WorksComputer Science
A mesh-free deep energy method for phase-field modeling of thermally driven brittle fracture
Abstract
Phase-field simulation of thermally driven fracture requires resolving localized crack bands whose paths are unknown in advance and thermal boundary layers whose thickness evolves during the transient. Neural energy-based computations of thermal fracture have so far started from defects placed in advance. This paper proposes a mesh-free deep energy method based on multiresolution neural representations that predicts spontaneous crack nucleation, crack spacing, and selective crack propagation and arrest under transient thermal loading. Two networks, each with its own multiresolution encoding, represent the displacement and phase fields and the transient temperature, and the mechanical energy and the incremental functional of the backward Euler heat equation are estimated on integration points redrawn at every iteration. Restricting irreversibility to the crack set, carrying the optimizer state across load steps and decaying the learning rates within each step help prevent spurious damage accumulation. On two notched plates the initiation peak loads agree closely with an independent finite element reference. On a quenched glass strip under an analytical temperature field the crack path changes from straight to oscillatory and then to branched, with an oscillation threshold close to the linear stability prediction. Under a prescribed boundary temperature a quenched ceramic slab forms a crack array of the type observed in experiment, with a spacing close to that of a published benchmark. On the slab and a quenched annulus the arrays are denser than those of a finite element computation on the same material field. No crack paths, nucleation sites or crack-adapted meshes are prescribed.
Introduction
Brittle solids subjected to rapid surface cooling form characteristic fracture patterns. A ceramic slab quenched into water develops a periodic array of parallel cracks along its cooled faces, a subset of which propagates into the depth while the rest arrest, so that the spacing of the array grows with the depth [1, 2, 3]. A glass strip notched at its leading edge and drawn at constant speed from a hot region into a cold bath develops a single crack, which starts at the notch and follows the immersion front. As the temperature jump or the drawing speed grows, the path changes from straight to oscillatory and then to branched [4, 5]. The same mechanism, a contraction that the surrounding material resists, produces the columnar joints of cooling lava [6], the crack networks of drying and cooling in rock, concrete and soil [7, 8, 9] and the thermal shock fracture of ceramic components [10, 11]. The classical theory of thermal shock and the experiments that tested it [8, 10, 12, 13, 14, 15] describe the initiation of fracture from flaws of given size and the propagation and selective arrest of an initial crack array assumed in advance. A computation of thermal shock has to predict the pattern itself, where the cracks nucleate, how many form, at what spacing, and which of them continue.
The sharp-crack description of linear elastic fracture mechanics represents the crack explicitly and advances it when the energy release rate [16], or the stress intensity factor [17], reaches a critical value. Remeshing [18], the extended finite element method [19] and element-free Galerkin methods [20, 21] carry such a crack with separate criteria for nucleation, for the direction of growth and for branching, and such cracks have been advanced under thermal loads [22, 23, 24]. On the thermal shock problem those criteria would have to supply the quantities the computation is meant to predict. Cohesive surfaces between all elements [25], peridynamic models [26, 27, 28] and lattice models [29] instead form cracks where bonds or cohesive surfaces fail.
Of the formulations that produce such patterns without assuming them, the one adopted here is the variational formulation of fracture of Francfort and Marigo [30] in the regularization of Bourdin et al. [31, 32]. In it the sharp crack is replaced by a phase field that localizes in a band of width proportional to a regularization length [33], and nucleation, propagation, branching and arrest result from minimizing one energy. The model has since been extended by energy decompositions that treat tension and compression differently [34, 35, 36], interpreted as a gradient damage model with a finite strength [37, 38, 39], and given a strength surface [40], higher-order densities [41] and penalized irreversibility [42], as reviewed in [43]. On the thermal shock problem Sicsic et al. [9] showed that the uniform damage band under the cooled surface bifurcates toward a periodic mode whose wavelength is proportional to the regularization length. The computations of Bourdin et al. [44] pass through this bifurcation to an array of short cracks whose selective propagation follows the spacing law of Bahr et al. [8] and the experiment of Shao et al. [45]. The problem tests the model both at nucleation, where the strength governs, and in propagation, where the toughness governs, and in the regularized model the two are not independent, since the strength is set by the toughness, the modulus and the regularization length together [39]. Lopez-Pamies et al. [46] argue that classical models of this class, without a strength surface, cannot describe nucleation in general, and Kamarei et al. [47] pose a set of challenge problems on which fracture models can be assessed in this respect. Zeng et al. [48] treat nucleation in thermal quenching with a model whose strength is independent of the regularization length, together with random fields of the strength.
Of the two experiments above, the glass strip is the one in which quasi-static crack propagation under a thermal load is controlled and reproducible. Yuse and Sano [4, 5] mapped its straight, oscillatory and branched regimes and reported the transition to oscillation as a super-critical Hopf bifurcation. Ronsin et al. [49] measured the temperature field and the thresholds as functions of the drawing speed, Ronsin and Perrin [50] the thresholds and the wavelength at onset, and Yang and Ravi-Chandar [51] followed the growth of the crack in real time. Linear stability analyses [52, 53, 54, 55] and an equation of motion for the tip [56] gave the threshold and, in most cases, the wavelength, and the paths were computed with lattice models, finite elements, the extended finite element method, universal meshes, peridynamics and phase-field models [11, 23, 24, 27, 29, 55, 57, 58]. The computation of Corson et al. [55] recovered the transition but not the complex paths well above the threshold, where the authors suggest that the quasi-static approximation may no longer hold.
Phase-field computations of thermo-elastic fracture have been made with finite elements in quasi-static and dynamic settings and with monolithic, explicit and staggered solvers [59, 60, 61, 62, 63], for the hot cracking of the melt pool in additive manufacturing [64], in three dimensions [65, 66], on meshes adapted around the growing cracks [67, 68, 69] and with a strength independent of the regularization length [11]. Each of them meets the resolution requirement of the model with a mesh, and a strong-form meshless method [70] meets it with a set of nodes. On the thermal shock problem the mesh has to resolve two narrow zones that change during the computation. One is the band of width proportional to , along a path the computation is meant to discover [34, 71]. The other is the thermal boundary layer, whose thickness after a cooling time is , with the thermal diffusivity, and is smallest early in the transient, when the array nucleates. A refinement that anticipates either is not neutral, since nonuniform meshes can favor the nucleation of cracks in their refined zones [37]. A uniform mesh avoids this bias at the cost of the finest resolution everywhere, as in the finite element computations of this paper. In purely elastic fracture the authors’ mesh-free deep energy method [72] is built so that no mesh follows the crack.
Neural networks with smooth activations can serve as global trial functions that are defined without a mesh and differentiated to any order by automatic differentiation [73, 74, 75]. For problems governed by a minimum principle the deep Ritz method [76] and the deep energy method [77, 78] minimize the corresponding functional over the network parameters. Essential conditions can be imposed exactly through lifts [79], the distribution of the training points affects the accuracy of the solution and the convergence of the training [80, 81], and input encodings let networks represent steep fields [82, 83] and can speed up training [84]. Phase-field modeling of brittle fracture has been treated in this way by Goswami et al. [85, 86] and Manav et al. [87] and in a number of variants [88, 89, 90, 91, 92, 93, 94]. An explicit crack has been advanced by the extended deep energy method of Wang et al. [95], networks have served as solution operators and surrogates [96, 97, 98], and networks that minimize a peridynamic energy have propagated cracks [99, 100]. The study of [72] found that the crack advanced only when the integration points were redrawn during training rather than held fixed.
In the thermal and thermo-elastic applications of networks [101, 102, 103, 104, 105], no crack was present, and where networks have been used for thermal fracture, the crack was mostly stationary and a stress intensity factor was the output [106, 107, 108]. Three recent studies allow the crack to grow under a thermal load. He et al. [109] apply the deep Ritz method to fracture under a steady or spatially uniform temperature field, on a finite element mesh and from initial cracks. Kong [110] grows cracks from notches under a steady temperature held fixed during loading, with an energy-based network, a second-order phase field and a history field. The authors’ extended deep energy method [111] solves the transient temperature on the cracked domain and advances an explicit crack from a notch. In all three the crack starts from a defect placed in advance, so the nucleation, the spacing and the selection of cracks under a transient temperature field are not computed.
This paper proposes a mesh-free deep energy method that computes them. It builds on the solver of [72] (§ 3) and is mesh-free in the sense that no connectivity is built and no point set is held fixed. Two difficulties arise when a transient temperature drives the cracks. The thermal boundary layer can be thinner than the crack band at the first time step and thickens as the transient advances, so the temperature is computed by a second network with a multiresolution encoding of its own, from an incremental functional of the backward Euler heat equation in which the convective condition is a natural term. Under a uniformly cooled surface the damage is diffuse over a large region before cracks nucleate, and a stochastic solver that retains every small increase of the phase field turns the cooled layer into a damaged film. Irreversibility is therefore imposed on the crack set only (§ 2.1), and, to keep the bands from widening, the optimizer state is carried across load steps and the learning rates decay within each step (§ 3.3). The resolution is set by the uniform grids of the encoding, and the integration points follow the crack band as it develops, so no mesh is fitted to the band or to the boundary layer and nothing is remeshed as the array grows. A curved surface is described exactly by a spline map. Where the heat transfer is little affected by the cracks and the compared computations use the same closed form, as for a strip drawn at constant speed into a bath, an analytical temperature field can be used instead of the computed one. It enters the mechanical problem as a known thermal strain. Where the temperature is computed, each example treats the conductivity as damage-dependent, if the cracks cross the heat flux, or as damage-independent, if they run along it.
The method is assessed on five problems. A single-edge-notched plate under tension and a functionally graded plate under shear, each at three thermal loads, verify the thermo-mechanical solver against an independent staggered finite element reference of the same phase-field model, up to the differences stated in § 4. The directional quenching of a glass strip tests the evolution of a crack path under an analytical temperature field and is compared with a finite element computation, the phase-field computation of Zeng and Dolbow [11], the linear stability analysis of Corson et al. [55], and the experiments of Yuse and Sano [4, 5] and of Ronsin and Perrin [50]. The quenched ceramic slab combines a computed thermal transient with the spontaneous nucleation of an array, and a quenched annulus extends it to a curved domain. Both are compared with a finite element computation on the same realization of the random material field of Tang et al. [112], which breaks the symmetry of the uniformly cooled surface. The slab is also compared with the experiments of Jiang et al. [3] and Shao et al. [45] and the benchmark computation of Bourdin et al. [44].
Section 2 states the model, Section 3 the method, Section 4 the five examples and Section 5 the conclusions, with the numerical parameters and the supporting studies in the appendices.
Thermo-elastic phase-field model
Consider a linearly elastic body occupying an open bounded domain , analyzed under plane strain or plane stress as each example states, with normal to the plane. The body is loaded quasi-statically through prescribed boundary displacements and through a temperature field , with the position and the time. The temperature is either computed from a heat conduction problem or given as an analytical field. Cracks are described by a scalar phase field , with in intact material and on a fully developed crack, and small strains are assumed for the displacement . The fracture energy density, the strain energy decomposition and the hybrid formulation are adopted from the purely elastic method of [72]. The fracture energy density is the second-order density in the approximation of Ambrosio and Tortorelli [33], commonly denoted AT2. The strain energy is split spectrally into a tensile part that drives the crack and a compressive part that does not, and in the hybrid formulation the stress is degraded isotropically while only the tensile part drives the phase field. These ingredients are stated below with the temperature included, followed by irreversibility, the heat conduction problem and the coupling of the two problems. Table 1 collects the main symbols of the paper.
| Symbol | Meaning | Symbol | Meaning |
| , | physical, parametric domain | , | mechanical network, its parameters |
| , , | position, parametric coordinates, time | , | temperature network, its parameters |
| , , | displacement, strain, stress | , , | feature vectors of the two networks, number of levels |
| , | density, specific heat | , , | nodal features of level , quadratic B-spline, nodes along |
| , , | elastic strain, scalar thermal expansion strain, expansion coefficient | effective in-plane eigenstrain of plane strain | |
| , , , , | elastic constants, plane-stress constant | Jacobian of the geometry map | |
| , | phase field, seeded notch profile | , | sampling density on , points per iteration |
| , , | regularization length, toughness, reference toughness at | driving-force indicator of the sampler | |
| , , | total energy, elastic density, its tensile and compressive parts | , | envelopes of the displacement and temperature network terms |
| , | degradation function, residual stiffness | , , | displacement and temperature scales, displacement lift |
| , , , | folded hybrid, fracture and penalty densities, their sum | squash function of the phase field | |
| , , | penalty weight, threshold and indicator of the crack set | , , , | plateau window length, tolerance, window, monitored estimate |
| , , , | loss of load step , load step, number of load steps, prescribed displacement | , , | reaction force, residuals of the mechanical problem |
| , , , | temperature, stress-free and initial temperature, bath temperature, thermal load ( on the plates, on the strip, the slab and the annulus) | , | Griffith length , peak stress of the homogeneous AT2 model |
| , , , | conductivity, conductivity of a load step, volumetric heat capacity, diffusivity | , , , , | Weibull modulus, modulus, strength and toughness of cell of the material field, nominal strength |
| , , , | film coefficient, convective boundary, boundary with prescribed temperature, prescribed boundary temperature and its extension into the domain, or the analytical field | , , , , , | incremental thermal functional, its density, time step, time of load step , points on edge and its parameter |
| , | drawing speed and half-width of the strip | , , | Péclet and thermal-shock numbers, threshold of oscillation |
| , , , | amplitude and measured wavelength of an oscillating path, number and position of its turning points | , | crack spacing, depth from the cooled face |
| critical stress intensity factor | , , , , | constants, reference, normalizing and normalized temperature of the toughness law | |
| , , , , | stratum weights, process-zone weights of the sampler | , , | initial, first and last time of the quench time grid |
| , , | thermal and mechanical updates of load step , iteration index | , , , | distance to the notch, distances to the cooled surfaces, decay length of the temperature lift |
Table 1.
Table 1. Main symbols. Total energies are written with an uppercase , densities with a lowercase , as in [72]. The load step index appears as a subscript, and descriptive subscripts are upright (, ). is the identity tensor, the trace, and the in-plane and the three-dimensional trace, the gradient, the divergence, the colon the double contraction, the Euclidean norm, the bold the outward normal of the boundary, the positive and negative parts, the expectation, a hat on , and an estimate from sampled points and a hat on , and the raw output of the mechanical network, and the stop-gradient. On , and , the subscript marks a cell of the material field, and is the nominal strength.
Energy functional
The temperature enters the mechanics through the scalar thermal expansion strain and the elastic strain ,
where is the coefficient of thermal expansion and the stress-free temperature at which the thermal strain vanishes.
The total energy of the body at a given temperature field reads
in which is the quadratic degradation function with the residual stiffness . The second integral is the fracture energy in the form of Bourdin et al. [31, 32], with the toughness and the regularization length . The densities are the tensile and compressive parts, in the spectral decomposition of Miehe et al. [34, 113], of the strain energy density of the undamaged material,
evaluated on the elastic strain, with its principal values and the Lamé constants and , which follow from Young’s modulus and Poisson’s ratio . Under plane strain the total out-of-plane strain vanishes, so the elastic strain contains the restrained component , which enters Eq. (3) through the three-dimensional trace and through the principal values. Under plane stress the split is the common two-dimensional approximation, in which is replaced by and the trace and the principal values are those of the in-plane elastic strain. The toughness may vary in space and with the temperature. Where it depends on the temperature, it follows the quadratic law of Ruan et al. [64],
in which , which may vary in space, is the toughness at the temperature , and are material constants, is the normalizing temperature of that law and is the normalized temperature.
Equation (2) defines the fracture energy and the tensile part that drives the crack and is not itself minimized. The equations solved are those of the hybrid formulation of Ambati et al. [36], in the folded form of [72] and applied to the elastic strain in place of the total strain. In this formulation the stress of the damaged material is degraded isotropically,
while the phase field is driven by alone. The stress and the driving force are both carried by a single density, folded in the sense that its derivative with respect to gives the degraded stress and its derivative with respect to the driving force of alone,
where denotes the stop-gradient operation, whose argument is treated as a constant during differentiation. These two derivatives give the equilibrium and the phase-field equations of the hybrid formulation, as shown in [72], and the value of is the degraded strain energy density , so the computed energies and the reaction force keep their mechanical meaning. The constraint of the formulation of Ambati et al., where , is not imposed.
The load is applied in steps , and irreversibility between successive steps is enforced by the one-sided penalty
with the penalty weight , the trial phase field of step , written once the step is accepted, the phase field accepted at the end of the previous step, and the indicator of the crack set with the crack-set threshold . The indicator restricts the penalty to the developed cracks, so that outside the crack set the phase field may decrease. The construction follows the crack-set irreversibility of Bourdin et al. [31, 32], in which partially damaged regions may heal [42, 114], except that the penalty acts against a decrease below the previous value rather than setting on the crack set.
With the penalty acts on the whole domain, the form of [72]. Where diffuse damage covers a large region, as under a uniformly cooled surface before cracks nucleate, each stochastic update of the optimizer (§ 3) perturbs the phase field slightly, and a penalty on the whole domain retains every increase, so that the increases accumulate into a damaged film. The threshold counters this, and two options of the mechanical update (§ 3.3) keep the bands from widening (Appendix § 10). Since the weight is finite, the constraint holds approximately (Appendix § 9).
At load step , with the temperature and, where the example has one, the prescribed boundary displacement , the mechanical problem is stated on the loss
with the fracture energy density of Eq. (2). The loss is written below where the step is clear. The solution of the step is the pair , with admissible for , at which the two residuals
vanish, both derivatives taken with the arguments of held constant. By the construction of Eq. (6), is the weak form of the equilibrium equation with the stress of Eq. (5), and that of the phase-field equation driven by , with the penalty added. As in [72], the two fields are advanced together by one gradient-based update per optimizer iteration on these derivatives. This update, like the hybrid formulation itself, does not minimize a single energy [36, 43], so the value of need not decrease from one iteration to the next. In the isothermal limit with the formulation reduces to that of [72]. Without the penalty, the strong forms of the two conditions are
with the natural condition on the boundary.
Heat conduction
Where the temperature is computed from the heat conduction problem below, each example treats the conductivity in one of two ways. Where the cracks cross the heat flux, their faces are taken as adiabatic through a damage-dependent conductivity [63, 64]. The conductivity of load step is degraded by the phase field of the previous step, so that the thermal problem of a step can be solved before its mechanical problem,
with the conductivity of the intact material and the residual stiffness . For a finite and the crack band retains a small conductance, so Eq. (11) is a phase-field approximation of adiabatic crack faces. Where the cracks run along the heat flux, the conductivity is damage-independent [44, 60, 62, 63], , so that the crack band conducts heat as intact material and the temperature does not depend on the cracks. Table A.1 lists the treatment of each example.
The temperature obeys the transient heat conduction equation. The boundary of is divided into a part with prescribed temperature, a convective part through which heat passes to the bath, and an insulated remainder,
with the volumetric heat capacity , the conductivity of the load step, the prescribed boundary temperature , the film coefficient and the bath temperature . The thermal diffusivity is , formed with the conductivity of the intact material, with the product of the density and the specific heat . The backward Euler difference over a step , with the time of load step , turns Eq. (12) into the stationarity condition of the incremental functional
with the temperature of the previous step and the conductivity of the step. The functional is written where the step is clear, with its domain integrand. In its stationarity condition the convective and the insulated conditions are natural and need no constraint on the field, like traction-free edges in the mechanical problem, so only the condition on is imposed. The functional is strictly convex in for and , so the temperature at is its unique minimizer over the fields that carry the prescribed data on , and it is computed by the same estimator and optimizer as the mechanical fields. The convexity does not carry over to the parameters of the network that represents the field. Incremental minimum principles for coupled heat conduction [115, 116] and implicit time steps computed as minimizations over networks [111, 117, 118] precede this formulation.
Coupling of the two problems
Within a load step the coupling is one way. The temperature acts on the mechanics through the thermal strain of Eq. (1) and, where the toughness depends on the temperature, through Eq. (4) evaluated at and held fixed during the mechanical problem. The mechanics acts on the temperature only through the damage-dependent conductivity of Eq. (11), evaluated on , and not at all where the conductivity is damage-independent. The temperature of a step is therefore computed first and passed to the mechanical problem as data (§ 3.3). Where the temperature is an analytical field, as on the strip, no thermal problem is solved and the thermal strain is a known eigenstrain. The thermo-elastic heat source [119] and the heat released by fracture [59] are neglected, as in the computations of Mandal et al. [62] and Bourdin et al. [44].
Mesh-free deep energy method
The incremental problems of § 2 are solved with the mesh-free deep energy method of [72]. Its representation of the displacement and the phase field by one network on a multiresolution encoding of B-spline feature grids, its stratified Monte Carlo estimator on integration points redrawn at every iteration and its training by Adam [120] are used unchanged and are restated only as far as the temperature requires. For the thermal problem the method adds a second network for the temperature on feature grids of its own (§ 3.1). The thermal functional is estimated on points drawn from the same density as those of the mechanical loss (§ 3.2). Each load step obtains the temperature first, from the thermal update or the analytical field, and passes it to the mechanical update, for which two options are introduced (§ 3.3). The numerical parameters of every run are tabulated in Appendix § 9.
Field representation and boundary conditions
The displacement and the phase field are represented as in [72] by one network on the parametric domain with the coordinates ,
a multilayer perceptron with smooth Gaussian error linear unit (GELU) activations [121] and the parameters . The perceptron returns the raw outputs , and and is fed with the coordinates and with the feature vector of levels of trainable feature grids. The subscript M marks the mechanical network and T the temperature network introduced below. The parameters comprise the weights of the perceptron and the feature values of the grids. Level stores a feature vector at each node of a uniform grid with nodes along (Appendix § 9). The feature vector of the level at a point is the quadratic B-spline blend of the nodal vectors at the nodes around it,
where and are the quadratic B-splines of level centered on node along and on node along , three of which are nonzero in each direction at any point. Figure 1 shows the construction. The blend makes every feature in the coordinates, so with the smooth activation the strains and the gradient of the phase field taken by automatic differentiation are continuous (a linear blend gives features). The spacing of the finest grid is the smallest scale the representation can express, uniform over the parametric domain and fixed before training. What concentrates on the crack is the density of the integration points (§ 3.2). The grids are a tensor-product parameterization of the field, on which no connectivity is built and to which no integration rule is tied. Curved domains enter through an exact non-uniform rational B-spline (NURBS) map of the parametric square, as in isogeometric analysis [122] and in [72].

Figure 1. Field representation of the proposed method. The feature grids of the levels, from coarse to fine (top left), are read at the same point (red). On each level (top right) the feature at the point is the quadratic B-spline blend of the nodes around it (blue), weighted by products of the one-dimensional basis functions of Eq. (15). The features of all levels form the vector , which the network takes together with (bottom). The mechanical and the temperature networks share this structure, each with feature grids of its own.
The raw outputs become fields that satisfy the essential conditions exactly. Each displacement component is written as , with the raw output of that component, a displacement scale, an envelope that vanishes where the displacement is prescribed and a lift, a function that carries the prescribed values, so the essential conditions hold for any network parameters. A displacement fixed at a single point is imposed instead by subtracting the network value there. The phase field is assembled around a seeded notch,
where is the squash function and is the optimal profile of the AT2 density about the notch at the distance from it, with where no notch is seeded. Since is positive, the intact state is approached but not represented exactly (Appendix § 9).
The temperature varies steeply within a boundary layer whose thickness may change by more than an order of magnitude over a transient, while the crack band has the fixed width set by . The temperature is therefore represented by a second network with a narrower perceptron, parameters disjoint from and feature grids of its own. Without feature grids the early layer may not be resolved (Appendix § 11). With their feature vector the temperature of step is
where the lift extends the prescribed temperatures of into the domain and vanishes on . The temperature scale normalizes the raw output to order one (Appendix § 9). The network takes neither the time nor an explicit crack geometry as input, and the temperature of step follows from the lift at and from the parameters obtained in that step. Where no boundary temperature is prescribed, and . On the notched plates interpolates linearly between the prescribed temperatures of the bottom and the top edge (Appendix § 9). Where a surface is held at from a uniform initial temperature ,
written for two cooled surfaces, with and the physical distances to them and the decay length chosen well below the thickness of the boundary layer at the first time step. The lift holds the cooled surfaces at from the outset, so the first backward Euler step forms the boundary layer from the initial jump. An analytical temperature field is set as with .
Monte Carlo integration
The loss and the thermal functional are both estimated in the same way as the energy in [72]. For any sampling density on , an integral over the domain is the expectation of its integrand times divided by the density, with the Jacobian of the geometry map. With the integrand of the loss in Eq. (8) and points drawn from ,
where denotes the expectation over points drawn from , and the estimator is unbiased for any density that is positive and evaluated exactly at the drawn points. Points drawn from the same density, independently at every iteration of the thermal solve, give the estimator of the thermal functional,
where is the thermal density of Eq. (13) and the second sum runs over the convective edges , each with points drawn uniformly in its parameter and weighted by the arc length per unit parameter. The importance weights and the weights for the arc length are treated as data and excluded from differentiation.
The interior points are drawn from the mixture of three strata of [72],
with stratum weights that sum to one, , and each stratum a normalized density weighted by , so that the uniform stratum is uniform in the physical domain. The crack stratum is proportional to , and where no notch is seeded it is instead a fixed band along the cooled faces. The process-zone stratum, which follows the growing cracks, peaks on the flanks of the crack band and ahead of the tip, with the weights of the notch profile and of the driving-force indicator . The indicator is a tensile strain energy of the previous state evaluated on its elastic strain, since under a thermal load the total strain does not measure the stress (Appendix § 9). The crack and process-zone strata are piecewise constant on an auxiliary grid of cells, so their densities are known exactly at every point, and since the uniform stratum keeps , the estimator stays unbiased for the mechanical loss and the thermal functional alike, which Appendix § 11 checks on the thermal solve. All strata are frozen within a step and rebuilt from the accepted state.
The points are redrawn at every optimizer iteration, as in the deep Ritz method [76], for the thermal problem as for the mechanical one. On a fixed point set the optimizer lowers the estimated functional at the points while the field deteriorates between them [123], as [72] measured on the purely elastic problem and Appendix § 11 measures on the thermal one.
Solution of a load step
Since the temperature of a step does not depend on its mechanical solution (§ 2.3), each load step solves the two problems in sequence,
The temperature at is found first by the thermal update , the iterations of Adam on over at the frozen phase field , starting from the parameters of the previous step, with the temperature supplied by a frozen copy of that step’s network. The displacement and the phase field are then advanced by the mechanical update , the iterations of Adam from on the derivatives of the estimate of the loss of Eq. (8), and its result is the accepted state of the step. The thermal strain is taken as
so that the mechanical gradient does not pass through the temperature network and the temperature enters the mechanical loss as data. No state is stored at integration points, since ,
and the stratum densities are fields that can be evaluated wherever the sampler draws next.
Each solve stops when the relative range of its estimated loss or functional over a trailing window of iterations falls below a tolerance, or when its iteration budget, the maximum number of iterations, is spent (Appendix § 9). The test measures the noise of the estimate as well as its drift and is not a residual test, so the result of a step depends on the budget, which is set per example. On the strip every step uses its full budget, which is chosen in Appendix § 13.
At the network , with its output layer initialized at zero, returns zero, so the initial temperature is the lift , which needs no training. The mechanical network is initialized and trained in the first step as in [72]. Both solves use Adam only, since the redraw changes the estimate of the objective at every iteration and a line search would act on the noise of one draw. Figure 2 shows one load step, with what each solve receives from the frozen state and what it passes on.

Figure 2. One load step of the thermo-mechanical solver. The thermal functional (orange) is minimized and the mechanical fields (blue) are advanced, the two updates of Eq. (22) running in sequence, each on its own parameters, on points redrawn at every iteration. The temperature enters the mechanical loss as the thermal strain of Eq. (23), through which no gradient passes, and the accepted state becomes the frozen state of the next step (dashed), until the last load step ends the run.
Two options of the mechanical update are introduced here for problems in which many cracks nucleate under a uniform load. In the update of [72] the network parameters are carried across load steps while Adam is restarted. With its running averages of the gradient reset, Adam moves every parameter by about the learning rate in the first iterations, which perturbs the phase field throughout the domain at the start of every step. The first option carries the optimizer state across load steps. In the second, the learning rates decay exponentially over the budget of each step to 10% of their initial values, so each step ends with small updates, and the rates return to their initial values at the next step, a learning rate schedule with warm restarts [124]. The two options are used together and with the crack-set threshold of Eq. (7), the only combination tested on the slab that kept its bands at a nearly constant width and its cooled face free of a damaged film (Appendix § 10).
At the end of a step the reaction force on a loaded edge is the derivative of the loss with respect to the prescribed displacement, taken with the network parameters held at their accepted values,
where is the component of the prescribed displacement that loads the edge, and is the loss of Eq. (8) with the fields of the network and the displacement lift of . Since the lift depends on while the parameters stay fixed, and the stop-gradient leaves only the degraded strain energy in this derivative, is the work conjugate of , which equals the resultant of the hybrid stress in Eq. (5) on the edge when equilibrium holds (Appendix § 9). It is evaluated on a dedicated sample several times larger than the points of an iteration, as in [72].
Numerical examples
The two notched plates, under tension and under shear, are reported first, followed by the directionally quenched glass strip, the quenched ceramic slab and the quenched ceramic annulus, the last two being the quench problems. The temperature is computed with a damage-dependent conductivity on the notched plates, given as an analytical field on the strip, and computed with a damage-independent conductivity on the slab and the annulus (§ 2.2). The cooled faces of the slab and the annulus carry a prescribed boundary temperature, and the slab under the convective condition is treated in Appendix § 15. The toughness depends on the temperature, through Eq. (4), only on the notched plates.
The finite element computations use an independent staggered phase-field solver built on scikit-fem [125] with the hybrid formulation, checked against the published results of Tangella et al. [63] (Appendix § 12). Where the temperature is computed, the solver advances the heat conduction problem by backward Euler on the same time grid as the proposed method. It enforces irreversibility through the history field on the notched plates and as a bound on the phase field on the strip and the quench problems. It serves as the reference on the notched plates and as a computation of the same model on the strip and the quench problems, where “the same model” means the same phase-field model up to the treatment of irreversibility and the representation of the notch. On the notched plates the reference also omits the temperature-dependent toughness and the corner toughening of § 4.1. On the strip and the quench problems the bound prevents any decrease of the phase field, whereas the penalty of the proposed method acts only on the crack set, so where diffuse damage localizes the two solutions may differ. § 4.4 tests this difference on the slab.
Each problem is computed with one set of numerical parameters at all of its loads, and these are listed in Table A.1. The strip, the slab and the annulus use the crack-set threshold together with the optimizer state carried across load steps and the learning rates decaying within a step (§ 3.3), while the notched plates use and the update of [72].
Single-edge-notched tension test
With this problem we verify the thermo-mechanical solver on the tension specimen of [72], here with a thermal load added. A square plate of side 1 mm carries a horizontal notch of length 0.5 mm from the left edge at mid-height, Figure 3(a). The bottom edge is fixed and held at the initial temperature K, and the top edge is displaced vertically by in 80 steps of mm while its temperature is ramped linearly to at step 20 and then held, with , 0 and K. The lateral edges are free and insulated. The material is the ceramic of Tangella et al. [63] under plane strain, Table 2, with the damage-dependent conductivity of Eq. (11), since the crack crosses the heat flux, and the toughness law of Eq. (4) with the constants of Ruan et al. [64]. As in [72], the toughness is doubled within about of the four corners of the fixed and the displaced edges, so that the singular driving force there does not nucleate damage. The displacement lift interpolates linearly between the bottom and the top edge (Appendix § 9). Physical time is s mm, so the thermal boundary layer under the driven edge reaches a thickness of 0.24 to 0.33 mm at the peak loads of the cooled and heated cases.
| Property | Symbol | Homogeneous plate | Graded plate | |
| mm | ||||
| Young’s modulus (GPa) | 340 | 340 | 450 | |
| Poisson’s ratio | 0.22 | 0.22 | 0.22 | |
| Toughness (J m) | 42.47 | 42.47 | 120 | |
| Expansion coefficient (K) | ||||
| Conductivity (W m K) | 300 | 300 | 350 | |
| Density (kg m) | 2450 | 2450 | 4000 | |
| Specific heat (J kg K) | 775 | 775 | 775 | |
| Toughness law constants | , | 1.80, 1.10 | 1.80, 1.10 | 1.80, 1.10 |
| Toughness law temperatures (K) | , | 300, 1000 | 300, 1000 | 300, 1000 |
| Regularization length (mm) | 0.01 | 0.01 | 0.01 |
Table 2. Material data of the notched plates, after Tangella et al. [63]. The tension plate is homogeneous. The graded shear plate varies linearly in from the values of the homogeneous plate at to those of the last column at mm, taken from the released code of Tangella et al. [126], with , and constant. The toughness law of Eq. (4) takes its constants from Ruan et al. [64]. The regularization length is chosen here, against 0.005 mm in Tangella et al.
Figure 3. Geometry and boundary conditions of the two notched plates. (a) Homogeneous plate under tension. (b) Graded plate under shear, with stiffness, toughness, conductivity and density increasing from left to right. In both the bottom edge is fixed and held at , the top edge is displaced and brought to , and the lateral edges are free and insulated. Dimensions are marked on the first panel, and the temperature shading is schematic.
Figure 4 compares the computed load-displacement curves with the reference at the three thermal loads, and Figure 4(b) enlarges the initiation peak. The peak loads agree with the reference to within 1.30%, 1.08% and 1.76% in the cooled, isothermal and heated cases, and the displacement at the peak to within one load step. The isothermal case, which has no thermal load, agrees about as closely as the purely elastic computation of [72], which agreed to within about 1% on the same test and material.
Figure 4. Notched plate under tension. Load-displacement curves of the proposed method against the finite element reference at three thermal loads. (a) Full response. (b) Detail of the initiation peak. The cooled case is blue, the isothermal case black and the heated case red, the reference dashed and the proposed method solid.
The reference peaks rise by 4.5% from the cooled to the heated case, a rise that the proposed method, whose peaks lie within 1.3% of one another, does not reproduce. The toughness law that the reference omits does not explain this difference, since it changes the reference peak of the cooled case by 0.23% (Appendix § 12). The thermal load changes the reference peak far less on this plate than on the graded plate, by 4.5% against 28%. The thermal solve is therefore verified here through the temperature field, and the coupling through the peak load on the graded plate of § 4.2.
Figure 5 shows the temperature, the vertical displacement and the phase field of the proposed method from the start to failure. In the cooled case the crack bends slightly toward the cooled edge near the right side, as in the reference, whereas the isothermal crack runs straight.
Figure 5. Notched plate under tension at K, the cooled edge at the top. Temperature, vertical displacement and phase field of the proposed method at load steps 0, 36, 42 and 80 of 80, the initial state, the last step before the peak load, a step during the propagation and the final state. Each row shares one color scale. The region below the crack stays at once the crack has crossed, since the crack faces are nearly adiabatic.
Away from the notch the temperature depends on and alone, and along a vertical line it agrees with an implicit finite difference solution of the one-dimensional transient to within 0.44 K until the crack crosses the plate (Appendix § 9). Across the notch it changes by a few kelvin over the width of the phase-field band, Figure 6, since the damage-dependent conductivity of Eq. (11) makes the notch faces nearly adiabatic, and the temperature network represents this jump with no crack geometry in its inputs.
Figure 6. Notched plate under tension at K, temperature of the proposed method along two vertical lines at the step before the peak load, through the notch and away from it. The jump on the notch plane is the effect of the nearly adiabatic notch faces.
Single-edge-notched shear test with functionally graded material
As the next example, the same notch is loaded in shear on a functionally graded plate, Figure 3(b). This example verifies the solver where the material properties, and with them the thermal stress, vary across the specimen. The top edge is displaced horizontally in 160 steps of mm while its temperature is ramped linearly to at step 12, with , 0 or K. The properties vary linearly in to the stiffer and tougher end values of Table 2, in the reference as well. Equation (4) is applied to the graded , and the remaining settings are those of § 4.1.
Figure 7 compares the load-displacement curves. The crack initiates at the notch at the first peak, arrests as it reaches the stiffer and tougher material, and the load rises to a second maximum before the plate fails. Both solvers pass through these four stages. The initiation peak, the global maximum of every curve, agrees to within 1.18%, 0.14% and 0.58% in the cooled, isothermal and heated cases, with the peak displacement within one load step. The second maximum differs by at most 7.3%, the plate fails at a displacement about 4% larger than in the reference, and on the reloading branch the force of the proposed method scatters by up to 2% about a smooth curve, since the energy is estimated stochastically. Figure 8 shows the fields of the heated case at these stages, with a downward kink of the crack at the notch tip.

Figure 7. Graded notched plate under shear. Load-displacement curves of the proposed method against the finite element reference at three thermal loads. (a) Full response through initiation, arrest, reloading and failure. (b) Detail of the initiation peak. The cooled case is blue, the isothermal case black and the heated case red, the reference dashed and the proposed method solid.
Figure 8. Graded notched plate under shear at K, the heated edge at the top. Temperature, horizontal displacement and phase field of the proposed method at load steps 0, 56, 76 and 132 of 160, which are the start, the initiation peak, the step near the arrest trough and the step after the second maximum. Each row shares one color scale.
The initiation peaks of the proposed method spread over 26%, from 92.1 N at K to 116.0 N at K, against 28% in the reference. Since the displacement of the driven edge normal to the notch is held at zero, the thermal expansion of the boundary layer compresses the notch plane in the heated case and opens it in the cooled one, which raises the shear load at initiation in the first and lowers it in the second.
Directionally quenched glass strip
With this problem we test the evolution of a crack path under an analytical temperature field. A hot glass strip drawn at constant speed into a cold bath carries a crack that follows the immersion front, and as the temperature jump grows the path changes from straight to oscillatory and then branches, while at larger jumps several cracks start along the leading edge [4, 5]. The strip of Zeng and Dolbow [11], Figure 9, is 30 by 16 mm with half-width mm and made of the glass in Table 3. It is computed under plane stress and drawn at mm s into a bath at K. Their model carries a tensile strength of 80 MPa independent of the regularization length, whereas the AT2 model at mm has the homogeneous peak stress MPa [39]. The comparison is therefore between two models on one problem, and Zeng and Dolbow show that the strength also affects the path. The control parameters are the Péclet number and the thermal-shock number , with in kelvin, where is the jump from the stress-free hot region to the bath and . This is not the resistance parameter of Hasselman [12]. The notch of Zeng and Dolbow, of length 5 mm at the leading edge and offset 0.25 mm from the centerline, is retained. The trailing edge carries zero horizontal displacement and its midpoint zero vertical displacement, and all other edges are traction free, so the thermal strain is the only load.
| Property | Symbol | Value |
| Young’s modulus (GPa) | 70 | |
| Poisson’s ratio | 0.22 | |
| Toughness (J m) | 8.25 | |
| Expansion coefficient (K) | ||
| Diffusivity (mm s) | 0.47 | |
| Regularization length (mm) | 0.1 |
Table 3. Material data of the glass strip, after Zeng and Dolbow [11]. The strip is computed under plane stress. The regularization length is chosen here, against 0.03 mm in Zeng and Dolbow.
Figure 9. Geometry of the glass strip, drawn at speed from the hot region into the bath. The shading and the curve are the analytical temperature of Eq. (25) at one instant, and the leading notch of length 5 mm is offset 0.25 mm from the centerline. The transition length, the offset and the width are not to scale.
The profile of the front is steady in the frame of the bath, in which advection balances diffusion ahead of the front, and reads in the coordinates of the strip
with the coordinate along the strip. It is set as in Eq. (17) with , so no thermal problem is solved, and it neglects the effect of the crack on the temperature. Zeng and Dolbow and Corson et al. [55] use the same profile. The temperature jumps are 130, 145, 155, 165, 170, 175, 190, 217 and 307 K, and at each of them the front travels 24 mm in 240 load steps, each advancing it by one regularization length. The increment and the iteration budget are chosen in Appendix § 13.
The oscillatory path is measured by two order parameters, the amplitude and the wavelength , read from the ridge of the phase field traced from the notch to the tip. With the successive turning points of this path, in number,
so that is the period of a nearly periodic path and half the largest half swing between two successive turning points. Both are read on the developed path, up to mm, which excludes the last quarter of the travel, where the tip approaches the end of the strip (Appendix § 14).
Figure 10 shows the final paths at six of the nine temperature jumps. The path is straight at 130 K, oscillates from 145 to 165 K with an amplitude that grows with the jump, and branches from 170 to 217 K. At 170 and 175 K it first oscillates and then forks, the fork moving toward the notch as the jump grows, and from 190 K the tip splits at the end of the notch. This is the sequence observed in experiment, obtained here with one set of numerical parameters. Zeng and Dolbow [11] find the same order at this speed, straight to 127 K, oscillatory from 147 to 167 K and branched above. At 307 K two further cracks nucleate along the leading edge, and at this jump the crack of Zeng and Dolbow turns back and follows an irregular wavy path. The finite element computation gives the same class of path at each jump at which it was run except 145 K, where its path stays straight. At 165 K the wavelength and the amplitude of the proposed method are 17% and 59% larger than those of the finite element computation at 0.05 mm per step, its finest increment (Appendix § 13). The faint band at the immersion front is partial damage, , which recedes behind the front since irreversibility acts on the crack set only.
Figure 10. Glass strip, phase fields at six temperature jumps, the bath on the left and the front advancing per step. (a) = 130 K, straight. (b) 165 K, oscillatory. (c) 175 K, oscillatory over about one and a half periods, then branching. (d) 190 K, branching at the end of the notch. (e) 217 K, branching at the end of the notch, with oscillating branches. (f) 307 K, two further cracks nucleated along the leading edge beside the crack from the notch. The faint vertical band near the right end marks the immersion front at the last step. All panels share one color scale from intact (dark) to fully damaged (light).
The path at 130 K stays within 0.09 mm of a straight line, and the amplitude and the wavelength of the three oscillating cases rise with , Figure 11. In these cases the fourth half swing, centered between and 15.5 mm, is the largest, and the later ones decrease. At 145 K the last half swing of the developed path is 0.94 of the largest and the last of the whole path 0.28 of it, so the oscillation there decays over the last quarter of the travel, which the developed path excludes. Yuse and Sano [5] reported the transition to oscillation as a supercritical Hopf bifurcation, in which the oscillation grows continuously from zero and its squared amplitude grows linearly with the distance from the threshold , as in the phase-field computation of Corson et al. [55]. The three oscillating cases are consistent with this linear growth, Figure 11(a), and the line gives , between the straight case at and the lowest oscillating case at . Leaving out one case moves by at most 0.1 (Appendix § 14), and over the five settings of Appendix § 13 at which 145 and 165 K both oscillate, lies between 8.97 and 9.18. Three cases cannot distinguish this growth from the first-order transition described by Bahr et al. [57]. The wavelength rises linearly with , Figure 11(b), and its value at the threshold, , is compared with experiment, which measures it at onset [50].
Figure 11. Glass strip at , order parameters of the proposed method on the developed path against the thermal-shock number , with the temperature jump on the upper axis. (a) Squared amplitude. (b) Wavelength. The lines are least-squares fits through the three oscillating cases. The line in (a) meets zero amplitude at the threshold , and the line in (b) is drawn from there. The cases from 170 K on branch or form several cracks and have neither parameter.
Table 4 compares the onset with the other computations and with experiment. The threshold lies inside the interval bracketed by the straight and the oscillating paths of Zeng and Dolbow [11] on the same strip, although in a run at the slightly higher Péclet number
Table 4.
Table 4. Onset of oscillation on the glass strip at . Zeng and Dolbow give the interval of their computations at this speed, and their wavelength is read here from their lowest oscillating path, at . The values of Corson et al. are read from their Figure 6 at . The finite element computation is that of Appendix § 13 at 0.05 mm per step, with its wavelength at . The experimental value is the fit of Ronsin and Perrin, (Appendix § 14).
their computation already oscillates at . The threshold of the proposed method lies about 13% above , 7% above that of the linear stability analysis of Corson et al. [55] for a sharp crack and below the finite element threshold for the same model. The wavelength at onset lies 11% below that of Corson et al., and at the line of Figure 11(b) gives , about 1% below the 0.53 read from the path of Zeng and Dolbow. The onset wavelengths of the proposed method and of Corson et al. lie 26% and 17% below the experimental value, which is measured on the real temperature field rather than the idealized front of Eq. (25).
Quenched ceramic slab
With this problem a thermal transient computed by the temperature network drives the spontaneous nucleation and selection of a crack array. A ceramic slab 50 mm long and 10 mm wide, at the uniform stress-free temperature , is quenched into water at K through all four faces, and the thermal load is . A periodic array of cracks nucleates along the cooled faces, and a subset of them propagates into the depth. Geometry and material follow Tang et al. [112], who modeled the alumina of the experiments of Jiang et al. [3], except that the toughness is that of Bourdin et al. [44]. The benchmark is the computation of Bourdin et al. on a slab of nearly the same proportions, which they compared with the experiment of Shao et al. [45]. One quarter, 25 by 5 mm, is computed, Figure 12, under plane stress with the material of Table 5, mm and no mechanical load. The symmetry cuts carry zero normal displacement and are insulated, and the cooled faces are traction free.
Figure 12. Geometry and boundary conditions of the quenched slab. The computed quarter is shaded, held at the bath temperature along its two outer faces and bounded by the two insulated symmetry cuts. The temperature shading is schematic.
| Property | Symbol | Value |
| Young’s modulus (GPa) | 370 | |
| Poisson’s ratio | 0.30 | |
| Toughness (J m) | 42.47 | |
| Expansion coefficient (K) | ||
| Conductivity (W m K) | 31 | |
| Density (kg m) | 3980 | |
| Specific heat (J kg K) | 880 | |
| Film coefficient (kW m K) | 50 | |
| Nominal strength (MPa) | 180 | |
| Weibull modulus | 15 | |
| Regularization length (mm) | 0.05 |
Table 5. Material data of the quenched ceramic, after Tang et al. [112] with the toughness of Bourdin et al. [44], used by the slab and the annulus under plane stress. The regularization length is chosen here. The film coefficient, chosen within the measured range, enters only the computations under the convective condition, in Appendix § 10, Appendix § 11 and Appendix § 15.
A uniformly cooled slab in a homogeneous model develops a uniform damage band that bifurcates to a periodic mode and localizes into cracks [9, 44], and a computation leaves the uniform state only through a perturbation of its symmetry. The quench problems take this perturbation from the random material field of Tang et al. [112], who regard the scatter of the local stiffness and strength as intrinsic to the ceramic. Square cells of side each receive independent Weibull draws of Young’s modulus and strength , of Weibull modulus with the nominal values of Table 5 as scale parameters. This Weibull modulus is the value that Tang et al. adopt, not one measured on specimens. Since the AT2 peak stress of § 4.3 is fixed by , and , the drawn strength enters through the toughness,
which is rescaled by one constant per realization so that its mean equals the nominal and the Griffith length is the same for every realization (Appendix § 9). The finite element computation evaluates the same realization at its quadrature points.
The cooled surfaces are held at the bath temperature from on, the limit of the convective condition, so the temperature is prescribed on the boundary and computed in the interior. The conductivity is damage-independent since the cracks run along the heat flux. Both choices follow Bourdin et al. [44], Chu et al. [60], Mandal et al. [62] and Tangella et al. [63]. The temperature is advanced on a time grid that concentrates the load steps in the early transient,
with load steps from s to s after the initial state at , so the first step forms a cooled layer of thickness mm. Four thermal loads, , 380, 480 and 580 K, are computed, each on one realization of the field and each beside a finite element computation on that realization (Appendix § 12). Crack counts and the spacing are read along lines parallel to the cooled long face at depths from 0.1 to 2 mm (Appendix § 9).
Since the temperature does not depend on the cracks, halfway along the long face it obeys the one-dimensional heat equation normal to that face. At K the temperature network agrees with a backward Euler solution of this transient on the same time steps to within 1.6% of at every load step, and at 480 K to within 1.4% (Appendix § 9).
Under the prescribed boundary temperature a uniform damage band forms in the cooled layer in the first load steps and breaks into a dense array of short cracks, whose number along the face stays fixed from load step 20 on, so the later steps select which cracks continue into the depth. Figure 13 shows the final states at the four loads beside the finite element computations and the photographs of plates of this ceramic quenched by Jiang et al. [3]. All three show parallel cracks normal to the cooled long faces with a hierarchy of lengths, a spacing of the long cracks that grows with the depth and decreases with the load, and shorter cracks from the end faces. The two computations differ in the positions of the cracks, and both show short cracks between the long ones that are difficult to observe in the photographs, as Tang et al. [112] note for their computation.
Figure 13. Quenched slab under a prescribed boundary temperature, final crack patterns of the 50 by 10 mm slab at four thermal loads. In each group, the proposed method (top) and the finite element computation on the same realization (middle) show the computed quarter reflected about its two symmetry cuts, with the phase field from 0 (black) to 1 (light). The bottom row is the photograph of a plate quenched from into a bath at 293 K by Jiang et al. [3], the cracks revealed by dye. Photographs reprinted from Acta Materialia, Vol. 60, C.P. Jiang et al., A study of the mechanism of formation and numerical simulations of crack patterns in ceramics subjected to thermal shock, Pages 4540–4550, Copyright (2012), with permission from Elsevier.
Bourdin et al. [44] report that the developed cracks propagate under the control of the Griffith length alone, so the spacing against the depth in units of should follow one curve at every load. Figure 14 compares it with their curve and with the experiment of Shao et al. [45]. Averaged geometrically over the depths covered by both curves, the spacing of the proposed method lies within 14% of the benchmark at every load, smaller at 280 K and larger at 580 K (Appendix § 9). That of the finite element computation lies 11% to 23% above the benchmark, and that of the proposed method 17% to 25% below that of the finite element computation. Since the benchmark uses a gradient damage model with dissipation linear in the damage on a homogeneous material, the agreement of the proposed method with the benchmark does not show which solver is more accurate. The benchmark tests the scaling with the Griffith length, and the finite element computation tests the solution of the same AT2 model up to the treatment of irreversibility. At 380 K three solver seeds on the same field, which differ in the sampler and the network initialization, give spacings from 5% to 11% below the benchmark. With every iteration budget doubled the spacing lies 1.8% below it, outside the range of the seeds and toward the finite element computation (Appendix § 10), so part of the difference from the finite element computation at this load depends on the budget.
Figure 14. Quenched slab under a prescribed boundary temperature, crack spacing against the depth below the cooled long face at the four thermal loads, both normalized by the Griffith length (Appendix § 9). The proposed method (blue, solid) and the finite element computation (orange, dashed) are shown beside the benchmark of Bourdin et al. [44] (black) and the experiment of Shao et al. [45] (green circles), both digitized from the figure of Bourdin et al.
Table 6 lists the counts against the depth. The counts of the two computations differ by at most 35% down to 0.5 mm at the three higher loads, while at 280 K the proposed method has nearly twice as many of the shallowest cracks, and from 1 mm on it has 1 to 6 more cracks at every load. With the bound of the finite element computation at 380 K restricted to the crack set , the rule of the proposed method, the finite element counts change by at most 3 and the difference remains, 2 to 7 cracks from 1 mm on.
| (mm) | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 | 0.75 | 1.0 | 1.5 | 2.0 |
| 280 K, proposed | 36 | 27 | 24 | 19 | 16 | 15 | 10 | 9 | 8 |
| 280 K, finite element | 20 | 19 | 18 | 16 | 14 | 11 | 9 | 6 | 5 |
| 380 K, proposed | 46 | 31 | 27 | 22 | 20 | 16 | 15 | 14 | 9 |
| 380 K, finite element | 38 | 31 | 23 | 19 | 17 | 14 | 9 | 8 | 7 |
| 480 K, proposed | 50 | 39 | 30 | 27 | 21 | 19 | 17 | 12 | 9 |
| 480 K, finite element | 51 | 36 | 24 | 20 | 19 | 16 | 13 | 9 | 7 |
| 580 K, proposed | 57 | 41 | 34 | 27 | 23 | 20 | 19 | 13 | 11 |
| 580 K, finite element | 56 | 37 | 30 | 24 | 19 | 17 | 13 | 9 | 8 |
Table 6. Quenched slab under a prescribed boundary temperature, number of cracks along the cooled long face that reach the depth , the proposed method against the finite element computation on the same realization of the field. A crack is a local maximum of the phase field above 0.7, counted on the computed quarter over the 22.5 mm of the face more than 2.5 mm from the end face (Appendix § 9).
The two experiments differ at depth. On the photographs of Jiang et al., Ruan et al. [64] counted 32 cracks longer than 1 mm on the central 30 mm of a long face at 380 K. The spacing measured by Shao et al. corresponds to 17 such cracks, and under the same rule the proposed method counts 20, between the two experiments, and the finite element computation 12. The agreement with experiment is qualitative.
Under the convective condition with a film coefficient of the measured magnitude [127, 128, 129], both computations nucleate later and form sparser arrays near the surface (Appendix § 15). Under the convective condition the proposed method forms 16 to 27 cracks near the surface at the four loads against 3 to 9 for the finite element computation, a difference that arises at nucleation. This condition is closer to the experiments, since the Biot number is only about 0.2 at load step 20, ms, when the array under the prescribed boundary temperature has formed. The comparison with experiment above therefore rests on the idealization of a prescribed boundary temperature.
Quenched ceramic annulus
With this problem we apply the quench to a curved domain. Quenched circular ceramic specimens develop periodic and hierarchical crack patterns in experiment [130]. An annulus with a bore of radius 2.5 mm and an outer radius of 10 mm, of the ceramic of Table 5, uniform at , is quenched through both rims, Figure 15. One quarter is computed on a single quadratic NURBS patch that represents both rims exactly, with no mesh. The radial cuts carry zero normal displacement and are insulated, and both rims are traction free and held at the bath temperature. The material field, the damage-independent conductivity, the time grid, the loads and the finite element computation on the same realization follow the slab, and counts are read along arcs at the depth from each rim (Appendix § 9).
Figure 15. Geometry and boundary conditions of the quenched annulus. The computed quarter is shaded, held at the bath temperature along both rims and bounded by the two insulated radial symmetry cuts. The temperature shading is schematic.
Figure 16 shows the final states, one thermal load per quadrant. Both computations form an array on each rim, with long cracks running inward from the outer rim and outward from the bore and shorter cracks between them, and several long cracks lie at the same angles with similar lengths in the two. At depths of 1 mm and more the counts agree within 2 cracks on both rims at every load (Table 7). Nearer the surface, where nucleation sets the spacing, the proposed method forms more cracks, about twice as many at 0.1 mm from the bore below 580 K. At 580 K the proposed method leaves a fully damaged layer about 0.25 mm thick along the bore, which the finite element computation does not form (Appendix § 10). The spacing of the proposed method averaged over the depth lies 10% to 17% below that of the finite element computation on the outer rim and 19% to 36% below at the bore, Figure 17. The longest outer cracks of the finite element computation reach 3 mm at the two higher loads, beyond those of the proposed method.
| bore | outer rim | |||||||||
| (mm) | 0.1 | 0.5 | 1.0 | 2.0 | 3.0 | 0.1 | 0.5 | 1.0 | 2.0 | 3.0 |
| 280 K, proposed | 7 | 5 | 3 | 1 | 0 | 22 | 11 | 7 | 2 | 0 |
| 280 K, finite element | 3 | 3 | 3 | 0 | 0 | 14 | 9 | 5 | 2 | 0 |
| 380 K, proposed | 16 | 4 | 4 | 1 | 0 | 28 | 14 | 8 | 4 | 0 |
| 380 K, finite element | 8 | 4 | 3 | 2 | 0 | 24 | 9 | 7 | 3 | 0 |
| 480 K, proposed | 19 | 5 | 4 | 2 | 0 | 37 | 14 | 10 | 5 | 0 |
| 480 K, finite element | 10 | 5 | 4 | 1 | 0 | 35 | 11 | 8 | 4 | 1 |
| 580 K, proposed | 17 | 6 | 5 | 3 | 0 | 46 | 16 | 10 | 6 | 0 |
| 580 K, finite element | 12 | 5 | 4 | 2 | 0 | 39 | 14 | 8 | 4 | 2 |
Table 7. Quenched annulus, number of cracks cut by the arc at the depth from each rim, counted at the level 0.7 as on the slab, the proposed method against the finite element computation on the same realization of the field. At 580 K the arc at 0.1 mm from the bore lies inside the fully damaged layer of the proposed method.
Figure 16. Quenched annulus, final phase fields at the four thermal loads assembled into one ring, one load per quadrant counterclockwise from the top right. (a) The proposed method. (b) The finite element computation on the same realization. The phase field ranges from 0 (black) to 1 (light).
Figure 17. Quenched annulus, crack spacing against the depth from the bore (top) and from the outer rim (bottom) at the four thermal loads, both normalized by the Griffith length (Appendix § 9). The proposed method (blue, solid) is compared with the finite element computation (orange, dashed).
Conclusions
This paper proposes a mesh-free deep energy method for phase-field modeling of thermally driven brittle fracture, built on the purely elastic solver of [72]. Unlike the existing neural energy-based approaches to thermal fracture, which start from defects placed in advance, the proposed method computes the spontaneous nucleation of cracks and the fracture patterns that develop under transient thermal loading. It captures the emergence of crack arrays, their propagation and arrest, and complex crack trajectories without prescribing crack paths or nucleation sites and without meshes adapted to the cracks. Where the temperature is not given as an analytical field, it is computed by a second network from an incremental functional of the heat equation, in which the convective condition is a natural boundary condition. Irreversibility on the crack set, together with an optimizer state carried across load steps and learning rates that decay within each step, controls the spurious accumulation of damage.
The performance of the proposed method was assessed on five numerical examples. The temperature was computed with a damage-dependent conductivity on the notched plates, given as an analytical field on the strip, and computed with a damage-independent conductivity on the slab and the annulus. On the two notched plates the initiation peak loads agree with an independent finite element reference to within 1.8% at every thermal load, and on the graded plate the initiation peak rises by 26% from −30 to +30 K, against 28% in the reference. On the glass strip the crack path changes from straight to oscillatory and then to branched as the temperature jump grows, in the order found by Zeng and Dolbow [11], with an oscillation threshold close to the linear stability prediction. On the quenched slab under a prescribed boundary temperature the crack array is of the type observed in experiment, with a spacing averaged over the depth within 14% of the benchmark computation of Bourdin et al. [44]. On the slab and the annulus the finite element computation on the same material field forms a pattern of the same type, and the proposed method forms a denser array.
Several limitations remain. The results depend on the resolution of the neural representations and on the optimization procedure, and a systematic convergence study has not been made. A larger iteration budget changes the crack spacing on the slab, and on the strip the paths near the transitions between propagation regimes change with every increase of the budget up to the one used. The settings that control the spurious accumulation of damage were tested under limited conditions and do not remove every such artifact. A damaged layer persists along the bore of the annulus at the largest thermal load. On the slab under the convective condition (Appendix § 15) the proposed method forms several times more cracks near the surface than the finite element computation, a difference that arises at nucleation. The results for the quench problems rest on single realizations of the material field, and the strength of the AT2 model is tied to the regularization length. The formulation is restricted to two-dimensional, quasi-static fracture, and the thermal problem of a step is solved before, and not iterated with, the mechanical one. The thermo-elastic heat source, the heat released by fracture [59] and the contact of the crack faces are not modeled, although a heat source from the rates of the volumetric strain and of the damage changes the crack spacing in a gradient damage model of a quenched ceramic plate [119]. The crack band is either nearly adiabatic or transmits heat as intact material, without a model of the effective conductivity of a cracking body [131]. The proposed method also takes longer than the finite element computation on the notched plates (Table A.1).
Future work should improve the numerical convergence, the robustness and the computational efficiency, in particular for problems with the spontaneous nucleation of many cracks. Further developments include a fuller thermo-mechanical coupling, the heat transfer across opening cracks and the extension to thermo-hydro-mechanical fracture. Since the material field, the network fields and the energy enter one differentiable computation, a spatially varying material property could also be identified from an observed fracture pattern, which requires reliable sensitivities of the computed fracture evolution to that property.
Data availability
Data will be made available on request.
Code availability
The solver, the staggered finite element reference implementation and the scripts reproducing all examples of this paper are available at https://github.com/HannnZH/Meshfree-DEM-ThermoMechanical-PhaseField.
Acknowledgments
This research was undertaken with the assistance of resources from the Gadi supercomputer [132] at the National Computational Infrastructure (NCI Australia), an NCRIS enabled capability supported by the Australian Government, and from the Katana cluster [133] at UNSW Sydney, supported by Research Technology Services.
Numerical parameters and verification of the reported runs
Table A.1 collects the numerical parameters of every run reported in § 4. All runs use the multilayer perceptrons of Eq. (14) with GELU activations, 4 hidden layers of width 128 in the mechanical network and 3 of width 64 in the temperature network. Every level of the feature grids stores 2 feature values at each node. Adam is used with the moment parameters 0.9 and 0.999, the stability constant and a learning rate of on the perceptron and on the grids. Where the optimizer state is carried across load steps, the step counter of Adam is carried with it, so the bias correction is not restarted. The thermal solve starts each step with a new Adam and no decay of the learning rates. The integration points are redrawn at every iteration, each stratum receiving a fixed share of them, and the stratum weights , and of Eq. (21) are 0.4, 0.3 and 0.3 for the uniform, crack and process-zone strata on the notched plates and the strip and 0.3, 0.3 and 0.4 on the quench problems. The first thermal solve has the largest budget, since it forms the boundary layer from the initial state of the lift. All runs are in double precision on one NVIDIA V100 GPU of Gadi, except the strip, which ran on one NVIDIA H200 GPU of Katana as consecutive jobs whose times are summed in the table. The finite element references of the notched plates ran on 8 cores of a CPU node of Gadi.
The remaining numerical parameters are common to all runs unless stated otherwise. The residual stiffness is , the displacement scale , and the temperature scale of Eq. (17) is the magnitude of the thermal load, or 1 K where it vanishes. A quadratic penalty on the feature values, the sum of their squares multiplied by , is added to each estimated loss. The bias of the phase-field output is initialized at , so that at the start in material without a notch. The decay length of the temperature lift in Eq. (18) is mm. The finest mechanical level has a node spacing of about on the plates, the slab and the strip, and on the annulus the physical spacing grows with the radius. On the notched plates the reaction force of Eq. (24) is estimated on 80 000 points.
Each solve stops when
or when its iteration budget is spent, with the current iteration, the window of iterations, the value of or of at iteration and the tolerance .
On the notched plates, whose parametric and physical coordinates coincide, the uniform stratum of Eq. (21) is drawn from a scrambled Sobol’ sequence [134], and on the mapped patches from the auxiliary grid. That grid has cells on the notched plates, on the slab and the strip and on the annulus. The crack and process-zone strata are held constant within each cell and normalized by the sum over the cells, and a point is drawn by choosing a cell with its probability and a uniform position inside it. On the quench problems the crack stratum is a Gaussian band along each cooled face or rim with a standard deviation of mm normal to it. The process-zone stratum uses and on the driving-force indicator
divided by its maximum over the cells, with the tensile part of the volumetric-deviatoric split of Amor et al. [35]. This tensile part serves the sampling only and is distinct from the spectral part that drives the crack. Since the estimator of Eq. (19) divides by the sampling density, this choice changes the variance of the estimate and not its expectation. As in [72], the volumetric-deviatoric form is used since it is smooth and, under shear, broader than the spectral one, so that the process zone is not left without points. On the slab under the convective condition of Appendix § 15, each of the two cooled faces has points of Eq. (20).
On the notched plates, with the bottom edge fixed and the top edge driven, the lifts read
with , and under tension and under shear, with the reaction force taken with respect to the nonzero component. In a profile of load step 50 of the slab at 380 K the backward pass takes 71% of a mechanical iteration, and at full budgets the thermal solve takes about 8% of a load step. A perceptron of 3 hidden layers of width 64 halves the time of a mechanical iteration on the slab. Its effect on the accuracy was measured on the cooled tension plate, where the same perceptron shortens the run by 3% and raises the error of the peak load from 1.30% to 3.6%, so the reported runs keep the wider perceptron.
Settings, measurements and thermal verification
The notched plates and their finite element references were computed with the volumetric heat capacity and the time scale both divided by , as reported by Tangella et al. [63], which leaves the product and every field unchanged, and the times reported are those of the physical problem.
Table A.1. Numerical parameters of the reported runs. Levels are the numbers of nodes of the three levels of feature grids along , from coarse to fine, and along each grid has as many nodes as the nearest integer to times the aspect ratio of the patch, so that the physical node spacing is about the same in the two directions. is the number of integration points per iteration. The mechanical iteration budgets are those of the first step, the second step and every later step, and the thermal budgets those of the first transient solve and of every later step. is the penalty weight of Eq. (7), and the parameter counts of the two networks include the feature values. The temperature is computed or given as an analytical field, and the conductivity is damage-dependent, Eq. (11), or damage-independent. The strip has no temperature network, since its temperature is an analytical field. The wall time of the proposed method is that of one run, and that of the finite element computation the range over the thermal loads. The rows below the dashed line refer to the finite element computations.
| Tension | Shear, graded | Strip | Slab | Annulus | |
| Temperature | computed | computed | analytical | computed | computed |
| Conductivity | |||||
| (mm) | 0.01 | 0.01 | 0.1 | 0.05 | 0.05 |
| Levels, mechanical | 32/128/384 | 32/128/384 | 64/256/624 | 64/192/384 | 64/256/1280 |
| Levels, thermal | 32/128/384 | 32/128/384 | 64/128/384 | 64/256/384 | 64/256/320 |
| 12 000 | 12 000 | 100 000 | 80 000 | 80 000 | |
| Iterations, mechanical | 1000/4000/2000 | 1000/4000/2000 | 10 000/5000/5000 | 1500/3000/2400 | 1500/3000/2400 |
| Iterations, thermal | 1200/400 | 1200/400 | 1600/800 | 1600/800 | |
| (MPa) | 20.6 | ||||
| Load steps | 80 | 160 | 240 | 100 | 100 |
| Parameters, mechanical | 380 803 | 380 803 | 1 772 355 | 1 935 235 | 2 661 379 |
| Parameters, thermal | 338 689 | 338 689 | 1 688 321 | 271 745 | |
| Wall time (h) | 5.0 | 10.4 | 25.6 | 16.8 | 17.0 |
| Finite element computation | |||||
| Element size | |||||
| Wall time (h) | 2.3 to 2.6 | 4.6 to 6.3 |
Table A.1. Numerical parameters of the reported runs.
On the quench problems the rescaling of the toughness field in Eq. (27) multiplies the strength implied by each cell by the square root of one constant, so the local strengths keep the shape of the drawn distribution. With the nominal values at mm the peak stress is 182 MPa, close to the nominal strength of 180 MPa. The Griffith length is written by Bourdin et al. [44], and at K the slab matches their setup to 4% in and to 2% in the half thickness over .
On the slab, crack counts are the local maxima of the phase field above the level 0.7 along a line at the depth parallel to the cooled long face, merged within 0.15 mm, over the part of the face more than 2.5 mm from the cooled end face. The spacing is the length of that part divided by the count. On the annulus the same rule is applied along the arc at the depth from each rim, merged within 0.15 mm of arc. The benchmark is the curve of Bourdin et al. [44] digitized from their Figure 1(c). As a check on the digitization, the semianalytical law of Bahr et al. [8], digitized from the same axes, gives the exponent 0.745 against its . A computed spacing is compared with the benchmark, and with the finite element computation, through the ratio of the two curves, averaged geometrically over logarithmically spaced depths in the range that both curves cover. The count under the rule of Ruan et al. [64] is the number of cracks reaching 1 mm on the 15 mm of the quarter next to its symmetry cut, doubled for the central 30 mm of the long face.
On the cooled tension plate at load steps 20, 36 and 37, the last at the peak, and on the isothermal plate at its peak, the resultant of the hybrid stress along the loaded edge differs from the reaction force of Eq. (24) by 1.6% to 4.0%. The reported force is the domain integral of the energy derivative, and the edge resultant, which samples the stress on one line, serves only as a check on equilibrium. The violation of irreversibility, , measured on a grid of 600 600 points, is at most on the tension plate at the same steps and at most on the crack set of the slab at 380 K. Outside the crack set the phase field decreases by up to 0.11 at nucleation, as the threshold allows.
The one-dimensional reference of the tension plate is an implicit finite difference solution on 2001 nodes with 200 substeps per load step. That of the slab is a backward Euler finite volume solution on 20 000 cells on the time grid of Eq. (28), along the line normal to the long face halfway along it, over the 100 load steps (§ 4.4). At 380 K the first backward Euler step leaves the layer up to 47 K warmer than the continuous transient, an error that decays to 5.5 K by load step 16 and that the finite element computation shares.
Solver settings, resolution and budget on the crack arrays
Table B.1 compares subsets of three settings on the slab of § 4.4 at K, with all other settings as in Table A.1. The settings are the crack-set threshold of Eq. (7), at , a value that was not varied, and the two options of the mechanical update in § 3.3 (the optimizer state carried across load steps and the decay of the learning rates within a step). With none of the three, 84 of the 100 load steps start with a rise of the loss, and a diffuse background and a damaged film accumulate along the cooled face. The threshold alone, the carried optimizer alone and the threshold with the decay still widen the bands or form a fully damaged layer, and only the three together keep the bands at a nearly constant width and the cooled face clean. The threshold with the carried optimizer and without the decay, tried on the slab under the convective condition of Appendix § 15, also widened the bands and formed a film along the end face. The decay alone and the carried optimizer with the decay were not computed.
| Subset | Cooled face | Width (mm) | Count | Rising steps |
| none | film and background | 0.125, 0.161 | 50, 50 | 84 |
| threshold | clean | 0.100, 0.123 | 48, 50 | 0 |
| carried optimizer | layer of 0.2 mm, background | 0.117, 0.226 | 48, 63 | 0 |
| threshold and decay | fully damaged layer, bulges | 0.127, 0.237 | 58, 59 | 74 |
| all three | clean | 0.091, 0.096 | 50, 50 | 1 |
Table B.1. Slab under a prescribed boundary temperature at K, one realization, subsets of the crack-set threshold and the two options of the mechanical update. State of the cooled face at the last load step, width at half maximum of the bands at the depth of 1 mm at load steps 55 and 100, count along the cooled face at the depth of 0.1 mm at load steps 50 and 100, and number of load steps whose loss rises by more than 15% in the first 300 iterations.
On the annulus of § 4.5 at K the bore retains a fully damaged layer, about 0.25 mm thick on the reported realization of the field and about 0.1 mm on a second one. A linear ramp of the learning rates over the first 200 iterations of every load step thins it to about 0.1 mm and changes the counts by at most 2 cracks.
The resolution and the iteration budget were tested on the slab at K, on the reported realization. The three solver seeds of § 4.4, which measure the scatter of the solver and not that of the material field, differ in their counts by at most 5 cracks at any depth. With the number of integration points and the node counts on every mechanical level and on the two finer thermal levels doubled, the crack counts stay within the range of the seeds except for one crack more at 2 mm, and the spacing lies 8.8% below the benchmark of Bourdin et al. [44]. With every iteration budget doubled, the counts lie one or two cracks below that range between 0.3 and 0.75 mm, and the spacing lies 1.8% below the benchmark, against 5.0% to 10.5% for the three seeds.
Sampler and representation controls
These controls, on the slab of § 4.4, test two choices of § 3, the redrawing of the points at every iteration and the feature grids of the temperature network, and check that the process-zone stratum, which follows the mechanical state, does not bias the thermal solve. On the slab the finest thermal level has a node spacing of 0.013 mm to resolve the layer of the first time step. That is finer than the mean distance between the 80 000 points of a fixed set even in the band along the cooled face, where the crack stratum concentrates them to about 0.02 mm, so a fixed set leaves that level unconstrained between the points.
| Variant | Parameters | 1 to 30 | 31 to 100 | Layer rms (K) | Layer integral (%) |
| Thermo-mechanical run | 1 688 321 | 1.35 to 3.30 | 1.67 to 3.29 | 0.22 to 0.69 | −0.37 to +0.50 |
| Redrawn points | 1 688 321 | 1.75 to 3.32 | 1.58 to 3.06 | 0.25 to 0.69 | −0.44 to +2.19 |
| Fixed set per step | 1 688 321 | 150 to 1434 | 121 to 203 | 30 to 158 | −10.3 to +1804 |
| No feature grids | 8577 | 11.3 to 61.5 | 1.66 to 8.79 | 0.54 to 9.29 | −13.7 to +0.40 |
Table C.1. Thermal controls on the quenched slab under the convective boundary condition, = 380 K, against the backward Euler finite volume solution. For each variant, the range over the load steps of three quantities, the largest departure (load steps 1 to 30 and 31 to 100), the root mean square weighted by area over the cooled layer and the departure of the temperature drop integrated over that layer (all 100 load steps). The thermo-mechanical run is a computation on the realization of Table 6 at this load under the same boundary condition.
The thermal problem of the slab alone, whose temperature does not depend on the cracks since the conductivity is damage-independent, was computed over the 100 load steps at = 380 K under the convective condition of Appendix § 15, with the budgets, the learning rates and the point count of a thermo-mechanical run, and compared with a backward Euler finite volume solution along the line normal to the long face halfway along it, Table C.1. On points redrawn at every iteration with no mechanical state in the sampler, the network reproduces the temperature of the thermo-mechanical run, so the process-zone stratum does not bias the thermal solve. On one point set per time step the temperature departs by at least 121 K at every load step, while the functional estimated at the fixed points lies below that of the redrawn run at 93 of the 100 steps, the mechanism of [72]. Without feature grids the departure reaches 16% of the thermal load while the cooled layer is thinner than about 0.3 mm. The effect of the representation on the crack pattern was not computed.
Verification of the finite element reference
The reference for the two notched plates (§ 4.1 and § 4.2) is a staggered phase-field solver built on scikit-fem [125]. It solves the plane-strain problem with the AT2 density and the hybrid formulation, in which the equilibrium equation uses the degraded isotropic energy and the phase field is driven by the tensile part of the spectral decomposition. Irreversibility is enforced through the history field, and the notch enters as a Dirichlet condition on the phase field. Each load step is solved by alternate minimization to a staggered tolerance of on the relative change of the nodal phase field in the Euclidean norm, with a cap of 2000 staggered iterations on the notched plates, which was not reached. The mesh is a uniform grid of bilinear quadrilaterals, 512 on a side after 9 refinements of the unit square, an element size of . The heat conduction problem is advanced by backward Euler on the time grid of § 4.1 with a lumped heat capacity, which keeps the temperature inside the range of its boundary values at the steep front where the consistent capacity does not. The conductivity is damage-dependent, Eq. (11). The thermal problem is solved first at every step and the mechanical problem on the frozen temperature, which is the order of the proposed method. The thermal strain enters through the elastic strain, and with and the in-plane stress of the undamaged material is
| Fraction of the height | 0.10 | 0.15 | 0.20 | 0.30 | 0.40 | 0.50 | 0.60 |
| Tangella et al., Fig. 11(c) | 54 | 41 | 32 | 22 | 16 | 14 | 13 |
| Finite element, their staggered loop | 39 | 30 | 23 | 19 | 16 | 14 | 11 |
| Finite element, alternate minimization | 36 | 29 | 21 | 18 | 15 | 14 | 11 |
Table D.1. Reproduction of the ceramic quenching test of Tangella et al. [63] at K and 50 s. Number of cracks along the cooled long face that reach the given fraction of the plate height, read from their Fig. 11(c) and from the finite element fields of Fig. D.1 by one rule.
the plane-strain law of a body carrying the isotropic in-plane eigenstrain . The spectral decomposition is the one of § 2.1 under plane strain and plane stress alike, and the graded plate carries the linear material fields of Table 2 at the quadrature points. The mesh was checked on the homogeneous plate under shear at K of Tangella et al. [63], where the peak falls from 79.9 N at 7 uniform refinements (elements of ) to 76.8 N at 8 () and to 75.9 N at 9 (), the refinement of the reference. At 9 refinements the crack initiates one load step earlier than at 8. At 7 refinements the plate separates at a displacement 2% below their published value, with mm against their 0.005 mm.
Added to the reference as an option, evaluated on the frozen temperature, the toughness law of Eq. (4) raises the peak of the cooled tension case by 0.23%, from 96.33 to 96.55 N.

Figure D.1. Finite element reproduction of the ceramic quenching test of Tangella et al. [63] at K and 50 s, quarter model with the cooled long face at the bottom and the cooled end face at the left, with the phase field from 0 (black) to 1 (light). (a) Their staggered loop. (b) Alternate minimization.
The same solver was run on quenched ceramic plates in two roles. The first is a reproduction of the ceramic quenching test of Tangella et al. [63] at K. It uses their nearly homogeneous material field, time step of s up to s and regularization length of 0.06 mm, with the bath temperature prescribed on the cooled faces, a damage-independent conductivity and elements of side , Fig. D.1 and Table D.1. Their staggered loop, following their released code [126], iterates the displacement and the phase field until both change by less than , with no cap. The cracks reaching 0.4 of the height and more agree with their Fig. 11(c) in number to within 2 and in depth. The end face has 4 short cracks in their field and in both finite element fields, while along the long face the finite element fields have fewer shallow cracks, 39 and 36 against 54 at 10% of the height. The total count depends on the time step, and no mesh refinement was made on this problem.
The second role is the computation of the same model on the slab and the annulus, under plane stress on the realization of the proposed method, with the bath temperature prescribed on the cooled boundary, a damage-independent conductivity and the time grid of Eq. (28). It enforces the bound on the phase field by an active set and uses bilinear quadrilaterals of side , on the annulus a tensor grid in radius and angle. Each step is solved by alternate minimization to a staggered tolerance of , with a cap of 50 iterations on the slab and 1000 on the annulus. On the slab at 380 K elements of side change the counts of Table 6 by at most 1 crack and the spacing at , 20, 30, 50 and 70 by at most 8%. At 480 K a first step 100 times shorter changes the counts by at most 2 cracks and the spacing at the same depths by less than 6%.
Loading increment and iteration budget on the glass strip
The loading increment on the strip (§ 4.3) was chosen on the finite element computation of the same model. The solver of Appendix § 12 was run on the strip under plane stress with the analytical temperature field of Eq. (25) and elements of side , at two increments for each class of path, Table E.1. Between 0.1 and 0.05 mm per step at 130, 145, 165 and 190 K, and between 0.2 and 0.1 mm per step at 217 and 307 K, no path changes class, the oscillation at 165 K keeps its order parameters to within 5%, and the forks and the cracks move by at most 0.1 mm. The finite element computation is therefore insensitive to the increment at 0.1 mm per step, one regularization length, the increment of § 4.3.
| (K) | 0.2 | 0.1 | 0.05 | 0.1 mm per step | |
| 130 | 8.24 | straight, 0.11 mm | straight, 0.04 mm | straight, 0.09 mm | |
| 145 | 9.19 | straight, 0.06 mm | straight, 0.10 mm | 0.0638, 0.510 | |
| 165 | 10.46 | 0.1047, 0.540 | 0.1094, 0.542 | 0.1742, 0.633 | |
| 190 | 12.05 | fork, mm | fork, mm | fork, mm | |
| 217 | 13.76 | fork, mm | fork, mm | fork, mm | |
| 307 | 19.46 | 3 cracks | 3 cracks | 3 cracks |
Table E.1. Loading increment on the glass strip, mm, travel 24 mm, both computations under the analytical temperature field of Eq. (25). Path at the last load step of the finite element computation at two increments and of the proposed method at 0.1 mm per step with 5000 iterations. The entries give for straight paths the largest transverse departure of the ridge from its position at mm, taken from there to the crack tip, for oscillating paths and on the developed path, for branched paths the position of the fork, and for multiple cracks their number.
Unlike the finite element computation, which is solved to a staggered tolerance of at every step with a cap of 200 iterations, the proposed method uses its full budget at every step of the strip, since the noise of the estimate exceeds the tolerance of Eq. (29), so its path depends on the budget as well as on the increment. Table E.2 lists the paths at three budgets Table E.2. Iteration budget and loading increment of the proposed method on the glass strip, ordered by the number of iterations per millimeter of travel, the first step carrying twice the budget in every run. The entries give for straight paths the largest transverse departure defined in Table E.1, for oscillating paths and on the developed path and for branched paths the position of the fork, all lengths in mm. At 145 K the ratio of the last half swing on the developed path to the largest follows in parentheses. At 307 K every setting forms three cracks. The reported runs are those at 0.1 mm per step with 5000 iterations.
| mm per step | Iterations | Iterations per mm | 130 K | 145 K | 165 K | 190 K | 217 K |
| 0.2 | 3000 | 15 000 | 0.04 | 0.08 | 0.1129, 0.587 | 10.8 | 5.4 |
| 0.1 | 1500 | 15 000 | 0.10 | 0.0142, 0.498 (0.43) | 0.1271, 0.613 | 8.0 | 5.3 |
| 0.1 | 3000 | 30 000 | 0.07 | 0.0250, 0.471 (0.58) | 0.1598, 0.617 | 7.9 | 5.2 |
| 0.05 | 2500 | 50 000 | 0.05 | 0.0556, 0.504 (1.00) | 0.1732, 0.626 | 5.4 | 5.1 |
| 0.1 | 5000 | 50 000 | 0.09 | 0.0638, 0.510 (0.94) | 0.1742, 0.633 | 5.5 | 5.1 |
| 0.05 | 3000 | 60 000 | 0.07 | 0.0681, 0.518 (0.93) | 0.1773, 0.630 | 5.4 | 5.1 |
Table E.2. Iteration budget and loading increment of the proposed method on the glass strip, ordered by the number of iterations per millimeter of travel, the first step carrying twice the budget in every run. The entries give for straight paths the largest transverse departure defined in Table E.1, for oscillating paths A/b and Λ/b on the developed path and for branched paths the position x of the fork, all lengths in mm. At 145 K the ratio of the last half swing on the developed path to the largest follows in parentheses. At 307 K every setting forms three cracks. The reported runs are those at 0.1 mm per step with 5000 iterations.
for 0.1 mm per step and at the increments of 0.05 and 0.2 mm per step. Budget and increment act mainly through the number of iterations per unit of travel, and up to 50 000 iterations per millimeter a larger number raises the amplitude at 165 K, sustains the oscillation at 145 K on the developed path and moves the fork at 190 K to the end of the notch. At 50 000 iterations per millimeter the two increments give the same class of path and forks within 0.1 mm, with order parameters within about 1% at 165 K and an amplitude 13% smaller at 0.05 mm per step at 145 K. The most expensive setting, 0.05 mm per step with 3000 iterations, changes no class of path and raises the amplitude by at most 7%, so the reported runs use 0.1 mm per step with 5000 iterations. Over the five settings at which both 145 and 165 K oscillate, the line through their squared amplitudes gives a threshold between 8.97 and 9.18.
Measurement of the order parameters
Table F.1 lists the successive half swings of the three oscillating cases on the strip (§ 4.3) over the whole path. The largest, the fourth, lies on the developed path in each case, and the threshold is 9.01 whether the amplitude is read on the developed or on the whole path. Leaving out one case gives 8.98 to 9.11, reading the amplitude as the mean of the three largest half swings gives 9.03, and as the last half swing of the whole path 9.18. The wavelength over the whole path differs from that on the developed path by less than 1% at 155 and 165 K and by 4.3% at 145 K. The experimental wavelength of Table 4 is the fit of Ronsin and Perrin [50], which reads with the width at onset and the thermal length , evaluated at .
Table F.1. Successive half swings of the traced path (mm) of the three oscillating cases, 0.1 mm per step with 5000 iterations, over the whole path, and the last half swing over the largest.
| (K) | Half swings (mm) | Last over largest | |||||||||
| 145 | 0.75 | 0.90 | 0.99 | 1.02 | 0.96 | 0.86 | 0.71 | 0.55 | 0.39 | 0.28 | 0.28 |
| 155 | 1.31 | 1.71 | 1.97 | 2.03 | 1.90 | 1.75 | 1.60 | 1.49 | 0.73 | ||
| 165 | 2.06 | 2.55 | 2.78 | 2.79 | 2.67 | 2.42 | 2.15 | 0.77 |
Table F.1. Successive half swings of the traced path (mm) of the three oscillating cases, 0.1 mm per step with 5000 iterations, over the whole path, and the last half swing over the largest.
Slab under a convective boundary
On the slab of § 4.4, under the film coefficient of Table 5, the cooled faces exchange heat with the bath by the convective condition of Eq. (12), a natural boundary condition of the functional of Eq. (13), and the problems of § 4.4 are otherwise unchanged. Table G.1 lists the counts. Film coefficients for alumina quenched into water peak at 15 to 60 [127, 128, 129], and the coefficient here is held constant at a value in that range. At 380 K the finite element computation forms its first cracks at load step 44, against step 11 under the prescribed boundary temperature. At every load it forms a sparse array of nearly equal cracks, unchanged at 380 K when the cap of its staggered loop is raised from 50 to 1000. Repeated at 380 K with film coefficients of 200, 1000 and 5000 , it forms an array that approaches that of the prescribed boundary temperature as the film coefficient grows.
| Computation | (K) | 0.1 | 0.5 | 1.0 | 2.0 |
| finite element | 280 | 3 | 3 | 3 | 3 |
| 380 | 5 | 5 | 5 | 4 | |
| 480 | 6 | 6 | 6 | 6 | |
| 580 | 9 | 9 | 9 | 6 | |
| proposed method | 280 | 16 | 16 | 11 | 5 |
| 380 | 17 | 13 | 11 | 7 | |
| 480 | 20 | 14 | 11 | 7 | |
| 580 | 27 | 17 | 13 | 9 | |
| finite element, | 380 | 12 | 12 | 11 | 5 |
| finite element, | 380 | 27 | 16 | 13 | 7 |
| finite element, | 380 | 35 | 18 | 10 | 7 |
| finite element, prescribed boundary temperature | 380 | 38 | 17 | 9 | 7 |
Table G.1. Quenched slab under the convective boundary condition, number of cracks along the cooled long face that reach the depth , on the realizations of Table 6. The film coefficient is that of Table 5 unless stated, and the last row of the sweep is the prescribed boundary temperature.
Under the convective condition both arrays are sparser near the surface, since the layer in tension is deeper when it cracks. At 380 K, when the cracks appear, the stress of a one-dimensional cooling falls to half its surface value at the depth of 0.05 mm under the prescribed boundary temperature and at 0.29 mm under the convective condition. The deeper layer releases its energy through fewer and longer cracks.
At 380 K the proposed method forms 17 nuclei in load step 47, in which its damage band localizes, and from 16 to 20 when the computation is restarted at load step 44 with 2 or 4 times the iterations or other optimizer settings. Started from the state of the proposed method at load step 44, before this localization, the finite element computation forms its own 5 cracks, while started from a state with 16 nuclei it retains them. The two solvers therefore differ at nucleation, and irreversibility then preserves the nuclei. With its bound applied on the crack set only, the finite element computation still forms 5 cracks near the surface, so within the variations tried none of the budget, the optimizer settings or the treatment of irreversibility removes the difference.
References
- [1]H.-A. Bahr, G. Fischer, and H.-J. Weiss. Thermal-shock crack patterns explained by single and multiple crack propagation. Journal of Materials Science, 21(8):2716–2720, 1986. doi:10.1007/BF00551478.DOI
- [2]Y. Shao, X. Xu, S. Meng, G. Bai, C. Jiang, and F. Song. Crack patterns in ceramic plates after quenching. Journal of the American Ceramic Society, 93(10):3006–3008, 2010. doi:10.1111/j.1551-2916.2010.03971.x.DOI
- [3]C. P. Jiang, X. F. Wu, J. Li, F. Song, Y. F. Shao, X. H. Xu, and P. Yan. A study of the mechanism of formation and numerical simulations of crack patterns in ceramics subjected to thermal shock. Acta Materialia, 60(11):4540–4550, 2012. doi:10.1016/j.actamat.2012.05.020.DOI
- [4]A. Yuse and M. Sano. Transition between crack patterns in quenched glass plates. Nature, 362(6418):329–331, 1993. doi:10.1038/362329a0.DOI
- [5]A. Yuse and M. Sano. Instabilities of quasi-static crack patterns in quenched glass plates. Physica D: Nonlinear Phenomena, 108(4):365–378, 1997. doi:10.1016/S0167-2789(97)00011–0.DOI
- [6]L. Goehring, L. Mahadevan, and S. W. Morris. Nonequilibrium scale selection mechanism for columnar jointing. Proceedings of the National Academy of Sciences, 106(2):387–392, 2009. doi:10.1073/pnas.0805132106.DOI
- [7]E. A. Jagla. Stable propagation of an ordered array of cracks during directional drying. Physical Review E, 65(4):046147, 2002. doi:10.1103/PhysRevE.65.046147.DOI
- [8]H.-A. Bahr, H.-J. Weiss, U. Bahr, M. Hofmann, G. Fischer, S. Lampenscherf, and H. Balke. Scaling behavior of thermal shock crack patterns and tunneling cracks driven by cooling or drying. Journal of the Mechanics and Physics of Solids, 58(9):1411–1421, 2010. doi:10.1016/j.jmps.2010.05.005.DOI
- [9]P. Sicsic, J.-J. Marigo, and C. Maurini. Initiation of a periodic array of cracks in the thermal shock problem: a gradient damage modeling. Journal of the Mechanics and Physics of Solids, 63:256–284, 2014. doi:10.1016/j.jmps.2013.09.003.DOI
- [10]T. J. Lu and N. A. Fleck. The thermal shock resistance of solids. Acta Materialia, 46(13):4755–4768, 1998. doi:10.1016/S1359-6454(98)00127-X.DOI
- [11]B. Zeng and J. E. Dolbow. A complete phase-field fracture model for brittle materials subjected to thermal shocks. arXiv preprint arXiv:2602.09031v2, 2026.arxiv.org/abs/2602.09031
- [12]D. P. H. Hasselman. Unified theory of thermal shock fracture initiation and crack propagation in brittle ceramics. Journal of the American Ceramic Society, 52(11):600–604, 1969. doi:10.1111/j.1151-2916.1969.tb15848.x.DOI
- [13]S. Nemat-Nasser, L. M. Keer, and K. S. Parihar. Unstable growth of thermally induced interacting cracks in brittle solids. International Journal of Solids and Structures, 14(6):409–430, 1978. doi:10.1016/0020-7683(78)90007-0.
- [14]Z. P. Bažant, H. Ohtsubo, and K. Aoh. Stability and post-critical growth of a system of cooling or shrinkage cracks. International Journal of Fracture, 15(5):443–456, 1979. doi:10.1007/BF00023331.DOI
- [15]J. F. Geyer and S. Nemat-Nasser. Experimental investigation of thermally induced interacting cracks in brittle solids. International Journal of Solids and Structures, 18(4): 349–356, 1982. doi: 10.1016/0020-7683(82)90059-2.DOI
- [16]A. A. Griffith. The phenomena of rupture and flow in solids. Philosophical Transactions of the Royal Society of London A, 221:163–198, 1921. doi: 10.1098/rsta.1921.0006.DOI
- [17]G. R. Irwin. Analysis of stresses and strains near the end of a crack traversing a plate. Journal of Applied Mechanics, 24(3):361–364, 1957. doi: 10.1115/1.4011547.DOI
- [18]T. N. Bittencourt, P. A. Wawrzynek, A. R. Ingraffea, and J. L. Sousa. Quasi-automatic simulation of crack propagation for 2D LEFM problems. Engineering Fracture Mechanics, 55(2):321–334, 1996. doi: 10.1016/0013-7944(95)00247-2.DOI
- [19]N. Moës, J. Dolbow, and T. Belytschko. A finite element method for crack growth without remeshing. International Journal for Numerical Methods in Engineering, 46(1):131–150, 1999. doi: 10.1002/(SICI)1097-0207(19990910)46:1<131::AID-NME726>3.0.CO;2-J.doi.org/10.1002/(sici)1097-0207(19990910)46:1<131::aid-nme726>3.0.co;2-j
- [20]T. Belytschko, Y. Y. Lu, and L. Gu. Element-free Galerkin methods. International Journal for Numerical Methods in Engineering, 37(2):229–256, 1994. doi: 10.1002/nme.1620370205.DOI
- [21]T. Rabczuk and T. Belytschko. Cracking particles: a simplified meshfree method for arbitrary evolving cracks. International Journal for Numerical Methods in Engineering, 61(13):2316–2343, 2004. doi: 10.1002/nme.1151.DOI
- [22]M. Duflot. The extended finite element method in thermoelastic fracture mechanics. International Journal for Numerical Methods in Engineering, 74(5):827–847, 2008. doi: 10.1002/nme.2197.DOI
- [23]T. Menouillard and T. Belytschko. Analysis and computations of oscillating crack propagation in a heated strip. International Journal of Fracture, 167(1):57–70, 2011. doi: 10.1007/s10704-010-9519-0.DOI
- [24]M. M. Chiaramonte, B. E. Grossman-Ponemon, L. M. Keer, and A. J. Lew. Numerical analyses of crack path instabilities in quenched plates. Extreme Mechanics Letters, 40: 100878, 2020. doi: 10.1016/j.eml.2020.100878.DOI
- [25]X.-P. Xu and A. Needleman. Numerical simulations of fast crack growth in brittle solids. Journal of the Mechanics and Physics of Solids, 42(9):1397–1434, 1994. doi: 10.1016/0022-5096(94)90003-5.DOI
- [26]B. Kilic and E. Madenci. Prediction of crack paths in a quenched glass plate by using peridynamic theory. International Journal of Fracture, 156(2):165–177, 2009. doi: 10.1007/s10704-009-9355-2.DOI
- [27]Z. Xu, G. Zhang, Z. Chen, and F. Bobaru. Elastic vortices and thermally-driven cracks in brittle materials with peridynamics. International Journal of Fracture, 209(1–2):203–222, 2018. doi: 10.1007/s10704-017-0256-5.DOI
- [28]H. Li, Z. Zhang, L. Wang, X. Gu, Y. Zhang, and X. Shao. Integration of peridynamics and deep learning for efficient and accurate thermomechanical modeling. Applied Sciences, 15(18):10032, 2025. doi: 10.3390/app151810032.DOI
- [29]Y. Hayakawa. Numerical study of oscillatory crack propagation through a two-dimensional crystal. Physical Review E, 49(3):R1804–R1807, 1994. doi: 10.1103/PhysRevE.49.R1804.DOI
- [30]G. A. Francfort and J.-J. Marigo. Revisiting brittle fracture as an energy minimization problem. Journal of the Mechanics and Physics of Solids, 46(8):1319–1342, 1998. doi: 10.1016/S0022-5096(98)00034–9.DOI
- [31]B. Bourdin, G. A. Francfort, and J.-J. Marigo. Numerical experiments in revisited brittle fracture. Journal of the Mechanics and Physics of Solids, 48(4):797–826, 2000. doi: 10.1016/S0022–5096(99)00028–9.DOI
- [32]B. Bourdin, G. A. Francfort, and J.-J. Marigo. The variational approach to fracture. Journal of Elasticity, 91(1–3):5–148, 2008. doi: 10.1007/s10659–007–9107–3.
- [33]L. Ambrosio and V. M. Tortorelli. Approximation of functionals depending on jumps by elliptic functionals via Γ-convergence. Communications on Pure and Applied Mathematics, 43(8):999–1036, 1990. doi: 10.1002/cpa.3160430805.DOI
- [34]C. Miehe, M. Hofacker, and F. Welschinger. A phase field model for rate-independent crack propagation: robust algorithmic implementation based on operator splits. Computer Methods in Applied Mechanics and Engineering, 199(45–48):2765–2778, 2010. doi: 10.1016/j.cma.2010.04.011.
- [35]H. Amor, J.-J. Marigo, and C. Maurini. Regularized formulation of the variational brittle fracture with unilateral contact: numerical experiments. Journal of the Mechanics and Physics of Solids, 57(8):1209–1229, 2009. doi: 10.1016/j.jmps.2009.04.011.DOI
- [36]M. Ambati, T. Gerasimov, and L. De Lorenzis. A review on phase-field models of brittle fracture and a new fast hybrid formulation. Computational Mechanics, 55(2):383–405, 2015. doi: 10.1007/s00466–014–1109–y.DOI
- [37]K. Pham, H. Amor, J.-J. Marigo, and C. Maurini. Gradient damage models and their use to approximate brittle fracture. International Journal of Damage Mechanics, 20(4):618–652, 2011. doi: 10.1177/1056789510386852.DOI
- [38]J.-Y. Wu. A unified phase-field theory for the mechanics of damage and quasi-brittle failure. Journal of the Mechanics and Physics of Solids, 103:72–99, 2017. doi: 10.1016/j.jmps.2017.03.015.DOI
- [39]E. Tanné, T. Li, B. Bourdin, J.-J. Marigo, and C. Maurini. Crack nucleation in variational phase-field models of brittle fracture. Journal of the Mechanics and Physics of Solids, 110:80–99, 2018. doi: 10.1016/j.jmps.2017.09.006.DOI
- [40]A. Kumar, B. Bourdin, G. A. Francfort, and O. Lopez-Pamies. Revisiting nucleation in the phase-field approach to brittle fracture. Journal of the Mechanics and Physics of Solids, 142:104027, 2020. doi: 10.1016/j.jmps.2020.104027.DOI
- [41]M. J. Borden, T. J. R. Hughes, C. M. Landis, and C. V. Verhoosel. A higher-order phase-field model for brittle fracture: formulation and analysis within the isogeometric analysis framework. Computer Methods in Applied Mechanics and Engineering, 273:100–118, 2014. doi: 10.1016/j.cma.2014.01.016.DOI
- [42]T. Gerasimov and L. De Lorenzis. On penalization in variational phase-field models of brittle fracture. Computer Methods in Applied Mechanics and Engineering, 354:990–1026, 2019. doi:10.1016/j.cma.2019.05.038.DOI
- [43]J.-Y. Wu, V. P. Nguyen, C. T. Nguyen, D. Sutula, S. Sinaie, and S. P. A. Bordas. Phase-field modeling of fracture. Advances in Applied Mechanics, 53:1–183, 2020. doi:10.1016/bs.aams.2019.08.001.DOI
- [44]B. Bourdin, J.-J. Marigo, C. Maurini, and P. Sicsic. Morphogenesis and propagation of complex cracks induced by thermal shocks. Physical Review Letters, 112(1):014301, 2014. doi:10.1103/PhysRevLett.112.014301.DOI
- [45]Y. Shao, Y. Zhang, X. Xu, Z. Zhou, W. Li, and B. Liu. Effect of crack pattern on the residual strength of ceramics after quenching. Journal of the American Ceramic Society, 94(9):2804–2807, 2011. doi:10.1111/j.1551-2916.2011.04728.x.DOI
- [46]O. Lopez-Pamies, J. E. Dolbow, G. A. Francfort, and C. J. Larsen. Classical variational phase-field models cannot predict fracture nucleation. Computer Methods in Applied Mechanics and Engineering, 433:117520, 2025. doi:10.1016/j.cma.2024.117520.
- [47]F. Kamarei, B. Zeng, J. E. Dolbow, and O. Lopez-Pamies. Nine circles of elastic brittle fracture: a series of challenge problems to assess fracture models. Computer Methods in Applied Mechanics and Engineering, 448:118449, 2026. doi:10.1016/j.cma.2025.118449.DOI
- [48]B. Zeng, J. Guilleminot, and J. E. Dolbow. Examining crack nucleation under spatially uniform stress states with a complete phase-field model for fracture. Theoretical and Applied Fracture Mechanics, 140:105170, 2025. doi:10.1016/j.tafmec.2025.105170.DOI
- [49]O. Ronsin, F. Heslot, and B. Perrin. Experimental study of quasistatic brittle crack propagation. Physical Review Letters, 75(12):2352–2355, 1995. doi:10.1103/PhysRevLett.75.2352.DOI
- [50]O. Ronsin and B. Perrin. Dynamics of quasistatic directional crack growth. Physical Review E, 58(6):7878–7886, 1998. doi:10.1103/PhysRevE.58.7878.DOI
- [51]B. Yang and K. Ravi-Chandar. Crack path instabilities in a quenched glass plate. Journal of the Mechanics and Physics of Solids, 49(1):91–130, 2001. doi:10.1016/S0022-5096(00)00022-3.
- [52]M. Marder. Instability of a crack in a heated strip. Physical Review E, 49(1):R51–R54, 1994. doi:10.1103/PhysRevE.49.R51.DOI
- [53]S.-i. Sasa, K. Sekimoto, and H. Nakanishi. Oscillatory instability of crack propagations in quasistatic fracture. Physical Review E, 50(3):R1733–R1736, 1994. doi:10.1103/PhysRevE.50.R1733.DOI
- [54]M. Adda-Bedia and Y. Pomeau. Crack instabilities of a heated glass strip. Physical Review E, 52(4):4105–4113, 1995. doi:10.1103/PhysRevE.52.4105.
- [55]F. Corson, M. Adda-Bedia, H. Henry, and E. Katzav. Thermal fracture as a framework for quasi-static crack propagation. International Journal of Fracture, 158(1):1–14, 2009. doi:10.1007/s10704-009-9361-4.DOI
- [56]E. Bouchbinder, H. G. E. Hentschel, and I. Procaccia. Dynamical instabilities of quasi-static crack propagation under thermal stress. Physical Review E, 68(3):036601, 2003. doi:10.1103/PhysRevE.68.036601.DOI
- [57]H.-A. Bahr, A. Gerbatsch, U. Bahr, and H.-J. Weiss. Oscillatory instability in thermal cracking: a first-order phase-transition phenomenon. Physical Review E, 52(1):240–243, 1995. doi:10.1103/PhysRevE.52.240.DOI
- [58]W. Pan, R. Abdelmoula, J. Li, and C. Cheng. Analysis of crack path instabilities in a quenched glass plate using the phase-field cohesive zone model. Journal of Mechanics of Materials and Structures, 19(2):235–250, 2024. doi:10.2140/jomms.2024.19.235.
- [59]C. Miehe, L.-M. Schänzel, and H. Ulmer. Phase field modeling of fracture in multi-physics problems. Part I. Balance of crack surface and failure criteria for brittle crack propagation in thermo-elastic solids. Computer Methods in Applied Mechanics and Engineering, 294:449–485, 2015. doi:10.1016/j.cma.2014.11.016.DOI
- [60]D. Chu, X. Li, and Z. Liu. Study the dynamic crack path in brittle material under thermal shock loading by phase field modeling. International Journal of Fracture, 208 (1–2):115–130, 2017. doi:10.1007/s10704-017-0220-4.DOI
- [61]T. Wang, X. Ye, Z. Liu, X. Liu, D. Chu, and Z. Zhuang. A phase-field model of thermo-elastic coupled brittle fracture with explicit time integration. Computational Mechanics, 65(5):1305–1321, 2020. doi:10.1007/s00466-020-01820-6.DOI
- [62]T. K. Mandal, V. P. Nguyen, J.-Y. Wu, C. Nguyen-Thanh, and A. de Vaucorbeil. Fracture of thermo-elastic solids: phase-field modeling and new results with an efficient monolithic solver. Computer Methods in Applied Mechanics and Engineering, 376:113648, 2021. doi:10.1016/j.cma.2020.113648.DOI
- [63]R. G. Tangella, P. Kumbhar, and R. K. Annabattula. Hybrid phase-field modeling of thermo-elastic crack propagation. International Journal for Computational Methods in Engineering Science and Mechanics, 23(1):29–44, 2022. doi:10.1080/15502287.2021.1904462.DOI
- [64]H. Ruan, S. Rezaei, Y. Yang, D. Gross, and B.-X. Xu. A thermo-mechanical phase-field fracture model: application to hot cracking simulations in additive manufacturing. Journal of the Mechanics and Physics of Solids, 172:105169, 2023. doi:10.1016/j.jmps.2022.105169.DOI
- [65]D. Li, P. Li, W. Li, W. Li, and K. Zhou. Three-dimensional phase-field modeling of temperature-dependent thermal shock-induced fracture in ceramic materials. Engineering Fracture Mechanics, 268:108444, 2022. doi:10.1016/j.engfracmech.2022.108444.DOI
- [66]T. Wang, Y. Zhang, H. Han, L. Wang, X. Ye, and Z. Zhuang. Phase field modeling of crack propagation in three-dimensional quasi-brittle materials under thermal shock. Engineering Fracture Mechanics, 302:110070, 2024. doi:10.1016/j.engfracmech.2024.110070.DOI
- [67]H. Badnava, M. A. Msekh, E. Etemadi, and T. Rabczuk. An h-adaptive thermo-mechanical phase field model for fracture. Finite Elements in Analysis and Design, 138:31–47, 2018. doi:10.1016/j.finel.2017.09.003.DOI
- [68]N. Ferro, S. Micheletti, and S. Perotto. Anisotropic mesh adaptation for crack propagation induced by a thermal shock in 2D. Computer Methods in Applied Mechanics and Engineering, 331:138–158, 2018. doi:10.1016/j.cma.2017.11.024.DOI
- [69]H. Chen, S. Natarajan, E. T. Ooi, and C. Song. Modeling of coupled thermo-mechanical crack propagation in brittle solids using adaptive phase field method with scaled boundary finite element method. Theoretical and Applied Fracture Mechanics, 129:104158, 2024. doi:10.1016/j.tafmec.2023.104158.DOI
- [70]I. Ali, G. Vuga, B. Mavrić, and B. Šarler. A fourth-order phase-field model of crack initiation and propagation under thermomechanical loading solved with strong-form meshless method. Theoretical and Applied Fracture Mechanics, 139:105078, 2025. doi:10.1016/j.tafmec.2025.105078.DOI
- [71]Z. Si, T. Yu, Y. Li, and S. Natarajan. Adaptive multi-patch isogeometric phase-field method for quasi-static brittle fracture based on Nitsche’s method. Computer Methods in Applied Mechanics and Engineering, 414:116154, 2023. doi:10.1016/j.cma.2023.116154.DOI
- [72]H. Zhang, M. Makki Alamdari, B. Shahbodagh, M. Wahab, C. Anitescu, T. Rabczuk, and E. Atroshchenko. A mesh-free multiresolution deep energy method with phase-field modeling of brittle fracture. arXiv preprint arXiv:2608.24126, 2026.arxiv.org/abs/2608.24126
- [73]I. E. Lagaris, A. Likas, and D. I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations. IEEE Transactions on Neural Networks, 9(5):987–1000, 1998. doi:10.1109/72.712178.DOI
- [74]M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019. doi:10.1016/j.jcp.2018.10.045.DOI
- [75]G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang. Physics-informed machine learning. Nature Reviews Physics, 3(6):422–440, 2021. doi:10.1038/s42254-021-00314-5.DOI
- [76]W. E and B. Yu. The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems. Communications in Mathematics and Statistics, 6(1):1–12, 2018. doi:10.1007/s40304-018-0127-z.DOI
- [77]E. Samaniego, C. Anitescu, S. Goswami, V. M. Nguyen-Thanh, H. Guo, K. Hamdia, X. Zhuang, and T. Rabczuk. 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, 362:112790, 2020. doi:10.1016/j.cma.2019.112790.DOI
- [78]V. M. Nguyen-Thanh, X. Zhuang, and T. Rabczuk. A deep energy method for finite deformation hyperelasticity. European Journal of Mechanics - A/Solids, 80:103874, 2020. doi:10.1016/j.euromechsol.2019.103874.
- [79]N. Sukumar and A. Srivastava. Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks. Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022. doi:10.1016/j.cma.2021.114333.DOI
- [80]C. Wu, M. Zhu, Q. Tan, Y. Kartha, and L. Lu. A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 403:115671, 2023. doi: 10.1016/j.cma.2022.115671.DOI
- [81]M. A. Nabian, R. J. Gladstone, and H. Meidani. Efficient training of physics-informed neural networks via importance sampling. Computer-Aided Civil and Infrastructure Engineering, 36(8):962–977, 2021. doi: 10.1111/mice.12685.DOI
- [82]M. Tancik, P. P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. T. Barron, and R. Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In Advances in Neural Information Processing Systems, volume 33, pages 7537–7547, 2020.arxiv.org/abs/2006.10739
- [83]T. Müller, A. Evans, C. Schied, and A. Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4):102:1–102:15, 2022. doi: 10.1145/3528223.3530127.DOI
- [84]X. Huang and T. Alkhalifah. Efficient physics-informed neural networks using hash encoding. Journal of Computational Physics, 501:112760, 2024. doi: 10.1016/j.jcp.2024.112760.DOI
- [85]S. Goswami, C. Anitescu, S. Chakraborty, and T. Rabczuk. Transfer learning enhanced physics informed neural network for phase-field modeling of fracture. Theoretical and Applied Fracture Mechanics, 106:102447, 2020. doi: 10.1016/j.tafmec.2019.102447.DOI
- [86]S. Goswami, C. Anitescu, and T. Rabczuk. Adaptive fourth-order phase field analysis using deep energy minimization. Theoretical and Applied Fracture Mechanics, 107:102527, 2020. doi: 10.1016/j.tafmec.2020.102527.DOI
- [87]M. Manav, R. Molinaro, S. Mishra, and L. De Lorenzis. Phase-field modeling of fracture with physics-informed deep learning. Computer Methods in Applied Mechanics and Engineering, 429:117104, 2024. doi: 10.1016/j.cma.2024.117104.DOI
- [88]A. Chakraborty, C. Anitescu, S. Goswami, X. Zhuang, and T. Rabczuk. Variational energy based XPINNs for phase field analysis in brittle fracture. arXiv preprint arXiv:2207.02307, 2022.arxiv.org/abs/2207.02307
- [89]Y. Ghaffari Motlagh, P. K. Jimack, and R. de Borst. Deep learning phase-field model for brittle fractures. International Journal for Numerical Methods in Engineering, 124(3):620–638, 2023. doi: 10.1002/nme.7135.DOI
- [90]B. Feng and X.-P. Zhou. The novel physics-enhanced graph neural network for phase-field fracture modelling. Computer Methods in Applied Mechanics and Engineering, 446:118284, 2025. doi: 10.1016/j.cma.2025.118284.
- [91]P. Zhang, K. Tang, and B. Xu. ChebyKAN-enhanced deep energy method for phase field fracture: from homogeneous to heterogeneous materials. Journal of the Mechanics and Physics of Solids, 214:106693, 2026. doi: 10.1016/j.jmps.2026.106693.DOI
- [92]N. Plunge, P. Brommer, R. S. Edwards, and E. G. Kakouris. Deep learning-based phase-field modelling of brittle fracture in anisotropic media. International Journal for Numerical Methods in Engineering, 127(13):e70381, 2026. doi: 10.1002/nme.70381.DOI
- [93]A. Dean and B. Bahtiri. A fracture-informed multiphysics self-sensing framework for phase-field fracture in piezoresistive materials using the deep energy method. Theoretical and Applied Fracture Mechanics, 148:105890, 2026. doi: 10.1016/j.tafmec.2026.105890.DOI
- [94]M. Yu, X. Long, C. Jiang, X. Wang, and K. Liu. Physics-informed neural networks for inverse length scale identification and forward phase-field fracture modeling. Engineering Fracture Mechanics, 337:111978, 2026. doi: 10.1016/j.engfracmech.2026.111978.DOI
- [95]Y. Wang, Y. Lin, S. Goswami, L. Zhao, H. Zhang, J. Bai, C. Anitescu, M. S. Eshaghi, X. Zhuang, T. Rabczuk, and Y. Liu. Towards unified AI-driven fracture mechanics: the extended deep energy method (XDEM). Nature Communications, 17:8492, 2026. doi: 10.1038/s41467-026-76748-1.DOI
- [96]E. Kiyani, M. Manav, N. Kadivar, L. De Lorenzis, and G. E. Karniadakis. Predicting crack nucleation and propagation in brittle materials using deep operator networks with diverse trunk architectures. Computer Methods in Applied Mechanics and Engineering, 441:117984, 2025. doi: 10.1016/j.cma.2025.117984.DOI
- [97]E. Hamdi and E. Lejeune. Towards robust surrogate models: benchmarking machine learning approaches to expediting phase field simulations of brittle fracture. Computer Methods in Applied Mechanics and Engineering, 449:118526, 2026. doi: 10.1016/j.cma.2025.118526.DOI
- [98]A. Ani, R. Nakka, G. Subhash, J.-F. Molinari, and S. A. Ponnusami. Machine learning for computational fracture and damage mechanics—status and perspectives. Engineering Fracture Mechanics, 332:111778, 2026. doi: 10.1016/j.engfracmech.2025.111778.DOI
- [99]X.-P. Zhou and X.-L. Yu. Transfer learning enhanced nonlocal energy-informed neural network for quasi-static fracture in rock-like materials. Computer Methods in Applied Mechanics and Engineering, 430:117226, 2024. doi: 10.1016/j.cma.2024.117226.DOI
- [100]X. Zhou and J. Ma. Nonlocal energy-encoded physics-informed neural networks for crack propagation in solid materials. Journal of the Mechanics and Physics of Solids, 216:106770, 2026. doi: 10.1016/j.jmps.2026.106770.DOI
- [101]S. Cai, Z. Wang, S. Wang, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks for heat transfer problems. Journal of Heat Transfer, 143(6):060801, 2021. doi: 10.1115/1.4050542.DOI
- [102]K.-C. Lin, K.-C. Wang, and C.-H. Hu. Investigating deep energy method applications in thermoelasticity. Engineering Analysis with Boundary Elements, 159:302–314, 2024. doi: 10.1016/j.enganabound.2023.12.012.DOI
- [103]A. Mojahedin, M. Hashemitaheri, M. Salavati, X. Fu, and T. Rabczuk. A deep energy method for the analysis of thermoporoelastic functionally graded beams. International Journal of Computational Methods, 19(8):2143020, 2022. doi: 10.1142/S0219876221430209.
- [104]M. Raj, P. Kumbhar, and R. K. Annabattula. Physics-informed neural networks for solving thermo-mechanics problems of functionally graded material. arXiv preprint arXiv:2111.10751, 2021.arxiv.org/abs/2111.10751
- [105]A. Harandi, A. Moeineddin, M. Kaliske, S. Reese, and S. Rezaei. Mixed formulation of physics-informed neural networks for thermo-mechanically coupled systems and heterogeneous domains. International Journal for Numerical Methods in Engineering, 125(4):e7388, 2024. doi: 10.1002/nme.7388.DOI
- [106]X.-L. Yu and X.-P. Zhou. A nonlocal energy-informed neural network for isotropic elastic solids with cracks under thermomechanical loads. International Journal for Numerical Methods in Engineering, 124(18):3935–3963, 2023. doi: 10.1002/nme.7296.DOI
- [107]Y. Wu, W. Liu, Z. Zhang, B. Liu, H. Sun, G. Wang, and X. Gong. Identification of rock fracture parameters under non-uniform temperature fields based on physics-informed neural networks. Acta Mechanica Sinica, 42(9):425672, 2026. doi: 10.1007/s10409-025-25672-x.DOI
- [108]S. K. Yadav, V. Yadav, and R. Patil. Fracture analysis of functionally graded plates under thermo-mechanical loading using mesh enriched physics-informed neural networks (M-PINNs). Research Square preprint, doi: 10.21203/rs.3.rs-9295110/v1, 2026.DOI
- [109]J.-N. He, H. Li, S. Natarajan, and T. Yu. Thermo-mechanical phase-field fracture modeling via deep Ritz method. Engineering Fracture Mechanics, 346:112651, 2026. doi: 10.1016/j.engfracmech.2026.112651.DOI
- [110]H.-e. Kong. A variational multiscale method for material heterogeneity of solid mechanics and an investigation of machine learning framework for fracture modeling. PhD dissertation, The Pennsylvania State University, 2026. URL https://etda.libraries.psu.edu/catalog/20420hqk5365.
- [111]H. Zhang, M. Makki Alamdari, B. Shahbodagh, M. Vahab, C. Anitescu, T. Rabczuk, and E. Atroshchenko. An extended deep energy method for thermo-mechanical crack propagation. arXiv preprint arXiv:2610.09433, 2026.arxiv.org/abs/2610.09433
- [112]S. B. Tang, H. Zhang, C. A. Tang, and H. Y. Liu. Numerical model for the cracking behavior of heterogeneous brittle solids subjected to thermal shock. International Journal of Solids and Structures, 80:520–531, 2016. doi: 10.1016/j.ijsolstr.2015.10.012.DOI
- [113]C. Miehe, F. Welschinger, and M. Hofacker. Thermodynamically consistent phase-field models of fracture: variational principles and multi-field FE implementations. International Journal for Numerical Methods in Engineering, 83(10):1273–1311, 2010. doi: 10.1002/nme.2861.DOI
- [114]T. Linse, P. Hennig, M. Kästner, and R. de Borst. A convergence study of phase-field models for brittle fracture. Engineering Fracture Mechanics, 184:307–318, 2017. doi: 10.1016/j.engfracmech.2017.09.013.DOI
- [115]Q. Yang, L. Stainier, and M. Ortiz. A variational formulation of the coupled thermo-mechanical boundary-value problem for general dissipative solids. Journal of the Mechanics and Physics of Solids, 54(2):401–424, 2006. doi: 10.1016/j.jmps.2005.08.010.DOI
- [116]B. Kamatagé, L. Cheng, R. Abdelmoula, E. Danho, and D. Kondo. An incremental variational method to the coupling between gradient damage, thermoelasticity and heat conduction. Comptes Rendus. Mécanique, 353:1063–1084, 2025. doi: 10.5802/crmeca.325.
- [117]E. H. Georgoulis, M. Loulakis, and A. Tsiourvas. Discrete gradient flow approximations of high dimensional evolution partial differential equations via deep neural networks. Communications in Nonlinear Science and Numerical Simulation, 117:106893, 2023. doi:10.1016/j.cnsns.2022.106893.DOI
- [118]M. S. Park, C. Kim, H. Son, and H. J. Hwang. The deep minimizing movement scheme. Journal of Computational Physics, 494:112518, 2023. doi:10.1016/j.jcp.2023.112518.DOI
- [119]H. Huang, H. Yu, W. Xu, Q. Wang, and H. Wu. A thermo-mechanical localizing gradient damage model with modified Mazars strain. International Journal for Numerical Methods in Engineering, 126(2):e7652, 2025. doi:10.1002/nme.7652.DOI
- [120]D. P. Kingma and J. Ba. Adam: a method for stochastic optimization. In International Conference on Learning Representations, 2015.arxiv.org/abs/1412.6980
- [121]D. Hendrycks and K. Gimpel. Gaussian error linear units (GELUs). arXiv preprint arXiv:1606.08415, 2016.
- [122]T. J. R. Hughes, J. A. Cottrell, and Y. Bazilevs. Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement. Computer Methods in Applied Mechanics and Engineering, 194(39–41):4135–4195, 2005. doi:10.1016/j.cma.2004.10.008.DOI
- [123]J. A. Rivera, J. M. Taylor, Á. J. Omella, and D. Pardo. On quadrature rules for solving partial differential equations using neural networks. Computer Methods in Applied Mechanics and Engineering, 393:114710, 2022. doi:10.1016/j.cma.2022.114710.DOI
- [124]I. Loshchilov and F. Hutter. SGDR: stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017.arxiv.org/abs/1608.03983
- [125]T. Gustafsson and G. D. McBain. scikit-fem: a Python package for finite element assembly. Journal of Open Source Software, 5(52):2369, 2020. doi:10.21105/joss.02369.DOI
- [126]R. G. Tangella, P. Kumbhar, and R. K. Annabattula. Thermo-elastic fracture using phase field method in FEniCS: source code for the single-edge notched specimen. Mechanics of Materials Group, Indian Institute of Technology Madras, 2022. https://home.iitm.ac.in/ratna/codes/thermo-elastic-fracture-fenics/index.html, accessed 2026-10-29.
- [127]P. F. Becher. Effect of water bath temperature on the thermal shock of Al₂O₃. Journal of the American Ceramic Society, 64(1):C17–C18, 1981. doi:10.1111/j.1151-2916.1981.tb09544.x.DOI
- [128]Y. Kim, W.-J. Lee, and E. D. Case. The measurement of the surface heat transfer coefficient for ceramics quenched into a water bath. Materials Science and Engineering: A, 145(1):L7–L11, 1991. doi:10.1016/0921-5093(91)90307-9.DOI
- [129]Z. Zhou, F. Song, Y. Shao, S. Meng, C. Jiang, and J. Li. Characteristics of the surface heat transfer coefficient for Al₂O₃ ceramic in water quench. Journal of the European Ceramic Society, 32(12):3029–3034, 2012. doi:10.1016/j.jeurceramsoc.2012.04.027.DOI
- [130]Y. Liu, X. Wu, Q. Guo, C. Jiang, F. Song, and J. Li. Experiments and numerical simulations of thermal shock crack patterns in thin circular ceramic specimens. Ceramics International, 41(1):1107–1114, 2015. doi:10.1016/j.ceramint.2014.09.036.DOI
- [131]L. Svolos, C. A. Bronkhorst, and H. Waisman. Thermal-conductivity degradation across cracks in coupled thermo-mechanical systems modeled by the phase-field fracture method. Journal of the Mechanics and Physics of Solids, 137:103861, 2020. doi: 10.1016/j.jmps.2019.103861.DOI
- [132]NCI Australia. Gadi supercomputer. doi: 10.25914/608bfd1838db2, 2019.
- [133]PVC (Research Infrastructure), UNSW Sydney. Katana. doi: 10.26190/669X-A286, 2010.DOI
- [134]I. M. Sobol. On the distribution of points in a cube and the approximate evaluation of integrals. USSR Computational Mathematics and Mathematical Physics, 7(4):86–112, 1967. doi: 10.1016/0041-5553(67)90144-9.DOI













