

Following the penalty treatment used in variational phase-field formulations, [20, 21, 30] and accounting for the present convention for cc, the constraint is enforced by adding

Pirr=∫Ω12κirr⟨cn−1−cn⟩2 dΩ.(32)\mathcal{P}^{\mathrm{irr}}=\int_{\Omega}\frac{1}{2}\kappa_{\mathrm{irr}}\langle c_{n-1}-c_n\rangle^2\,\mathrm{d}\Omega. \tag*{(32)}

The penalty is active only when cn>cn−1c_n>c_{n-1}, which represents an inadmissible recovery of material integrity.

For the AT2 crack density in Eq. (9), the penalty parameter is chosen using the AT2 scaling of Gerasimov and De Lorenzis, [30] with the undegraded fracture toughness used as the reference scale:

κirr=Gc0l0(1tolirr2−1),(33)\kappa_{\mathrm{irr}}=\frac{\mathcal{G}_c^0}{l_0}\left(\frac{1}{\mathrm{tol}_{\mathrm{irr}}^2}-1\right), \tag*{(33)}

where 0<tolirr≤10<\mathrm{tol}_{\mathrm{irr}}\leq1 is the prescribed irreversibility tolerance, set to 0.02 in this work. Using Gc0\mathcal{G}_c^0 provides a fixed global penalty scale and prevents the irreversibility penalty from decreasing as the local fracture toughness degrades.

Non-dimensionalisation scheme

A non-dimensionalisation scheme is introduced in this work for the governing quantities prior to neural optimisation to improve numerical conditioning, building on the scaling approaches of Manav et al. [20] and Plunge et al. [21] Let LL be a characteristic dimension of the body and let Eref=EE_{\mathrm{ref}}=E be the Young’s modulus of the isotropic solid, where

E=9κμ3κ+μ.(34)E=\frac{9\kappa\mu}{3\kappa+\mu}. \tag*{(34)}

The scaling is based on the undegraded fracture toughness Gc0\mathcal{G}_c^0 and the AT2 length scale l0l_0, providing fixed reference quantities throughout the loading history. The reference strain, stress, displacement and energy scales are

εref=Gc0Erefl0,σref=Erefεref=ErefGc0l0,uref=Lεref,Πref=LdGc0l0.(35)\varepsilon_{\mathrm{ref}}=\sqrt{\frac{\mathcal{G}_c^0}{E_{\mathrm{ref}}l_0}},\qquad\sigma_{\mathrm{ref}}=E_{\mathrm{ref}}\varepsilon_{\mathrm{ref}}=\sqrt{\frac{E_{\mathrm{ref}}\mathcal{G}_c^0}{l_0}},\qquad u_{\mathrm{ref}}=L\varepsilon_{\mathrm{ref}},\qquad\Pi_{\mathrm{ref}}=\frac{L^d\mathcal{G}_c^0}{l_0}. \tag*{(35)}

The dimensionless spatial, kinematic and history variables are defined by

x~=xL,l~0=l0L,u~=uuref,ε~=εεref,ε~p=εpεref,α~=αεref,ζ~=ζεref,ζ~cr=ζcrεref,ξ~=ξεref.(36)\begin{aligned} \widetilde{\mathbf{x}}&=\frac{\mathbf{x}}{L},\qquad\widetilde{l}_0=\frac{l_0}{L},\qquad\widetilde{\mathbf{u}}=\frac{\mathbf{u}}{u_{\mathrm{ref}}},\qquad\widetilde{\boldsymbol{\varepsilon}}=\frac{\boldsymbol{\varepsilon}}{\varepsilon_{\mathrm{ref}}},\qquad\widetilde{\boldsymbol{\varepsilon}}^p=\frac{\boldsymbol{\varepsilon}^p}{\varepsilon_{\mathrm{ref}}},\\ \widetilde{\alpha}&=\frac{\alpha}{\varepsilon_{\mathrm{ref}}},\qquad\widetilde{\zeta}=\frac{\zeta}{\varepsilon_{\mathrm{ref}}},\qquad\widetilde{\zeta}^{\mathrm{cr}}=\frac{\zeta^{\mathrm{cr}}}{\varepsilon_{\mathrm{ref}}},\qquad\widetilde{\xi}=\frac{\xi}{\varepsilon_{\mathrm{ref}}}. \tag*{(36)} \end{aligned}

The differential operators and integration measures transform according to

∇~=L∇,dΩ=Ld dΩ~,dS=Ld−1 dS~.(37)\widetilde{\nabla}=L\nabla,\qquad\mathrm{d}\Omega=L^d\,\mathrm{d}\widetilde{\Omega},\qquad\mathrm{d}S=L^{d-1}\,\mathrm{d}\widetilde{S}. \tag*{(37)}

The phase field cc and the residual toughness fraction bb are already dimensionless. The dimensionless material parameters are

κ~=κEref,μ~=μEref,H~=HEref,σ~y=σyσref,σ~=σσref.(38)\widetilde{\kappa}=\frac{\kappa}{E_{\mathrm{ref}}},\qquad\widetilde{\mu}=\frac{\mu}{E_{\mathrm{ref}}},\qquad\widetilde{H}=\frac{H}{E_{\mathrm{ref}}},\qquad\widetilde{\sigma}_y=\frac{\sigma_y}{\sigma_{\mathrm{ref}}},\qquad\widetilde{\boldsymbol{\sigma}}=\frac{\boldsymbol{\sigma}}{\sigma_{\mathrm{ref}}}. \tag*{(38)}

For isotropic elasticity,

κ~=13(1−2ν),μ~=12(1+ν),(39)\widetilde{\kappa}=\frac{1}{3(1-2\nu)},\qquad\widetilde{\mu}=\frac{1}{2(1+\nu)}, \tag*{(39)}

where ν\nu is Poisson’s ratio.

The arguments of the toughness-degradation law are scaled consistently. Since ζ=εrefζ~\zeta=\varepsilon_{\mathrm{ref}}\widetilde{\zeta}, ζcr=εrefζ~cr\zeta^{\mathrm{cr}}=\varepsilon_{\mathrm{ref}}\widetilde{\zeta}^{\mathrm{cr}} and ξ=εrefξ~\xi=\varepsilon_{\mathrm{ref}}\widetilde{\xi},

G~c=GcGc0=f(ζ~;ζ~cr,ξ~,b).(40)\widetilde{\mathcal{G}}_c=\frac{\mathcal{G}_c}{\mathcal{G}_c^0}=f\left(\widetilde{\zeta};\widetilde{\zeta}^{\mathrm{cr}},\widetilde{\xi},b\right). \tag*{(40)}

The dimensionless degradation function therefore retains the same piecewise form as Eq. (23).

The penalty parameter becomes

κ~irr=l0κirrGc0=1tolirr2−1.(41)\widetilde{\kappa}_{\mathrm{irr}}=\frac{l_0\kappa_{\mathrm{irr}}}{\mathcal{G}_c^0}=\frac{1}{\mathrm{tol}_{\mathrm{irr}}^2}-1. \tag*{(41)}

With Π~=Π/Πref\widetilde{\Pi}=\Pi/\Pi_{\mathrm{ref}}, the dimensionless internal and irreversibility contributions are

W~int=∫Ω~[g(c)ψ~e++ψ~e−+σ~yα~+12H~α~2+f(ζ~)((1−c)22+l~022∣∇~c∣2)]dΩ~,P~irr=∫Ω~12κ~irr⟨cn−1−cn⟩−2 dΩ~.(42)\begin{aligned} \widetilde{\mathcal{W}}^{\mathrm{int}}&=\int_{\widetilde{\Omega}}\left[g(c)\widetilde{\psi}^{e+}+\widetilde{\psi}^{e-}+\widetilde{\sigma}_y\widetilde{\alpha}+\frac{1}{2}\widetilde{H}\widetilde{\alpha}^{2}+f(\widetilde{\zeta})\left(\frac{(1-c)^2}{2}+\frac{\widetilde{l}_0^2}{2}\lvert\widetilde{\nabla}c\rvert^2\right)\right]\mathrm{d}\widetilde{\Omega},\\ \widetilde{\mathcal{P}}^{\mathrm{irr}}&=\int_{\widetilde{\Omega}}\frac{1}{2}\widetilde{\kappa}_{\mathrm{irr}}\langle c_{n-1}-c_n\rangle_-^2\,\mathrm{d}\widetilde{\Omega}. \tag*{(42)} \end{aligned}

where

ψ~e±=l0Cc0ψe±.(43)\widetilde{\psi}^{e\pm}=\frac{l_0}{C_c^0}\psi^{e\pm}. \tag*{(43)}

The dimensionless traction and body force are

h~=hσref,f~b=Lfbσref.(44)\widetilde{\mathbf{h}}=\frac{\mathbf{h}}{\sigma_{\mathrm{ref}}},\qquad\widetilde{\mathbf{f}}^{b}=\frac{L\mathbf{f}^{b}}{\sigma_{\mathrm{ref}}}. \tag*{(44)}

Finally, defining

F~=Fσref,Δχ~=Δχεref,(45)\widetilde{\mathcal{F}}=\frac{\mathcal{F}}{\sigma_{\mathrm{ref}}},\qquad\Delta\widetilde{\chi}=\frac{\Delta\chi}{\varepsilon_{\mathrm{ref}}}, \tag*{(45)}

allows the yield function and radial-return equations to retain their original algebraic form. For notational convenience, the tilde notation is omitted in the subsequent governing and computational equations. Dimensional material parameters are nevertheless reported in physical units in the numerical examples and are non-dimensionalised before optimisation.

DEM-based formulation for phase-field ductile fracture

Deep energy method

The DEM is a variational physics-informed approach in which trainable functions are used as trial approximations of the solution fields and their parameters are determined through minimisation of an energy-based objective. [15, 16] It follows the general structure of the Deep Ritz Method, combining a trainable approximation space, numerical integration of the variational functional and optimisation of the resulting parameterised problem. [14] Unlike standard residual-based PINNs, the governing equations are not imposed through pointwise strong-form residuals. Instead, the solution fields are determined through minimisation of the variational functional. [20, 21]

In the present problem, the global trainable fields are the displacement and phase fields,

u(x)≈uθu(x),c(x)≈cθc(x),(46)\mathbf{u}(\mathbf{x})\approx\mathbf{u}_{\boldsymbol{\theta}_u}(\mathbf{x}),\qquad c(\mathbf{x})\approx c_{\boldsymbol{\theta}_c}(\mathbf{x}), \tag*{(46)}

where θu\boldsymbol{\theta}_u and θc\boldsymbol{\theta}_c collect the trainable parameters of the corresponding representations. The two fields use separate trainable representations, as described in § 1.3.2.

The elastoplastic internal variables are not introduced as additional trainable fields. Instead, for a given trial state they are evaluated locally at the quadrature points using the radial-return algorithm of § 1.1.5. Denoting the local constitutive state by

qn⋆=(εnp⋆,αn⋆,ζn⋆),(47)\mathbf{q}_n^\star=\left(\boldsymbol{\varepsilon}_n^{p\star},\alpha_n^\star,\zeta_n^\star\right), \tag*{(47)}

the local constitutive update may be written symbolically as

qn⋆=R(ε(uθu),cθc;qn−1),(48)\mathbf{q}_n^\star=\mathcal{R}\left(\boldsymbol{\varepsilon}(\mathbf{u}_{\boldsymbol{\theta}_u}),c_{\boldsymbol{\theta}_c};\mathbf{q}_{n-1}\right), \tag*{(48)}

where R\mathcal{R} denotes the radial-return operator and qn−1\mathbf{q}_{n-1} is the constitutive history committed at the preceding loading increment. Every trial constitutive evaluation within increment nn starts from this same committed history. The updated variables are committed only after convergence of the global fields.

The trainable fields are evaluated at the nodes of the background mesh and interpolated to the quadrature points, where the strain, phase-field gradient and constitutive response are calculated. The resulting energy contributions define the training objective introduced in § 1.3. Numerical integration is discussed separately in § 1.3.1.

Figure 3 summarises the incremental procedure. The trainable fields are updated through alternating optimisation, while the constitutive variables remain local quantities governed by the material model.

Figure 3. Overview of the proposed incremental DEM. (a) Evaluation of a training block using the trainable field representations, admissibility and Dirichlet mappings, finite-element interpolation, the energy-based loss and the block-level stopping criterion. (b) Alternating optimisation of the phase and displacement fields, followed by physics-based convergence checks and commitment of the converged constitutive history.

Training loss

The training objective is constructed from the elastic, fracture and irreversibility contributions introduced in § 1.1. For the displacement-controlled problems considered in this work, loading is imposed through prescribed displacements and no prescribed tractions or body forces are applied. Consequently, the explicit external-work term Wext\mathcal{W}^{\mathrm{ext}} defined in Eq. (2) vanishes. The reduced training energy is therefore

En(θu,θc)=Ψne+Ψnf+Pnirr,(49)\mathcal{E}_n(\boldsymbol{\theta}_u,\boldsymbol{\theta}_c)=\Psi_n^{\mathrm{e}}+\Psi_n^{\mathrm{f}}+\mathcal{P}_n^{\mathrm{irr}}, \tag*{(49)}

where the elastic contribution is evaluated using the local constitutive state returned by Eq. (48).

The plastic contribution Ψp\Psi^{\mathrm{p}} is not included explicitly in the global training objective. This reflects the local treatment of the plastic internal variables rather than their omission from the constitutive model. During optimisation of the displacement field, the radial-return algorithm evaluates the plastic state and corresponding stress for each candidate displacement field. The resulting constitutive state is treated as fixed during differentiation, so gradients are not propagated through the return-mapping update. The resulting first variation is therefore obtained from the elastic contribution evaluated at the return-mapped constitutive state.

During optimisation of the phase field, the most recently available trial constitutive state is held fixed, according to the adopted staggered constitutive treatment. Consequently, Ψp\Psi^{\mathrm{p}} is constant with respect to the active phase-field parameters. The toughness-degradation factor f(ζn⋆)f(\zeta_n^{\star}) is treated in the same manner. Its value is retained locally within the fracture-energy integral, but gradients are not propagated through the constitutive history, consistently with § 1.1.4. Thus, plastic deformation influences the phase-field problem through the spatially varying degraded toughness without introducing the plastic variables into the global trainable parameter space.

With the non-dimensionalisation introduced in § 1.1.7, En\mathcal{E}_n is dimensionless. For En>0\mathcal{E}_n>0, the scalar loss used for optimisation is

Ln=log⁡10[En(θu,θc)].(50)\mathcal{L}_n=\log_{10}\left[\mathcal{E}_n(\boldsymbol{\theta}_u,\boldsymbol{\theta}_c)\right]. \tag*{(50)}

The use of a logarithmic energy objective follows the scaling strategy used in previous energy-based neural fracture formulations. [20, 21] It reduces the dynamic range of the objective and facilitates the use of consistent optimisation and stopping criteria across loading increments.

Differentiating the loss gives

∇θLn=1ln⁡(10)En∇θEn,(51)\nabla_{\boldsymbol{\theta}}\mathcal{L}_n=\frac{1}{\ln(10)\mathcal{E}_n}\nabla_{\boldsymbol{\theta}}\mathcal{E}_n, \tag*{(51)}

where θ\boldsymbol{\theta} denotes the active trainable parameter block. The logarithmic transformation therefore multiplies the energy gradient by a positive scalar. It preserves the stationary points and the component-wise signs of the gradient while rescaling its magnitude.

Numerical quadrature

The energy functional is evaluated over a background spatial discretisation comprising four-node quadrilateral elements. [31] The trainable displacement and phase fields are first evaluated at the element nodes, while their values and spatial gradients at the quadrature points are obtained using the corresponding bilinear shape functions. This follows the numerical-gradient strategy of Manav et al., [20] who reported spurious high-frequency oscillations in the learned fields when their gradients were computed using automatic differentiation. A related interpolation-based differentiation strategy using higher-order B-spline basis functions was adopted in Ref. 21. The energy contributions are then integrated element-wise using Gauss quadrature. At each quadrature point, the interpolated strain and phase-field value are supplied to the local constitutive update, and the resulting constitutive state is used to evaluate the elastic response and the plasticity-dependent fracture toughness. The interpolated phase-field gradient enters directly into the fracture-energy contribution.

Trainable field representations

The displacement and phase fields exhibit markedly different spatial behaviour. The displacement field is relatively smooth away from the fracture zone, although steep gradients may develop near crack localisation. The phase field develops a localised transition whose characteristic width is controlled by the phase-field length scale l0l_0. Preliminary calculations in the present study using a single shared global neural-network representation for both fields did not reproduce this localisation reliably. In particular, the displacement response could converge while the phase field remained excessively diffuse. Separate trainable representations are therefore adopted for u\mathbf{u} and cc. This choice also supports the alternating optimisation strategy introduced in § 1.3.4.

Displacement

The displacement field is represented by a ReZero-based residual NN adapted from the architecture previously developed for brittle anisotropic fracture. [21] Residual networks learn corrections to an identity mapping through shortcut connections. [32] ReZero modifies this structure by introducing a trainable scalar coefficient on each residual branch and initialising it to zero. [33] Each residual block therefore acts initially as an identity mapping and progressively develops non-trivial transformations during optimisation.

Following the non-dimensionalisation introduced in § 1.1.7, the spatial coordinates supplied to the displacement network are additionally standardised component-wise. This removes differences in coordinate scale associated with the spatial extent of the domain in each direction. For coordinate component i=1,…,di = 1, \ldots, d, the network input is defined as

xiin=sxxi−μiσi,i=1,…,d,(52)x_i^{\mathrm{in}} = s_x \frac{x_i-\mu_i}{\sigma_i}, \qquad i = 1, \ldots, d, \tag*{(52)}

where μi\mu_i and σi\sigma_i are the mean and standard deviation, respectively, of the nodal coordinates in direction ii. The scalar sxs_x is fixed at sx=1s_x = 1.

The standardised coordinate vector is first projected onto a hidden representation S0∈Rv\mathbf{S}_0 \in\mathbb{R}^v. For residual block I∈{1,…,Nb}I \in\{1, \ldots, N_b\},

SI=SI−1+βIFI(D)(SI−1),zI,0=SI−1,zI,j=ϕ[AI,j(zI,j−1)],j=1,…,D−1,FI(D)(SI−1)=AI,D(zI,D−1),AI,j(y)=WI,jy+bI,j,(53)\begin{aligned} \mathbf{S}_I &= \mathbf{S}_{I-1} + \beta_I \mathcal{F}_I^{(D)}\left(\mathbf{S}_{I-1}\right),\\ \mathbf{z}_{I,0} &= \mathbf{S}_{I-1},\\ \mathbf{z}_{I,j} &= \phi\left[\mathbf{A}_{I,j}\left(\mathbf{z}_{I,j-1}\right)\right], \qquad j = 1, \ldots, D-1,\\ \mathcal{F}_I^{(D)}\left(\mathbf{S}_{I-1}\right) &= \mathbf{A}_{I,D}\left(\mathbf{z}_{I,D-1}\right),\\ \mathbf{A}_{I,j}(\mathbf{y}) &= \mathbf{W}_{I,j}\mathbf{y} + \mathbf{b}_{I,j}, \tag*{(53)} \end{aligned}

where vv is the hidden width, DD is the depth of the residual branch and βI\beta_I is the ReZero scaling parameter. The coefficients βI\beta_I are initialised to zero. An output projection maps the final hidden representation to the two displacement components.

For the numerical examples considered in this work, the architecture parameters are set to

D=3,v=256,Nb=4.(54)D = 3, \qquad v = 256, \qquad N_b = 4. \tag*{(54)}

These values were selected empirically through preliminary numerical experiments. Smaller architectures exhibited greater variability across random initialisations, whereas larger architectures did not provide a clear improvement in the obtained solutions and incurred greater computational cost. The adopted configuration therefore provides a practical balance between robustness across initialisations and computational efficiency.

Activation function

A scaled hyperbolic-tangent activation is employed,

ϕ(z)=tanh⁡(mz),(55)\phi(\mathbf{z}) = \tanh(m\mathbf{z}), \tag*{(55)}

where the operation is applied component-wise and mm is a shared trainable parameter. The coefficient mm controls the slope of the activation and is initialised as m=1m = 1. The same form of adaptive activation was adopted previously for variational brittle-fracture modelling. [21]

Weight regularisation

Weight regularisation was required in previous neural brittle-fracture formulations to control excessive displacement gradients and associated strain magnitudes near localised fracture zones. [20, 21] No comparable instability was observed in the present numerical experiments when the displacement representation was coupled with the local elastoplastic return mapping. No additional parameter-norm regularisation is therefore applied to the displacement network. This choice is specific to the present formulation and numerical examples and is not assumed to hold generally for ductile-fracture problems.

Boundary-condition enforcement

Following the strong-imposition strategy used in variational neural fracture formulations, [20, 21] Dirichlet boundary conditions are imposed exactly through smooth boundary-window functions. For displacement component i=1,…,di = 1, \ldots, d, the admissible field is constructed as

uθ,i(x)=[1−wiF(x)][1−wL(x)]u^θ,i(x)+wL(x)uˉi,(56)u_{\theta,i}(\mathbf{x}) = \left[1-w_i^{\mathrm{F}}(\mathbf{x})\right]\left[1-w^{\mathrm{L}}(\mathbf{x})\right]\hat{u}_{\theta,i}(\mathbf{x}) + w^{\mathrm{L}}(\mathbf{x})\bar{u}_i, \tag*{(56)}

where u^θ,i\hat{u}_{\theta,i} denotes the unconstrained network output. The functions wFw^{\mathrm{F}} and wLw^{\mathrm{L}} equal unity on the corresponding fixed and loaded boundaries and decay smoothly away from them. Their supports are chosen not to overlap. Eq. (56) therefore imposes ui=0u_i=0 on the fixed boundary and ui=uˉiu_i=\bar{u}_i on the loaded boundary without introducing additional penalty terms into the loss.

Phase field

A different representation is adopted for the phase field. In the present ductile model, accumulated equivalent plastic strain can reduce the fracture toughness over a region wider than the final localised crack. Preliminary calculations using the global ReZero representation for cc tended to produce diffuse phase-field profiles within these regions of reduced toughness rather than the required localised fracture zone. Wang et al. 25 similarly employ separate representations for the displacement and phase fields and recommend a radial basis function (RBF) network for the phase field in the continuous formulation. Motivated by these observations, the present work represents the phase field using spatially localised Gaussian RBFs. [34]

The unconstrained phase-field representation is

c^θc(x)=bc+∑k=1Kwkcexp⁡[−∥x−μkη∥22σk2](57)\hat{c}_{\theta_c}(\mathbf{x}) = b_c + \sum_{k=1}^{K} w_k^c \exp\left[-\frac{\lVert\mathbf{x}-\boldsymbol{\mu}_k^\eta\rVert^2}{2\sigma_k^2}\right] \tag*{(57)}

where wkcw_k^c are trainable coefficients, bcb_c is a trainable bias (set to −3.99-3.99) and μkη\boldsymbol{\mu}_k^\eta denotes the centre of the kkth basis function. One initial centre is assigned to each node of the background mesh, such that K=NnodeK=N_{\mathrm{node}}.

The width of each Gaussian is fixed according to the local nodal spacing,

σk=sRBFhk,(58)\sigma_k=s_{\mathrm{RBF}}h_k, \tag*{(58)}

where hkh_k is the mean distance from the initial centre μk0\boldsymbol{\mu}_k^0 to its Nnb=12N_{\mathrm{nb}}=12 nearest neighbouring centres. The dimensionless coefficient sRBFs_{\mathrm{RBF}} controls the overlap between neighbouring basis functions. It is treated as a fixed hyperparameter, with sRBF=1s_{\mathrm{RBF}}=1 used throughout this work. This definition allows the Gaussian widths to reflect local variations in the resolution of the background mesh.

The centres are also allowed to move within a bounded neighbourhood of their initial positions. For spatial component i=1,…,di=1,\ldots,d,

μk,iη=μk,i0+ρktanh⁡(ηk,iρk),ρk=hk2,(59)\mu_{k,i}^\eta=\mu_{k,i}^0+\rho_k\tanh\left(\frac{\eta_{k,i}}{\rho_k}\right),\qquad\rho_k=\frac{h_k}{2}, \tag*{(59)}

where ηk,i\eta_{k,i} is trainable and is initialised to zero. This parameterisation allows each centre to adjust to the evolving phase-field profile while limiting its movement in each coordinate direction to ρk\rho_k from its initial position. The Gaussian widths are kept fixed during optimisation to reduce additional parameter coupling and numerical sensitivity.

Phase-field constraint mapping and initialisation

The unconstrained RBF field is mapped through the piecewise linear soft-bound function adopted previously in variational brittle-fracture modelling. [21]

fc(z)={βc(z+ℓs),z<−ℓs,z2ℓs+12,∣z∣≤ℓs,1+βc(z−ℓs),z>ℓs,cθc(x)=fc(c^θc(x)).(60)f_c(z)= \begin{cases} \beta_c(z+\ell_s), & z<-\ell_s,\\ \frac{z}{2\ell_s}+\frac{1}{2}, & |z|\leq\ell_s,\\ 1+\beta_c(z-\ell_s), & z>\ell_s, \end{cases} \qquad c_{\theta_c}(\mathbf{x})=f_c\left(\hat{c}_{\theta_c}(\mathbf{x})\right). \tag*{(60)}

The parameter ℓs\ell_s defines the half-width of the central linear interval and is unrelated to the material length scale ℓ0\ell_0. The parameter βc\beta_c controls the slope of the two outer branches. The values ℓs=4.0\ell_s=4.0 and βc=10−3\beta_c=10^{-3} are used throughout this work, based on the empirical observations made in previous works. [20, 21] The mapping is therefore a soft constraint rather than an exact projection onto [0,1][0,1]. The small but non-zero slope outside the central interval avoids the zero gradients that would arise from hard clipping.

The RBF coefficients wkcw_k^c and centre parameters ηk,i\eta_{k,i} are initialised to zero. The bias bcb_c is selected close to the upper end of, but within, the central interval so that the initial field represents intact material, c≈1c\approx1, while retaining a non-zero optimisation gradient. No additional nonlinear activation function or parameter-norm regularisation is applied to the RBF representation.

Optimisation strategy

The solution is advanced incrementally under prescribed displacement loading. Transfer learning is utilised by each block for computational efficiency. Each block has its own loss-based stopping criterion, while convergence of the coupled increment is assessed after each complete phase-field–displacement cycle using physics-based convergence criteria.

Transfer learning

The converged trainable parameters from increment n−1n-1 are used to initialise the displacement and phase-field representations at increment nn. The optimiser state associated with each field, including the parameter-wise step sizes, is also transferred between increments. The committed plastic history is carried forward separately at the quadrature points. This warm start exploits the similarity between neighbouring loading states and reduces the optimisation required at successive increments. [20, 21]

At the first loading increment, the ReZero coefficients of the displacement network are initialised to zero, while the RBF representation is initialised to the intact phase-field state described in § 1.3.2.2.

Alternating training scheme

The displacement and phase-field representations are optimised alternately. Both blocks minimise the same loss defined in Eq. (50); no separate energy functional is introduced for either block. Instead, only the parameters of the active representation are varied:

Lnc(θc)=Ln(θufixed,θc),Lnu(θu)=Ln(θu,θcfixed)(61)\mathcal{L}_n^c(\boldsymbol{\theta}_c) = \mathcal{L}_n(\boldsymbol{\theta}_u^{\mathrm{fixed}},\boldsymbol{\theta}_c), \qquad\mathcal{L}_n^u(\boldsymbol{\theta}_u) = \mathcal{L}_n(\boldsymbol{\theta}_u,\boldsymbol{\theta}_c^{\mathrm{fixed}}) \tag*{(61)}

Each coupled cycle begins with the phase-field block. The displacement representation and the current trial constitutive state are held fixed, while θc\boldsymbol{\theta}_c is optimised. The fracture-toughness degradation factor f(ζn⋆)f(\zeta_n^{\star}) is therefore also held fixed during this subproblem, consistent with § 1.1.4.

The phase-field parameters are then frozen and the displacement block is optimised. For each candidate displacement state, the plastic variables are evaluated through the radial-return algorithm. Every return-mapping evaluation starts from the history committed at increment n−1n-1. The resulting trial plastic state is not committed during the optimisation iterations.

After both blocks have been completed, the convergence criteria in § 1.3.4.5 are evaluated. If they are not satisfied, another phase-field–displacement cycle is performed. Once convergence is reached, the final plastic state is committed as the constitutive history for increment n+1n+1.

For computational efficiency, when one representation is frozen, its nodal values are cached and reused until its parameters are updated again. This avoids unnecessary repeated evaluations without modifying the optimisation problem.

Optimiser

Both parameter blocks are optimised using resilient backpropagation (RPROP). [35] RPROP assigns an individual step size to each trainable parameter and determines the update direction from the sign of the corresponding loss gradient. If the gradient retains the same sign between successive iterations, the step size is increased. A change in sign reduces the step size.

RPROP has shown favourable performance in previous variational neural fracture studies. [20, 21] In particular, Manav et al. [20] found RPROP more effective than the alternative optimisation algorithms considered in their study for resolving localised phase-field fracture.

The parameters used in the present work are

Δ0=10−3,η+=1.2,η−=0.5,Δmin⁡=10−9,Δmax⁡=50,(62)\Delta_0 = 10^{-3}, \qquad\eta^{+} = 1.2, \qquad\eta^{-} = 0.5, \qquad\Delta_{\min} = 10^{-9}, \qquad\Delta_{\max} = 50, \tag*{(62)}

where Δ0\Delta_0 is the initial parameter-wise step size, η+\eta^{+} and η−\eta^{-} are the increase and decrease factors, and Δmin⁡\Delta_{\min} and Δmax⁡\Delta_{\max} define the allowable step-size range. Automatic differentiation is applied to the discretely integrated loss to obtain its gradient with respect to the active trainable parameters. This is distinct from the spatial differentiation of the displacement and phase fields, which is performed using the interpolation procedure described in § 1.3.1.

Block-level stopping

Optimisation of an individual displacement or phase-field block is terminated when the relative change in the loss remains below tol⁡L=10−6\operatorname{tol}_{\mathcal{L}} = 10^{-6} over a patience window of Npat=50N_{\mathrm{pat}} = 50 consecutive epochs. A prescribed maximum number of epochs, NepochsN_{\mathrm{epochs}}, provides an additional termination condition.

Physics-based convergence

Loss stagnation within the individual optimisation blocks does not by itself establish convergence of the coupled problem. Convergence of load increment nn is therefore assessed after each complete phase-field–displacement cycle using the phase-field evolution and two displacement-based indicators.

Phase field

Convergence monitoring between successive staggered iterations is commonly employed in variational phase-field fracture formulations. In the present scheme, stabilisation of the phase field is assessed from its change between successive coupled cycles using the maximum norm:

∥c(j)−c(j−1)∥∞<tol⁡c,tol⁡c=1.5×10−3,(63)\left\lVert c^{(j)}-c^{(j-1)}\right\rVert_{\infty}<\operatorname{tol}_{c},\qquad\operatorname{tol}_{c}=1.5\times10^{-3}, \tag*{(63)}

where c(j)c^{(j)} denotes the vector of nodal phase-field values at the end of coupled cycle jj. To avoid terminating on a temporary plateau, this condition must be satisfied for Nplateau,c=3N_{\mathrm{plateau},c}=3 consecutive cycles.

Displacement field

The displacement convergence indicators are constructed from the internal forces associated with the stresses returned by the radial-return algorithm. The assembled internal-force vector is

fint=∑e=1Nel∑g=1Ngpwg∣det⁡Jeg∣BegTσeg,(64)\mathbf{f}^{\mathrm{int}}=\sum_{e=1}^{N_{\mathrm{el}}}\sum_{g=1}^{N_{\mathrm{gp}}}w_g\left\lvert\det\mathbf{J}_{eg}\right\rvert\mathbf{B}_{eg}^{\mathsf{T}}\boldsymbol{\sigma}_{eg}, \tag*{(64)}

where Beg\mathbf{B}_{eg} is the strain–displacement matrix at quadrature point gg of element ee. [31] Since the numerical examples are quasi-static and displacement controlled, no external nodal forces act on the unconstrained degrees of freedom. From the admissible displacement mapping in Eq. (56), variations of the unconstrained network output are weighted by the factor

ωai=(1−waL)(1−waiF),(65)\omega_{ai}=\left(1-w_{a}^{L}\right)\left(1-w_{ai}^{F}\right), \tag*{(65)}

where aa denotes the node index and ii the displacement component. The corresponding weighted equilibrium residual is therefore

ffreeint=fint⊙ω,(66)\mathbf{f}_{\mathrm{free}}^{\mathrm{int}}=\mathbf{f}^{\mathrm{int}}\odot\boldsymbol{\omega}, \tag*{(66)}

where ⊙\odot denotes the Hadamard product.

The first displacement criterion is

∥ffreeint∥∞<ftolint,ftolint=10−5,(67)\left\lVert\mathbf{f}_{\mathrm{free}}^{\mathrm{int}}\right\rVert_{\infty}<f_{\mathrm{tol}}^{\mathrm{int}},\qquad f_{\mathrm{tol}}^{\mathrm{int}}=10^{-5}, \tag*{(67)}

where the force is non-dimensional according to § 1.1.7.

A second indicator monitors stabilisation of the generalised reaction force associated with the prescribed displacement. For a prescribed displacement magnitude UpU_p acting in the unit direction e^\hat{\mathbf{e}}, Eq. (56) gives ∂uai/∂Up=waLe^i\partial u_{ai}/\partial U_p=w_{a}^{L}\hat{e}_i. The reaction work-conjugate to UpU_p is therefore

R=∑ie^i∑a=1NnodewaLfaiint.(68)R=\sum_i\hat{e}_i\sum_{a=1}^{N_{\mathrm{node}}}w_{a}^{L}f_{ai}^{\mathrm{int}}. \tag*{(68)}

This reaction measure is also used to construct the force-displacement response reported in § 2.

Over the most recent Nplateau,uN_{\mathrm{plateau},u} coupled cycles, the reaction-force spread is defined as

Rspread(j)=max⁡k∈WjR(k)−min⁡k∈WjR(k),Wj={j−Nplateau,u+1,…,j}.(69)R_{\mathrm{spread}}^{(j)}=\max_{k\in\mathcal{W}_j}R^{(k)}-\min_{k\in\mathcal{W}_j}R^{(k)},\qquad\mathcal{W}_j=\{j-N_{\mathrm{plateau},u}+1,\ldots,j\}. \tag*{(69)}

The displacement response is considered stationary when

Rspread(j)<RtolRmax⁡,Rtol=5×10−4,Nplateau,u=3,(70)R_{\mathrm{spread}}^{(j)}<R_{\mathrm{tol}}R_{\max},\qquad R_{\mathrm{tol}}=5\times10^{-4},\qquad N_{\mathrm{plateau},u}=3, \tag*{(70)}

where Rmax⁡R_{\max} is the maximum absolute reaction force reached up to the current stage of the simulation. Using the spread over several cycles reduces sensitivity to small oscillations in the reaction force.

The displacement requirement is satisfied when either Eq. (67) or Eq. (70) is met. The complete increment is accepted as converged only when the phase-field criterion in Eq. (63) and one of the displacement criteria are satisfied.

During the coupled optimisation, the toughness-degradation factor f(ζn⋆)f(\zeta_{n}^{\star}) is evaluated from the current trial constitutive state rather than being lagged to the value from increment n−1n-1. Its current value therefore enters the forward evaluation of the fracture-energy contribution. However, the dependence of ff on ζn⋆\zeta_{n}^{\star} is excluded from automatic differentiation, so gradients are not propagated through the degradation law or the constitutive history. During the phase-field block, the most recently evaluated f(ζn⋆)f(\zeta_n^{\star}) is held fixed. This treatment is consistent with the assumption ∂f/∂ζ=0\partial f/\partial\zeta= 0 adopted by Yin and Kaliske, [10], while retaining the current spatially varying degraded fracture toughness.

A maximum of Nmax⁡=500N_{\max}=500 coupled optimisation cycles is imposed to prevent indefinite iteration. If this limit is reached before the physics-based criteria are satisfied, the current state is retained and the simulation proceeds to the next loading increment. Reaching Nmax⁡N_{\max} is therefore an algorithmic termination condition and is not interpreted as convergence of the coupled increment.

Numerical results

Five ductile fracture specimens are considered to assess the proposed DEM formulation. The DEM predictions are compared with reference finite element method (FEM) solutions in terms of the force-displacement response, energy evolution, accumulated equivalent plastic strain and phase-field distributions. Parameters common to all DEM simulations are summarised in Table 1. Specimen-specific geometries, material parameters and loading schedules are reported in the corresponding subsections.

CategoryParameterNotationValue
Displacement representationResidual-branch depthDD3
Displacement representationHidden widthν\nu256
Displacement representationNumber of ReZero blocksNbN_{\mathrm{b}}4
Displacement representationInitial ReZero scalingβI\beta_{\mathrm{I}}0
Displacement representationCoordinate scalingsxs_x1
Displacement representationInitial activation scalingmm1
Phase-field representationNumber of nearest neighboursNnbN_{\mathrm{nb}}12
Phase-field representationRBF overlap coefficientsRBFs_{\mathrm{RBF}}1
Phase-field representationSoft-bound half-widthℓs\ell_s4.0
Phase-field representationOuter-branch slopeβc\beta_c10−310^{-3}
Phase-field representationIrreversibility tolerancetolirr\mathrm{tol}_{\mathrm{irr}}0.02
RPROP optimiserInitial step sizeΔ0\Delta_010−310^{-3}
RPROP optimiserStep-increase factorη+\eta^{+}1.2
RPROP optimiserStep-decrease factorη−\eta^{-}0.5
RPROP optimiserMinimum step sizeΔmin⁡\Delta_{\min}10−910^{-9}
RPROP optimiserMaximum step sizeΔmax⁡\Delta_{\max}50
Block stoppingRelative loss-improvement tolerancetolL\mathrm{tol}_{\mathcal{L}}10−610^{-6}
Block stoppingPatience windowNpatN_{\mathrm{pat}}50
Block stoppingMaximum number of epochsNepochsN_{\mathrm{epochs}}250
Coupled convergencePhase-field tolerancetolc\mathrm{tol}_c1.5×10−31.5 \times10^{-3}
Coupled convergencePhase-field plateau lengthNplateau,cN_{\mathrm{plateau},c}3
Coupled convergenceInternal-force toleranceftolf_{\mathrm{tol}}10−510^{-5}
Coupled convergenceReaction-force toleranceRtolR_{\mathrm{tol}}5×10−45 \times10^{-4}
Coupled convergenceReaction plateau lengthNplateau,uN_{\mathrm{plateau},u}3
Coupled convergenceMaximum coupled cyclesNmax⁡N_{\max}500

Table 1. Numerical parameters adopted for the DEM field representations, optimisation and convergence criteria. All numbers are dimensionless.

The FEM reference solutions are obtained using a staggered iterative scheme. Convergence of each staggered cycle is assessed using the normalised energy-iteration measure γ\gamma, defined in Eq. (43) of Ambati et al. 36`` and subsequently used in their ductile-fracture formulation. [6] A convergence tolerance of γ≤5∘\gamma\le5^\circ is adopted in the present calculations.

The FEM computations were performed on an Intel Core i9-14900K CPU using the MUMPS sparse direct solver with 20 threads, whereas the DEM simulations were carried out on NVIDIA A100 40 GB PCIe GPUs.

All specimens are analysed in two dimensions under plane-strain conditions. For computational efficiency, the prescribed displacement history is divided into coarse and fine loading stages. Coarser increments are used during the predominantly elastic regime, followed by finer increments during material softening and rupture. The displacement ranges and corresponding increment counts are specified separately for each specimen.

In DEM, each specimen is simulated using eight random seeds to assess sensitivity to random initialisation. The displacement-step-size study in § 2.3.3 uses 6 seeds for each increment count. Where ensemble curves are shown, they represent the mean response, with shaded bands spanning ±1\pm1 standard deviation.

The code and data used to generate the FEM and DEM results are available through the associated Zenodo repository. [37]

Ductile fracture of asymmetrically notched and symmetrically notched specimens

Three notched specimens with the same material parameters are considered to investigate the influence of geometry on plastic localisation and fracture evolution. The geometries and boundary conditions are shown in Figure 4, and the corresponding material parameters are given in Table 2. The asymmetrically notched specimen is used to assess crack propagation and linking between offset notches. The two symmetric specimens, with notch radii r=2.5r=2.5 mm and r=5r=5 mm, are used to examine the effect of notch radius on the localisation pattern and subsequent rupture response.

ParameterNotationValueUnit
Young’s modulusEE72 624.3MPa
Poisson’s ratioν\nu0.3311–
Yield stressσy\sigma_y345MPa
Hardening modulusHH250MPa
Undegraded fracture toughnessGc0\mathcal{G}_c^095N/mm
Phase-field length scalel0l_00.3mm
Toughness-degradation intervalξ\xi0.4–
Residual toughness fractionbb1×10−61 \times10^{-6}–
Degradation thresholdζcr\zeta^{\mathrm{cr}}0.02–

Table 2. Material parameters adopted for the asymmetrically and symmetrically notched specimens.

Figure 4. Geometry and boundary conditions for the notched specimens: (a) asymmetrically notched specimen, (b) symmetrically notched specimen with r=2.5r = 2.5 mm, and (c) symmetrically notched specimen with r=5r = 5 mm.

Asymmetrically notched specimen

The FEM reference used 10 000 load increments. The first 100 increments covered loading up to Up=0.9U_p=0.9 mm, while the remaining 9900 reached Up=1.65U_p=1.65 mm. The computation required 6.23 h on the hardware described above. Each DEM run used 230 loading increments: 30 increments up to Up=0.9U_p=0.9 mm and 200 increments up to Up=1.65U_p=1.65 mm. Across the eight seeds, the mean DEM wall-clock time was (16.5±1.1)(16.5\pm1.1) h, with a range of 15.5–18.6 h. Because the FEM and DEM computations were performed on different hardware, these wall-clock times are reported as computational costs on the stated hardware rather than as a hardware-normalised efficiency comparison. No DEM run reached the 500-cycle limit before rupture. Seed 8 reached the limit once, during the rupture increment. The limit therefore did not affect the pre-rupture response.

The global responses are compared in Figure 5. The DEM ensemble closely reproduces the FEM force-displacement curve throughout the stable loading regime. It captures the peak load and the gradual post-peak reduction with little variation among the seeds. The elastic, fracture, and plastic energy contributions also remain close to the FEM reference over most of the loading history.

Figure 5. Asymmetrically notched specimen: global response obtained with the proposed DEM and the FEM reference: (a) force-displacement response; (b) elastic energy Ψe\Psi^{\mathrm{e}}; (c) fracture energy Ψf\Psi^{\mathrm{f}}; (d) plastic energy Ψp\Psi^{\mathrm{p}}; and (e) relative difference in total energy, (ΨDEMtot−ΨFEMtot)/ΨFEMtot(\Psi_{\mathrm{DEM}}^{\mathrm{tot} - \Psi_{\mathrm{FEM}}^{\mathrm{tot}})/\Psi_{\mathrm{FEM}}^{\mathrm{tot}}}, where Ψtot=Ψe+Ψf+Ψp\Psi^{\mathrm{tot}} = \Psi^{\mathrm{e}} + \Psi^{\mathrm{f}} + \Psi^{\mathrm{p}}. All quantities are plotted against the prescribed displacement UpU_p. The DEM curves represent the mean over eight seeds, and the shaded bands span ±1\pm1 standard deviation.

Rupture occurs at a slightly smaller prescribed displacement in the DEM simulations. This difference appears only during the final unstable stage and may partly reflect the coarser DEM load discretisation. The plastic history and the associated toughness degradation are updated incrementally. Consequently, the predicted onset of an abrupt terminal event can be sensitive to the number and sequence of load increments. This interpretation is consistent with the load-step sensitivity examined in § 2.3.3.

The earlier rupture is accompanied by a slightly lower final plastic contribution and an earlier increase in fracture energy. The magnitude of the total-energy difference remains small throughout stable loading, increases briefly during rupture, and then decreases after complete separation. The discrepancy is therefore concentrated around the terminal instability and does not affect the close pre-rupture agreement.

The fields at the respective rupture states are shown in Figure 6. Both methods predict a curved crack joining the two notch roots, with similar overall trajectories. This crack-linking mechanism is consistent with that reported by Ambati et al. [6] The FEM crack has somewhat greater local variation in the central linking region, whereas the DEM path is smoother. The distributions of accumulated equivalent plastic strain ζ\zeta are also similar. The slightly lower DEM accumulation in the central region may be associated with its earlier rupture and coarser load-increment history. These local differences are concentrated in the central linking region and do not change the overall fracture mechanism. The small spread across the 8 seeds further indicates that the predicted global response is not strongly affected by network initialisation.

Figure 6. Asymmetrically notched specimen. Phase-field and accumulated equivalent plastic strain distributions at the respective rupture states. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM. The DEM fields correspond to seed 8.

Symmetrically-notched specimen r=2.5r = 2.5 mm

The FEM reference used 10 000 load increments. The first 100 increments covered the loading range up to Up=0.4U_p = 0.4 mm, followed by 9900 increments up to Up=3.6U_p = 3.6 mm. The computation required 6.08 h. Each DEM simulation used 230 load increments, with 30 increments up to Up=0.4U_p = 0.4 mm and 200 increments up to Up=3.6U_p = 3.6 mm. Across the eight seeds, the mean DEM wall-clock time was (28.0±3.7)(28.0 \pm3.7) h, with a range of 23–34.3 h. This was the most computationally demanding of the 5 benchmark cases because of the extended softening regime, during which approximately 15–30 coupled cycles were required per increment. Seeds 2 and 7 each reached the 500-cycle limit once, during their respective rupture increments. No seed reached this limit before rupture.

The global responses are compared in Figure 7. The DEM reproduces the loading branch and the extended softening response closely, but predicts final rupture at a somewhat larger prescribed displacement than the FEM reference. The plastic-energy evolution remains in close agreement, while the elastic and fracture-energy responses exhibit a corresponding shift as rupture is approached. The relative difference in total energy becomes increasingly negative beyond approximately Up=1.4U_p = 1.4 mm and changes sign after the FEM rupture event, reaching a small positive difference close to the final DEM rupture. The spread across the eight DEM seeds remains narrow over most of the loading history. The delayed rupture is therefore reproducible across the random initialisations considered here, rather than being associated with an isolated optimisation trajectory. Given the incremental and path-dependent constitutive evolution, the substantially coarser DEM loading discretisation may also contribute to the small shift in the onset of the final instability.

Figure 7. Symmetrically notched r=2.5r = 2.5 mm specimen. Global response obtained with the proposed DEM and the FEM reference: (a) force-displacement response; (b) elastic energy Ψe\Psi^{\mathrm{e}}; (c) fracture energy Ψf\Psi^{\mathrm{f}}; and (d) plastic energy Ψp\Psi^{\mathrm{p}}; (e) relative difference in total energy, 100(ΨDEMtot−ΨFEMtot)/ΨFEMtot100(\Psi_{\mathrm{DEM}}^{\mathrm{tot}} - \Psi_{\mathrm{FEM}}^{\mathrm{tot}})/\Psi_{\mathrm{FEM}}^{\mathrm{tot}}, where Ψtot=Ψe+Ψf+Ψp\Psi^{\mathrm{tot}} = \Psi^{\mathrm{e}} + \Psi^{\mathrm{f}} + \Psi^{\mathrm{p}}. All quantities are plotted against the prescribed displacement UpU_p. Shaded bands span ±1\pm1 standard deviation over the eight DEM seeds.

The phase-field and accumulated equivalent plastic strain distributions are shown in Figure 8 at Up=3.407U_p = 3.407 mm and in Figure 9 after rupture. Before rupture, the two methods predict closely comparable crack initiation from the notch roots and similar plastic-localisation patterns. This behaviour is also consistent with the failure mechanism reported for the smaller-radius notched specimen by Ambati et al. [6]. The main local difference is the narrower phase-field band obtained with the DEM. Since the RBF widths are defined by σk=sRBFhk\sigma_k = s_{\mathrm{RBF}} h_k, the locally refined nodal spacing reduces the individual kernel widths for a fixed sRBFs_{\mathrm{RBF}} and may contribute to this sharper representation. The same value of sRBFs_{\mathrm{RBF}} is retained for all specimens, avoiding specimen-specific tuning. The physical phase-field length scale l0l_0 is unchanged.

Figure 8. Symmetrically notched r=2.5r = 2.5 mm specimen. Phase-field and accumulated equivalent plastic strain distributions in the ligament region at Up=3.407U_p = 3.407 mm. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM. The DEM fields correspond to seed 1.

Figure 9. Symmetrically notched r=2.5r = 2.5 mm specimen. Phase-field and accumulated equivalent plastic strain distributions in the ligament region after rupture. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM.

After rupture, both methods predict the same overall inward crack-linking mechanism, although the final transition is resolved differently. The FEM solution develops the final connection over two load increments, whereas the DEM completes the corresponding linking event within a single increment. The additional intermediate load state in the FEM solution is accompanied by a localised region of higher accumulated equivalent plastic strain.

Symmetrically-notched specimen r=5r = 5 mm

The FEM reference used 2000 load increments. The first 100 increments covered the loading range up to Up=1.6U_p = 1.6 mm, followed by 1900 increments up to Up=1.8U_p = 1.8 mm. The computation required 0.98 h. Each DEM simulation used 130 load increments, with 30 increments up to Up=1.6U_p = 1.6 mm and a further 100 increments up to Up=1.8U_p = 1.8 mm.

Five of the eight DEM runs completed in (2.3±0.1)(2.3 \pm0.1) h. Seeds 6, 7 and 8 required 13.4 h, 8.2 h and 13.9 h, respectively. For these runs, the reaction force became very small after complete separation and increasingly sensitive to numerical fluctuations. As a result, the reaction-force stabilisation criterion was no longer satisfied within the prescribed cycle limit, and the affected post-rupture increments reached the maximum of 500 coupled cycles. No seed reached this limit before rupture. When all eight runs are included, the mean DEM wall-clock time is (5.9±5.2)(5.9 \pm5.2) h. The large variation in runtime is therefore associated with the post-rupture termination criterion rather than with the prediction of the loading response or onset of rupture.

The global responses are compared in Figure 10. The DEM and FEM solutions remain closely aligned throughout loading and softening, apart from a small initial difference associated with the different load discretisations. The main discrepancy develops during the terminal fracture event and is concentrated in the fracture-energy contribution. The elastic and plastic energies remain closely matched, whereas the two methods select different final crack-linking branches. The resulting DEM state has a total energy approximately 1.3 % below the FEM reference.

Figure 10. Symmetrically notched r=5r = 5 mm specimen. Global response obtained with the proposed DEM and the FEM reference: (a) force–displacement response; (b) elastic energy Ψe\Psi^{e}; (c) fracture energy Ψf\Psi^{f}; and (d) plastic energy Ψp\Psi^{p}; (e) relative difference in total energy, 100(ΨDEMtot−ΨFEMtot)/ΨFEMtot100(\Psi^{\mathrm{tot}}_{\mathrm{DEM}}-\Psi^{\mathrm{tot}}_{\mathrm{FEM}})/\Psi^{\mathrm{tot}}_{\mathrm{FEM}}, where Ψtot=Ψe+Ψf+Ψp\Psi^{\mathrm{tot}} = \Psi^{e}+\Psi^{f}+\Psi^{p}. All quantities are plotted against the prescribed displacement UpU_{p}. Shaded bands span ±1\pm1 standard deviation over the eight DEM seeds.

All eight DEM seeds select the same terminal fracture branch. The difference is therefore reproducible with respect to random initialisation and is not associated with an isolated optimisation trajectory. We interpret this as the selection of an alternative fracture branch during the abrupt final instability, rather than as evidence that either discrete solution represents a unique global energy minimum. Before this instability, the two approaches remain in close agreement.

The phase-field and accumulated equivalent plastic strain distributions immediately before and after rupture are shown in Figure 11 and Figure 12, respectively. Before rupture, both methods predict closely comparable localisation patterns. The accumulated equivalent plastic strain changes only slightly across the final transition, while the phase field undergoes the dominant change associated with complete crack linking. This suggests that the plastic localisation is largely established before complete separation. The subsequent failure is therefore dominated by rapid phase-field evolution within the plastically affected ligament. The abrupt failure and inclined crack development are also qualitatively consistent with the behaviour previously reported for the r=5r = 5 mm notched benchmark.[6]

Figure 11. Symmetrically notched r=5r = 5 mm specimen. Phase-field and accumulated equivalent plastic strain distributions in the ligament region immediately before rupture. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM. A common colour scale is used for each field to facilitate direct comparison between the two methods.

Figure 12. Symmetrically notched r=5r = 5 mm specimen. Phase-field and accumulated equivalent plastic strain distributions in the ligament region immediately after rupture. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM.

Ductile fracture of a V-notch specimen

The V-notch specimen is used to assess whether the proposed DEM framework can reproduce symmetric crack propagation from opposing notch roots and subsequent linking across the central ligament. The predicted force-displacement response is also compared with the experimental measurements reported by Li et al. [38] The specimen geometry follows the V-notch benchmark considered by Miehe et al. [9] The geometry and boundary conditions are shown in Figure 13a. The material parameters adopted in the present simulations are listed in Table 3.

ParameterNotationValueUnit
Young’s modulusEE68 899.4MPa
Poisson’s ratioν\nu0.33–
Yield stressσy\sigma_{\mathrm{y}}530MPa
Hardening modulusHH561MPa
Undegraded fracture toughnessGc0\mathcal{G}_{\mathrm{c}}^{0}90N/mm
Phase-field length scalel0l_{0}0.2mm
Toughness-degradation intervalξ\xi0.2–
Residual toughness fractionbb5×10−35 \times10^{-3}–
Degradation thresholdζcr\zeta^{\mathrm{cr}}0–

Table 3. Material parameters adopted for the V-notch specimen.

Figure 13. Geometry and boundary conditions for (a) the V-notch specimen and (b) the diagonally notched specimen.

The FEM reference used 10 000 load increments. The first 100 increments covered the range up to Up=0.2U_p = 0.2 mm, and the remaining 9900 increments covered the range up to Up=0.48U_p = 0.48 mm. The computation required 5.79 h. Each DEM run used 230 load increments. Of these, 30 increments covered the range up to Up=0.2U_p = 0.2 mm and 200 covered the range up to Up=0.48U_p = 0.48 mm.

Across the eight seeds, the mean DEM wall-clock time was (19.0±2.5)(19.0 \pm2.5) h, with a range of 16.0–22.8 h. No run reached the maximum of 500 coupled cycles at any increment.

The global responses are compared in Figure 14. The DEM and FEM force-displacement curves agree closely through the loading branch, peak load and most of the softening regime. Both numerical solutions also follow the experimental measurements well over this range. Near final rupture, the experiment shows a more gradual loss of load-carrying capacity, whereas the numerical responses drop more sharply. Overall, the DEM captures the main features of the measured response, with the largest difference occurring only during the final stage of failure.

Figure 14. V-notch specimen. Global response obtained with the proposed DEM and the FEM reference: (a) force-displacement response, including the experimental measurements of Li et al. [38]; (b) elastic energy Ψe\Psi^{e}; (c) fracture energy Ψf\Psi^{f}; and (d) plastic energy Ψp\Psi^{p}; (e) relative difference in total energy, 100(ΨDEMtot−ΨFEMtot)/ΨFEMtot100(\Psi_{\mathrm{DEM}}^{\mathrm{tot}}-\Psi_{\mathrm{FEM}}^{\mathrm{tot}})/\Psi_{\mathrm{FEM}}^{\mathrm{tot}}, where Ψtot=Ψe+Ψf+Ψp\Psi^{\mathrm{tot}}=\Psi^{e}+\Psi^{f}+\Psi^{p}. All quantities are plotted against the prescribed displacement UpU_{p}. The DEM curves represent the mean over eight seeds, and the shaded bands span ±1\pm1 standard deviation.

The elastic-energy predictions remain in close agreement throughout the loading history. Near rupture, the DEM shows a somewhat larger fracture-energy contribution and a correspondingly smaller plastic contribution than the FEM reference. These differences partly offset one another, so the total-energy response remains close over most of the simulation. The largest deviation is confined to the short interval associated with final rupture. The narrow ensemble bands before this point also indicate that the stable response is only weakly affected by random initialisation. The increased spread is concentrated around the terminal instability, where the solution becomes more sensitive to the precise evolution of the fracture process.

The phase-field and accumulated equivalent plastic strain distributions at rupture are shown in Figure 15. Both methods predict fracture localisation within the central ligament and a similar overall linking pattern. The FEM solution shows several small, localised degraded regions along the fracture zone, whereas the DEM produces a more continuous crack band. A similar difference is visible in the accumulated equivalent plastic strain field. The FEM solution exhibits stronger local concentrations in these regions, while the DEM distribution is smoother overall.

Phase-field and accumulated equivalent plastic strain distributions for the V-notch specimen, shown for FEM and DEM

Figure 15. V-notch specimen. Phase-field and accumulated equivalent plastic strain distributions at rupture. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM.

Ductile fracture of a diagonally-notched specimen

The diagonally notched specimen is used to assess the ability of the proposed DEM to reproduce a strongly curved crack path between two offset notches. The configuration requires progressive crack reorientation across the ligament under coupled plastic localisation and toughness degradation. Since the trajectory is not prescribed, this example provides a demanding test of the spatial field representations. The geometry and boundary conditions are shown in Figure 13b, and the material parameters are given in Table 4.

ParameterNotationValueUnit
Young’s modulusEE195 000MPa
Poisson’s ratioν\nu0.30–
Yield stressσy\sigma_{y}443MPa
Hardening modulusHH300MPa
Undegraded fracture toughnessGc0\mathcal{G}_{c}^{0}209N/mm
Phase-field length scalel0l_{0}0.2mm
Toughness-degradation intervalξ\xi0.35–
Residual toughness fractionbb1×10−21 \times10^{-2}–
Degradation thresholdζcr\zeta^{\mathrm{cr}}0–

Table 4. Material parameters adopted for the diagonally notched specimen.

The FEM reference used 10 000 load increments. The first 100 increments covered loading up to Up=0.23U_{\mathrm{p}} = 0.23 mm, while the remaining 9900 increments reached Up=0.53U_{\mathrm{p}} = 0.53 mm. The computation required 6.79 h. Each DEM run used 230 load increments. Of these, 30 increments covered loading up to Up=0.23U_{\mathrm{p}} = 0.23 mm and 200 reached Up=0.53U_{\mathrm{p}} = 0.53 mm. Across the eight seeds, the DEM wall-clock time was (22.8±1.7)(22.8 \pm1.7) h, with a range of 19.6–24.3 h. Seeds 3, 4 and 7 each reached the 500-cycle limit once, during their rupture increments. No seed reached the limit before rupture. The cycle limit was therefore confined to the terminal event and did not affect the stable pre-rupture response.

The global responses are compared in Figure 16. The DEM ensemble closely reproduces the FEM force-displacement curve through loading, peak load and the subsequent softening regime. Rupture occurs at a slightly smaller prescribed displacement in DEM. The difference is limited to the terminal instability, while the peak load and preceding response remain closely matched.

Figure 16. Diagonally notched specimen. Global response obtained with the proposed DEM and the FEM reference: (a) force-displacement response; (b) elastic energy Ψe\Psi^{e}; (c) fracture energy Ψf\Psi^{f}; (d) plastic energy Ψp\Psi^{p}; and (e) relative difference in total energy, 100(ΨDEMtot−ΨFEMtot)/ΨFEMtot100(\Psi^{\mathrm{tot}}_{\mathrm{DEM}}-\Psi^{\mathrm{tot}}_{\mathrm{FEM}})/\Psi^{\mathrm{tot}}_{\mathrm{FEM}}, where Ψtot=Ψe+Ψf+Ψp\Psi^{\mathrm{tot}}=\Psi^{e}+\Psi^{f}+\Psi^{p}. All quantities are plotted against the prescribed displacement UpU_p. The DEM curves represent the mean over eight seeds, and the shaded bands span ±1\pm1 standard deviation.

The narrow ensemble band over most of the loading history also indicates limited sensitivity to random initialisation before rupture.

The elastic and plastic contributions remain close to the FEM reference. The most noticeable difference occurs in the fracture energy. Once crack growth begins, the DEM predicts a somewhat larger fracture contribution and a slightly smaller terminal plastic contribution. The relative total-energy difference remains within a few per cent throughout the displayed loading range. It is slightly positive over most of the pre-rupture response and becomes modestly negative following the earlier DEM rupture. The differences between the individual energy components therefore partly compensate for one another.

The larger DEM fracture contribution accompanies a marginally more diffuse representation of the phase-field band. The substantially coarser DEM loading schedule may also contribute to the difference in energy partitioning. The load-step study in Section 4.4 shows that the fracture and plastic energy contributions are sensitive to increment size, although it does not establish the cause of the discrepancy in the present specimen.

The phase-field and accumulated equivalent plastic strain distributions at rupture are shown in Figure 17. Both methods predict the same curved crack connecting the two offset notch roots. The linking locations and overall trajectory agree closely. The DEM phase-field band is marginally more diffuse, while its accumulated equivalent plastic strain is slightly lower in the central linking region. These differences are local and do not alter the predicted fracture mechanism. The close agreement is particularly relevant because the curved trajectory emerges from the coupled solution and is not imposed in advance.

Figure 17. Diagonally notched specimen. Phase-field and accumulated equivalent plastic strain distributions at rupture. The phase field cc is shown for (a) FEM and (b) DEM, while the accumulated equivalent plastic strain ζ\zeta is shown for (c) FEM and (d) DEM. The DEM fields correspond to seed 8.

Displacement-step-size convergence study

A displacement-step-size study is performed to assess the sensitivity of the DEM solution to the loading discretisation. The V-notch specimen from § 2.3.1 is selected because its gradual softening is followed by a rapid terminal rupture. The first 30 increments, covering the initial loading response up to Up=0.2U_p = 0.2 mm, are retained in every simulation. The subsequent range up to Up=0.48U_p = 0.48 mm is discretised using N∈{50,100,150,200,250,300}N \in\{50, 100, 150, 200, 250, 300\} increments. 6 random seeds are considered for each

value of NN. This arrangement isolates the influence of the refined-stage discretisation while preserving the same initial loading history.

The results are summarised in Figure 18. Since the plastic internal variables and the associated toughness degradation are updated incrementally, the terminal solution can depend on the sequence of prescribed displacement states. The effect is most apparent during rapid crack linking. Nevertheless, the mean rupture displacement becomes largely insensitive to further refinement once a sufficient number of increments is used.

Figure 18. Displacement-step-size convergence of the DEM for the V-notch specimen. The refined loading stage from Up=0.2U_p = 0.2 mm to Up=0.48U_p = 0.48 mm is discretised using NN increments, with 6 seeds considered for each value of NN. (a) Rupture displacement, shown as the mean ±1\pm1 standard deviation, together with the FEM reference; (b) signed relative error in rupture displacement, 100(UpDEM−UpFEM)/UpFEM100(U_p^{\mathrm{DEM}} - U_p^{\mathrm{FEM}})/U_p^{\mathrm{FEM}}, with whiskers spanning the minimum and maximum values across the 6 seeds; (c) wall-clock time per seed, shown as the mean ±1\pm1 standard deviation, together with the FEM runtime; and (d) DEM-FEM differences in the final elastic, fracture, plastic and total energies.

The mean rupture displacement changes noticeably between N=50N = 50 and N=100N = 100, but stabilises from approximately N=150N = 150 onwards. For N≥150N \ge150, the mean rupture displacement remains within approximately 0.2 % of the FEM reference. The largest error among all individual seeds and discretisations remains below approximately 1.8 %. The seed-to-seed range does not decrease monotonically with NN. The mean and the ensemble spread should therefore both be considered when assessing convergence of the abrupt terminal event.

The DEM wall-clock time increases approximately linearly with the number of refined-stage increments. The mean time rises from approximately 4 h at N=50N = 50 to approximately 24 h at N=300N = 300. Each additional increment requires a separate coupled optimisation and constitutive update, which is consistent with the observed scaling. The FEM wall-clock time is included in Figure 18c for context. However, the FEM and DEM calculations were performed on different hardware. The comparison should therefore not be interpreted as a hardware-normalised measure of relative computational efficiency.

The magnitudes of the final energy differences in Figure 18d generally decrease towards zero as the loading discretisation is refined. At the coarsest discretisation, the DEM predicts lower elastic and plastic contributions and a higher fracture contribution than FEM. This pattern is consistent with coarse increments resolving fewer intermediate coupled states during crack linking and with the lower plastic accumulation observed before complete fracture.

As NN increases, the elastic, fracture and plastic contributions all approach their FEM values. The magnitude of the final total-energy difference decreases from approximately 0.07 J at N=50N = 50 to less than 0.01 J at N=300N = 300. The fracture and plastic differences have opposite signs and therefore partly compensate. Consequently, agreement in total energy alone provides a less stringent measure than agreement in the individual components.

The N=200N = 200 discretisation used for the principal V-notch comparison lies within the regime in which the mean rupture displacement has stabilised. Its mean rupture-displacement error is approximately 0.1 %, and its final total-energy difference is small. It therefore provides a practical balance between loading resolution and computational cost for this example. The study does not establish a universal optimal increment count, which will depend on the loading history, fracture instability and required accuracy.

Conclusions

The numerical results show that the proposed DEM reproduces the principal features of the FEM solutions across the five ductile-fracture benchmarks considered. The force-displacement responses, energy evolution, crack trajectories and accumulated equivalent plastic strain fields generally remain in close agreement throughout stable loading and softening. The largest differences are concentrated around abrupt terminal fracture, where the incremental plastic history, toughness degradation and crack-linking process make the solution more sensitive to the loading discretisation. The V-notch step-size study confirms this behaviour: for N≥150N \ge150, the mean rupture displacement remains within approximately 0.2 % of the FEM reference, while the largest error across the individual seeds and discretisations considered remains below approximately 1.8 %. Across the random initialisations considered, seed-to-seed variability is generally small during the stable response and becomes more pronounced near abrupt fracture events.

Computational cost remains a significant limitation of the present implementation. The displacement-step-size study shows an approximately linear increase in wall-clock time with the number of refined-stage loading increments because every additional increment requires a new coupled optimisation and constitutive update. Extended softening may also require many alternating phase-field-displacement cycles, while the reaction-force stabilisation criterion can become ineffective after complete separation when the reaction approaches numerical noise. Future work should therefore investigate adaptive displacement stepping, more efficient coupled stopping criteria and alternative optimisation strategies that reduce the number of field evaluations required per increment. Improved use of GPU parallelism and more efficient training algorithms will also be important if the method is to become computationally competitive for larger problems.

Further development of the trainable field representations is also required. Preliminary architecture studies showed that simply increasing the size of the displacement network did not provide a clear improvement relative to the additional computational cost. Improving equilibrium-residual convergence may therefore require enriched or adaptive displacement representations rather than larger networks alone. For the phase field, the symmetrically notched r=2.5r = 2.5 mm example produced a narrower localisation band. than the FEM reference. Since the Gaussian width is defined through σk=sRBFhk\sigma_k = s_{\mathrm{RBF}}h_k, future work should examine the selection of sRBFs_{\mathrm{RBF}}, adaptive or trainable RBF widths, and the interaction between RBF resolution and the physical phase-field length scale. The present framework also provides a basis for incorporating more complex constitutive physics. Natural extensions include gradient-enhanced plasticity, finite-strain elastoplasticity and thermomechanically coupled ductile fracture formulations. [9, 12] Such developments would provide a more demanding assessment of the advantages and limitations of the proposed treatment of local constitutive history.

CREDIT AUTHORSHIP CONTRIBUTION STATEMENT

N. Plungè: Conceptualisation, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft. P. Brommer: Conceptualisation, Funding acquisition, Methodology, Supervision, Writing - review & editing. R. S. Edwards: Conceptualisation, Methodology, Supervision, Writing - review & editing. E. G. Kakouris: Conceptualisation, Funding acquisition, Methodology, Supervision, Writing - review & editing.

DECLARATION OF COMPETING INTEREST

The authors declare that they have no competing financial interests or personal relationships that could have appeared to influence the research presented in this article.

DATA AVAILABILITY STATEMENT

The input data supporting this study, together with the code used to produce the FEM and DEM results, are made available via the associated Zenodo repository. [37] The main output data are reported in the figures, while additional data may be made available from the corresponding author upon reasonable request.

ACKNOWLEDGEMENTS

The authors acknowledge the computing resources provided by the Scientific Computing Research Technology Platform at the University of Warwick. GPU computations were undertaken on the Sulis Tier 2 HPC platform, hosted by the University of Warwick Scientific Computing Research Technology Platform. Sulis is funded by Engineering and Physical Sciences Research Council (EPSRC) Grant EP/T022108/1 and the HPC Midlands+ consortium. The first author is supported by a PhD studentship funded by the EPSRC Centre for Doctoral Training in Modelling of Heterogeneous Systems, HetSys, Grant EP/S022848/1, project reference 2886009.

References

  1. [1]Benzerga AA, Leblond JB, Needleman A, Tvergaard V. Ductile failure modeling. International Journal of Fracture 2016; 201(1): 29–80. doi: 10.1007/s10704-016-0142-6DOI
  2. [2]Noell PJ, Sills RB, Benzerga AA, Boyce BL. Void nucleation during ductile rupture of metals: A review. Progress in Materials Science 2023; 135: 101085. doi: 10.1016/j.pmatsci.2023.101085DOI
  3. [3]Besson J. Continuum models of ductile fracture: A review. International Journal of Damage Mechanics 2010; 19(1): 3–52. doi: 10.1177/1056789509103482DOI
  4. [4]Noell PJ, Carroll JD, Boyce BL. The mechanisms of ductile rupture. Acta Materialia 2018; 161: 83–98. doi: 10.1016/j.actamat.2018.09.006DOI
  5. [5]Bao Y, Wierzbicki T. On fracture locus in the equivalent strain and stress triaxiality space. International Journal of Mechanical Sciences 2004; 46(1): 81–98. doi: 10.1016/j.ijmecsci.2004.02.006DOI
  6. [6]Ambati M, Gerasimov T, De Lorenzis L. Phase-field modeling of ductile fracture. Computational Mechanics 2015; 55: 1017–1040. doi: 10.1007/s00466-015-1151-4DOI
  7. [7]Borden MJ, Hughes TJR, Landis CM, Anvari A, Lee IJ. A phase-field fracture formulation for fracture in ductile materials: Finite deformation balance law derivation, plastic degradation, and stress triaxiality effects. Computer Methods in Applied Mechanics and Engineering 2016; 312: 130–166. doi: 10.1016/j.cma.2016.09.005
  8. [8]Zhang H, Peng H, Pei XY, et al. Phase-field modeling of coupled spall and adiabatic shear banding and simulation of complex cracks in ductile metals. Journal of the Mechanics and Physics of Solids 2023; 172: 105186. doi: 10.1016/j.jmps.2022.105186DOI
  9. [9]Miehe C, Aldakheel F, Raina A. Phase field modeling of ductile fracture at finite strains: A variational gradient-extended plasticity-damage theory. International Journal of Plasticity 2016; 84: 1–32. doi: 10.1016/j.ijplas.2016.04.011DOI
  10. [10]Yin B, Kaliske M. A ductile phase-field model based on degrading the fracture toughness: Theory and implementation at small strain. Computer Methods in Applied Mechanics and Engineering 2020; 366: 113068. doi: 10.1016/j.cma.2020.113068DOI
  11. [11]Han J, Matsubara S, Moriguchi S, Kaliske M, Terada K. Crack phase-field model equipped with plastic driving force and degrading fracture toughness for ductile fracture simulation. Computational Mechanics 2022; 69: 151–175. doi: 10.1007/s00466-021-02087-1DOI
  12. [12]Syvolos L, Tran OT, Boureima ID, Anghel V, Garikipati K, Mourad HM. A phase-field fracture formulation for generalized standard materials: The interplay between thermomechanics and damage. Journal of the Mechanics and Physics of Solids 2025; 201: 106154. doi: 10.1016/j.jmps.2025.106154
  13. [13]Raissi M, Perdikaris P, Karniadakis GE. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 2019; 378: 686–707. doi: 10.1016/j.jcp.2018.10.045DOI
  14. [14]E W, Yu B. The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems. Communications in Mathematics and Statistics 2018; 6: 1–12. doi: 10.1007/s40304-018-0127-zDOI
  15. [15]Samaniego E, Anitescu C, Goswami S, et al. An energy approach to the solution of partial differential equations in computational mechanics via machine learning: Concepts, implementation and applications. Computer Methods in Applied Mechanics and Engineering 2020; 362: 112790. doi: 10.1016/j.cma.2019.112790DOI
  16. [16]Nguyen-Thanh VM, Zhuang X, Rabczuk T. A deep energy method for finite deformation hyperelasticity. European Journal of Mechanics - A/Solids 2020; 80: 103874. doi: 10.1016/j.euromechsol.2019.103874
  17. [17]Goswami S, Anitescu C, Chakraborty S, Rabczuk T. Transfer learning enhanced physics informed neural network for phase-field modeling of fracture. Theoretical and Applied Fracture Mechanics 2020; 106: 102447. doi: 10.1016/j.tafmec.2019.102447DOI
  18. [18]Goswami S, Anitescu C, Rabczuk T. Adaptive fourth-order phase field analysis for brittle fracture. Computer Methods in Applied Mechanics and Engineering 2020; 361: 112808. doi: 10.1016/j.cma.2019.112808DOI
  19. [19]Ghaffari Motlagh Y, Jimack PK, Borst dR. Deep learning phase-field model for brittle fractures. International Journal for Numerical Methods in Engineering 2023; 124(3): 620–638. doi: 10.1002/nme.7135DOI
  20. [20]Manav M, Molinaro R, Mishra S, De Lorenzis L. Phase-field modeling of fracture with physics-informed deep learning. Computer Methods in Applied Mechanics and Engineering 2024; 429: 117104. doi: 10.1016/j.cma.2024.117104DOI
  21. [21]Plunge N, Brommer P, Edwards RS, Kakouris EG. Deep Learning-Based Phase-Field Modelling of Brittle Fracture in Anisotropic Media. International Journal for Numerical Methods in Engineering 2026. doi: 10.1002/nme.70381DOI
  22. [22]Zhang P, Tang K, Xu B. ChebyKAN-enhanced deep energy method for phase field fracture: From homogeneous to heterogeneous materials. Journal of the Mechanics and Physics of Solids 2026; 214: 106693. doi: 10.1016/j.jmps.2026.106693DOI
  23. [23]Pantidis P, Svolos L, Abueidda D, Mobasher ME. Integrated Finite Element Neural Network (IFENN) for phase-field fracture with minimal input and generalized geometry–load handling. Computer Methods in Applied Mechanics and Engineering 2026; 448: 118485. doi: 10.1016/j.cma.2025.118485DOI
  24. [24]Pantidis P, Amin F, Abueidda D, Mobasher ME. A hybrid IFENN solver for generalizable modeling of phase-field fracture initiation and propagation. arXiv 2026. doi: 10.48550/arXiv.2606.27177arxiv.org/abs/2606.27177
  25. [25]Wang Y, Lin Y, Goswami S, et al. Towards unified AI-driven fracture mechanics: the extended deep energy method (XDEM). Nature Communications 2026; 17(1): 8492. doi: 10.1038/s41467-026-76748-1DOI
  26. [26]Zhang H, Makki Alamdari M, Shahbodagh B, et al. An Extended Deep Energy Method for Thermo-Mechanical Crack Propagation. arXiv 2026. doi: 10.48550/arXiv.2610.09433DOI
  27. [27]He J, Abueidda D, Abu Al-Rub R, Koric S, Jasiuk I. A deep learning energy-based method for classical elastoplasticity. International Journal of Plasticity 2023; 162: 103531. doi: 10.1016/j.ijplas.2023.103531DOI
  28. [28]Aldakheel F, Elsayed ES, Heider Y, Weeger O. Physics-based machine learning for computational fracture mechanics. Machine Learning for Computational Science and Engineering 2025; 1: 18. doi: 10.1007/s44379-025-00019-xDOI
  29. [29]Han W, Reddy BD. Plasticity: Mathematical Theory and Numerical Analysis. 9 of Interdisciplinary Applied Mathematics. New York: Springer. 2 ed. 2013DOI
  30. [30]Gerasimov T, De Lorenzis L. On penalization in variational phase-field models of brittle fracture. Computer Methods in Applied Mechanics and Engineering 2019; 354: 990–1026. doi: 10.1016/j.cma.2019.05.038DOI
  31. [31]Borst dR, Crisfield MA, Remmers JJC, Verhoosel CV. Non-Linear Finite Element Analysis of Solids and Structures. Wiley Series in Computational MechanicsChichester, UK: John Wiley & Sons. 2 ed. 2012
  32. [32]He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. arXiv 2015. doi: 10.48550/arXiv.1512.03385DOI
  33. [33]Bachlechner T, Majumder BP, Mao HH, Cottrell GW, McAuley J. ReZero is All You Need: Fast Convergence at Large Depth. arXiv 2020. doi: 10.48550/arxiv.2003.04887DOI
  34. [34]Broomhead DS, Lowe D. Multivariable functional interpolation and adaptive networks. Tech. Rep. RSRE-MEMO-4148, Royal Signals and Radar Establishment; Malvern, UK: 1988.
  35. [35]Riedmiller M, Braun H. A direct adaptive method for faster backpropagation learning: The RPROP algorithm. In: IEEE. ; 1993: 586–591.
  36. [37]Plungè N, Brommer P, Edwards RS, Kakouris EG. Variational deep learning for phase-field modelling of ductile fracture. Zenodo; 2026. Software and data, version 1.0.0. doi: 10.5281/zenodo.23348846.DOI
  37. [38]Li H, Fu MW, Lu J, Yang H. Ductile fracture: Experiments and computations. International Journal of Plasticity 2011; 27(2): 147–180. doi: 10.1016/j.ijplas.2010.04.001DOI

Paper details

vlnc:W_6fvhvuha
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