WorksPhysics and Astronomy
Data-Driven Cohesive Zone Modeling within the Generalized Standard Materials Framework
Abstract
Cohesive zone models are widely used to describe fracture and interfacial failure, yet most formulations prescribe problem-specific analytical traction-separation laws together with phenomenological rules for unloading and reloading, leading to specialized models for different cohesive behaviors. This work develops a unified learnable cohesive formulation within the generalized standard materials framework, in which the response is generated from learned constitutive functions while the underlying thermodynamic structure remains fixed. The surface free energy is decomposed into active and contact contributions, and irreversible damage evolution is governed by a learned mode-dependent damage resistance. The active energy is represented by an input-convex neural network, while the inverse damage resistance is represented by a monotone neural network. Convexity, monotonicity, normalization, and damage irreversibility are incorporated directly into the constitutive representation. Direct parameterization of the inverse resistance yields an explicit damage update and avoids local nonlinear inversion during constitutive evaluation. Material-point studies show that the formulation can represent qualitatively distinct cohesive responses, including plateaus, extended softening tails, irregular softening, nonlinear unloading, distinct Mode I and Mode II behaviors, and several classical mixed-mode cohesive laws. The framework therefore replaces law-specific model construction with a single thermodynamically structured representation capable of learning cohesive responses of broad functional complexity from data.
Introduction
Fracture in engineering materials often involves a finite process zone and cannot be described solely by the loss of cohesion across an ideally sharp crack. Within this zone, separation may involve mechanisms such as microcracking, plastic deformation, fiber bridging, interfacial debonding, or adhesive degradation. Cohesive zone models (CZMs) provide a reduced description of these mechanisms by replacing the detailed process-zone physics with a traction–separation relation acting across a displacement jump between separating surfaces [7, 9, 15, 22, 37, 39, 54, 63].
CZMs are widely used to model fracture and interfacial failure when the structure and dissipation of the process zone influence the macroscopic response. Many formulations prescribe the traction–separation relation through fixed analytical forms [42], including bilinear [19, 22], piecewise-linear [13], and trapezoidal [54] laws. These phenomenological models are simple and effective, but their prescribed functional forms can restrict the range of process-zone responses that can be represented in complex interfaces and composite materials [60].
A further challenge arises in mixed-mode fracture, where normal and tangential separations evolve jointly and their coupling must be represented consistently in the cohesive response [14, 35, 41]. Potential-based formulations provide a systematic way to introduce this coupling by deriving the traction vector from a scalar interface potential. For a sufficiently smooth potential, the resulting traction field is integrable and satisfies reciprocal normal–tangential coupling. Early examples include the polynomial potential of Needlemann [37] and the exponential potential of Xu and Needleman [64], later used by Xu and Needleman [63] to model dynamic crack growth and branching. The Park–Paulino–Roesler (PPR) model subsequently introduced separate control of the Mode I and Mode II fracture energies, cohesive strengths, and shape parameters [40, 41, 42].
Despite these advantages, the range of cohesive responses represented by a potential-based CZM remains limited by the prescribed form of the potential. Classical formulations typically employ low-parameter analytical functions, which may be restrictive for experimentally identified cohesive responses exhibiting plateaus [49], extended softening tails [48], or irregular and strongly mode-dependent shapes [4, 49]. This limitation is particularly important under large-scale bridging, where the process zone is not small relative to the structural dimensions and the response can depend on the detailed shape of the cohesive law rather than only on its peak strength and fracture energy [6, 48].
A separate limitation concerns irreversible evolution. An interface potential that depends only on the current separation describes a reversible response and retains no memory of prior decohesion. It therefore cannot by itself represent damage accumulation, hysteresis, or irreversible fracture dissipation. Classical CZMs introduce such behavior through damage or history variables together with prescribed unloading–reloading rules [23, 29, 38, 65]. A thermodynamically consistent formulation instead requires internal variables and evolution laws that distinguish recoverable interface energy from irreversible dissipation and satisfy the dissipation inequality [14, 36]. Moreover, the existence of an interface potential alone does not guarantee a physically admissible mixed-mode response; particular potential forms can produce residual or repulsive tractions and other nonphysical behavior [14, 35]. These considerations motivate a constitutive structure in which the stored energy, internal variables, and irreversible evolution are treated within a common thermodynamic framework.
The generalized standard material (GSM) framework [21] provides a systematic thermodynamic structure by combining a Helmholtz free-energy potential for the recoverable response with a convex dissipation potential governing the evolution of internal variables. Under the standard GSM assumptions, this construction ensures non-negative dissipation and provides evolution laws for irreversible internal variables, while the free-energy potential determines the recoverable response at fixed internal state. The framework has been extended to cohesive interfaces in adhesion–friction models [45], rate-dependent formulations [31], and rate-independent delamination models [47, 57]. Related thermodynamic interface models have also been formulated through damage surfaces and consistency conditions [61]. These approaches provide a consistent basis for irreversible cohesive fracture, but their constitutive potentials and evolution functions are generally prescribed analytically. Their representational capacity therefore remains tied to the assumed functional forms.
Recent advances in data-driven constitutive modeling provide a route to increase constitutive flexibility while retaining physical structure. Since the early use of neural networks for learning material response directly from data [20], the field has progressed from black-box constitutive approximations toward models that incorporate physical principles either through the training objective or directly through the model architecture [18]. Thermodynamics-informed approaches impose conditions such as non-negative dissipation through physics-based loss terms [56], whereas structure-preserving architectures can enforce properties such as convexity, monotonicity, and polyconvexity by construction [5, 28]. In particular, convex neural networks and neural GSM formulations have been used to learn free-energy and dissipation potentials for a range of bulk constitutive models [2, 16, 17, 52, 62].
Neural constitutive modeling of interfaces has received comparatively less attention. Existing approaches include automated discovery of traction–separation laws through deep reinforcement learning [58], thermodynamics-informed neural representations of mixed-mode traction–separation behavior [59], and neural cohesive laws for monotonic and cyclic fracture [50, 51]. Related developments have also introduced learned potential representations for rate-and-state friction [33]. Despite these advances, a general data-driven cohesive formulation that combines flexible representation of the local traction–separation response with the thermodynamic structure of GSM remains lacking. In particular, thermodynamic restrictions introduced through penalty terms in the training objective require balancing additional loss contributions and do not hold exactly by construction. This motivates a formulation in which the required constitutive properties are embedded directly in the neural parametrization of the free-energy and dissipation mechanisms.
To address this gap, this paper develops a thermodynamically structured, data-driven cohesive zone model within the GSM framework. Rather than prescribing the traction–separation law and unloading rules directly, the cohesive response is generated from learned constitutive functions. The active potential is represented by an input-convex neural network (ICNN) [3], constructed to satisfy the required convexity and monotonicity conditions, while the inverse damage resistance is represented by a monotone neural network (MNN) [26]. The latter directly parametrizes the mode-mixity-dependent inverse resistance and therefore avoids the numerical inversion of the resistance relation during constitutive evaluation. We consider both an MNN with unrestricted dependence on mode mixity and a full-MNN that additionally imposes increasing resistance with mode mixity, motivated by interfaces whose fracture toughness increases from Mode I toward Mode II. In contrast to approaches that impose thermodynamic conditions through additional terms in the training objective [59], the required convexity, monotonicity, and normalization properties are embedded directly in the constitutive parametrization.
The resulting formulation provides an explicit irreversible damage update without a local nonlinear solve for the damage state. The constitutive update is differentiable almost everywhere, which permits the neural constitutive functions to be identified directly from traction–separation data by gradient-based optimization through the complete loading history.
Numerical experiments demonstrate that the formulation can reproduce a broad range of cohesive responses, including bilinear, trapezoidal, exponential, half-sine, irregular softening, and PPR laws, while retaining the imposed thermodynamic structure. The examples also show that a single identified model can represent distinct Mode I and Mode II responses and predict the response at intermediate mode mixities not included in the calibration data.
The remainder of the paper is organized as follows. § 2 develops the cohesive formulation within the GSM framework, and § 3 introduces the ICNN and MNN parametrizations of the active potential and inverse damage resistance. § 4 defines the calibration dataset and identification objective. § 5 presents the numerical examples for Mode I and mixed-mode cohesive responses, including a comparison of the MNN and full-MNN representations. § 6 summarizes the main findings. The appendices collect the convex-analysis background (Appendix § 10), the convexity results used in the formulation (Appendix § 11), implementation and hyperparameter details (Appendix § 12), additional mixed-mode results (Appendix § 13), and the reference cohesive laws used in the classical-model study (Appendix § 14).
A GSM-Based Cohesive Zone Model
This section first briefly reviews the fundamental ingredients of cohesive zone models and then presents a flexible formulation of cohesive interfaces as an interfacial damage model within the GSM framework. This formulation naturally decomposes the constitutive description into three independent functions, each of which can be parametrized arbitrarily, provided that fundamental thermodynamic requirements — such as convexity, monotonicity, and other admissibility constraints — are satisfied.
Cohesive Zone Models
A cohesive zone model represents the fracture process zone as an internal interface that transmits a cohesive traction . Let and denote the displacement traces on the two sides of the interface, as shown in Figure 1(a). The displacement jump, or separation vector, is defined by

Figure 1. Cohesive zone model of an opening crack. (a) A process zone ahead of the traction-free crack bridges the two faces with a traction resisting their separation ; the interface itself is denoted . (b) The behavior is governed by a traction–separation law : the traction rises to a cohesive strength and then softens to zero, with the area under the curve equal to the fracture energy . The schematic shows a 1D model for clarity. In general, and are vectors.
A cohesive zone model can be specified by a traction–separation law . The traction typically increases to a cohesive strength and subsequently decreases to zero, at which point a traction-free crack forms, as illustrated in Figure 1(b). In one dimension, the area under the traction–separation curve equals the fracture energy .
In two dimensions,
where and are the unit normal and tangent to the interface. Here, denotes opening, while measures sliding and its sign specifies the sliding direction. The traction is decomposed similarly as
Pure opening, defined by and , is termed Mode I, whereas pure sliding, defined by and , is termed Mode II. General separation is mixed-mode. Following Turon et al. [53], the mode mixity is quantified by
with for Mode I and for Mode II. By convention, when the denominator vanishes.
The remainder of this section formulates the traction–separation law within the generalized standard materials framework, which ensures thermodynamic admissibility by construction.
A Generalized Standard Formulation
This section formulates the proposed rate-independent cohesive law within the generalized standard materials framework [21, 45]. The free energy determines the state laws, while a convex dissipation potential and the associated normality relation govern damage evolution. The model is specified by an active energy, a contact energy, and a scalar damage-resistance function. Relevant concepts from convex analysis are summarized in Appendix § 10.
All energetic quantities and powers are measured per unit reference area of the interface . The local state is described by the separation and a scalar damage variable
Surface Free Energy
Within the GSM framework, the constitutive response is derived from a surface free energy . The free surface energy is required to be non-negative, convex, and continuously differentiable with respect to for every fixed .
To distinguish the interface mechanisms associated with separation and contact, we adopt the tension–compression decomposition [25, 61]
The active energy represents the recoverable energy associated with opening and sliding and is degraded by damage. The contact energy represents the compressive normal penalty response and is not degraded. This decomposition is a modeling assumption of the present work. It permits the cohesive and contact responses to be treated separately and prevents damage from degrading the compressive normal response.
A sufficient structural condition for the total free energy in Eq. (6) to satisfy the stated properties is that the two composed contributions, regarded as functions of , are individually non-negative, convex, and continuously differentiable. We further impose the normalization
Since , these conditions ensure that is non-negative, convex, and continuously differentiable for every fixed .
The interface traction and the damage-driving force follow from the free energy as
Here, is the traction transmitted across the interface, and is the thermodynamic force conjugate to damage. Under pure normal compression, and , one has and hence . Closed sliding may still drive damage because the tangential component remains active. At , the active opening and shear tractions vanish, whereas the contact response remains active.
Cohesive laws commonly assume symmetry with respect to reversal of the tangential separation [14, 41, 55, 61, 63]. Under this assumption, the active energy can be expressed as
The corresponding composed active energy is convex in if and only if is jointly convex and non-decreasing in both arguments; see lemma B.3. Consequently, for ,
Since is non-decreasing in , the tangential traction has the same sign as . This structure precludes a reversal of the tangential traction relative to the tangential separation, a nonphysical response that can arise in some potential-based cohesive laws under mixed-mode loading [35].
If the two tangential directions are instead distinguished, the active energy takes the form
In this case, the corresponding composed energy is convex in if and only if is jointly convex and non-decreasing in its first argument; see lemma B.2. No monotonicity condition is required with respect to the signed tangential argument .
Similarly, the contact contribution is convex in if and only if is convex and non-decreasing in its argument.
Dissipation Potential
We next introduce the dissipation potential , where is the current mode mixity defined in Eq. (4). For each fixed , the function is assumed to be proper, non-negative, convex, and lower semicontinuous in , with . Rate independence is imposed through positive homogeneity of degree one,
Because is restricted to , the admissible damage rates depend on the current state. Irreversibility requires for , while saturation at complete decohesion requires at . We therefore define
The dissipation potential is finite only for . For a scalar damage variable, convexity and positive homogeneity of degree one then imply that is linear in on the admissible branch.
For , convexity and positive homogeneity of degree one imply that the dissipation potential is linear in on the admissible half-line . Including the constraint , it takes the form
where is the damage resistance. The function specifies the rate-independent dissipation and its dependence on the mode mixity.
The generalized normality relation [21, 45] gives
Since is convex and positively homogeneous of degree one, Eq. (15) implies
Together with the state relations in Eqs. (8) and (9), this yields the local isothermal dissipation inequality
For , the normality relation is equivalent to the KKT conditions
Hence, damage remains fixed while and evolves only when . The activation function is therefore derived from the dissipation potential rather than prescribed independently. At , the constraint prevents further damage evolution.
Because the affine degradation law in Eq. (6) makes independent of , the KKT conditions admit an explicit update when is continuous and strictly increasing. Define
where denotes the inverse of with respect to , i.e., . The damage variable then evolves according to
The running maximum guarantees damage irreversibility, while the definition of ensures that . Damage initiates when the driving force reaches . The strict monotonicity of with respect to guarantees that the inverse in Eq. (19) is single-valued. If were only non-decreasing, several values of could correspond to the same value of , so the inverse would not be uniquely defined. In addition, a forward parameterization of would generally require solving for during each constitutive update. This additional local nonlinear solve would increase the computational cost of inference and complicate training because gradients must be propagated through the solver.
For complete decohesion at fixed mode mixity, the dissipated energy per unit reference area is defined as,
The integral of therefore determines the mode-dependent fracture energy, with and , while the variation of with determines how this energy is dissipated during damage growth. If varies during the process, the dissipated energy becomes and is generally path-dependent.
Remark 2.1. Thermodynamic admissibility requires . Continuity and strict monotonicity of with respect to are additional modeling assumptions that ensure a well-defined, single-valued inverse . These assumptions permit the explicit damage update above and avoid an additional local nonlinear solve.
Specialization of the Constitutive Functions
The constitutive model is defined by three functions: the active energy , the contact energy , and the damage resistance . Their specific forms are introduced below.
Contact potential. The compressive contact response is represented using the standard quadratic penalty potential [25],
where is the penetration measure and is the normal penalty stiffness. The potential is non-negative, convex, non-decreasing in its argument, and continuously differentiable, with zero value and slope at . A sufficiently large limits negative normal separation and provides a penalty approximation of the impenetrability constraint.
Active potential. We consider the shear-symmetric case discussed in § 2.2.1 and represent the active potential by an input-convex neural network (ICNN) [3] with inputs . The network is constrained to be jointly convex and non-decreasing in both inputs. The tangent approximation introduced in § 3.1 preserves these properties while imposing zero energy and zero traction at the undeformed state. In the numerical examples, we use for brevity.
Damage resistance. For the reasons discussed in § 2.2.2, we directly parameterize the inverse resistance and require it to be strictly increasing in . This condition is equivalent to strict monotonicity of with respect to . We represent using the partially monotone neural network of Kim and Lee [26], with monotonicity imposed with respect to and unrestricted dependence on . We refer to this architecture as the MNN. For interfaces whose fracture toughness increases with mode mixity [4, 11, 24, 32], we additionally consider a fully monotone variant, denoted as the full-MNN, in which is constrained to be increasing in and decreasing in . The latter condition is equivalent to requiring to increase with at fixed damage. This additional monotonicity is introduced as an empirically motivated inductive bias rather than a general constitutive requirement. It is particularly useful when the calibration data sparsely sample the mode-mixity space.
Remark 2.2. In three dimensions, the tangential separation is a vector in the interface plane. Tangential symmetry then generalizes to in-plane isotropy, whereby the potentials depend on only through its magnitude [31]. The composition result in lemma B.2 extends directly to this vector-valued setting. Under this assumption, the model does not distinguish between Mode II and Mode III and cannot represent in-plane shear anisotropy or non-coaxial tangential traction–separation responses; see § 6.
Neural Network Parametrization
We use neural-network parametrizations for the active potential and the inverse damage resistance , both inferred indirectly from measured traction–separation data. The architectures are constructed to satisfy the required convexity and monotonicity conditions by construction. The physical normalizations, including a stress-free undeformed state for and zero damage at zero driving force for , are also imposed exactly. Thus, the constitutive structure is embedded directly into the parametrization, without additional penalty terms in the data-misfit loss. The two network architectures are described below.
Input-Convex Neural Network
The active potential is constructed from a raw ICNN potential [3] with the active kinematic measures introduced in § 2.2.1 as inputs. The ICNN propagates the input through a convex hidden path together with direct input connections. For ,
where the output is scalar. The constrained weights are parameterized as
and is a convex and non-decreasing activation function, taken here as softplus. These constraints make jointly convex and non-decreasing with respect to . In the original ICNN construction [3], non-negativity is required for the hidden-to-hidden weights, whereas the direct input weights may remain unconstrained. We additionally constrain the direct input weights to be non-negative so that the network is both convex and non-decreasing in , as required by the present constitutive formulation. The architecture of the ICNN in this work is illustrated in Figure 2.
Figure 2. Schematic of the ICNN used to evaluate .
A raw ICNN does not, in general, satisfy zero energy and zero traction at the undeformed state. We impose these conditions through the first-order tangent correction
Since the correction is affine in , it preserves convexity; see lemma B.1. Moreover, for the positive-weight softplus ICNN used here, is componentwise non-decreasing. Therefore, for ,
so the required non-decreasing dependence on the active kinematic measures is also preserved. By construction,
Finally, convexity of implies that its tangent plane at is a global underestimator. Hence, . The resulting active potential is therefore convex, non-negative, and non-decreasing, with zero energy and zero traction at the undeformed state. Related normalization strategies have been used in convex neural-network constitutive models to impose the stress-free reference state [16, 52].
The learnable parameters are
The construction above permits coupling between the normal and tangential responses. Many analytical cohesive laws [1, 10, 53] instead take the elastic response to be uncoupled so that tangential sliding does not alter the normal traction. A generic ICNN does not enforce this structure and may therefore learn artificial normal–tangential coupling from limited or noisy data. When elastic mode coupling is not desired, we use a partitioned architecture in which and are passed through independent ICNNs and their outputs are summed to form . This construction eliminates mixed normal–tangential derivatives by construction while preserving the convexity and monotonicity properties described above.
Remark 3.1. The tangent correction preserves convexity for both the coupled and partitioned architectures. A subtlety arises in smoothness with respect to because the inputs and are non-smooth at zero. At the pristine state, the zero-gradient normalization removes the associated non-differentiability and makes the composed potential differentiable. For a coupled ICNN, however, this property does not generally extend along the coordinate axes. A discontinuity in the normal traction may occur across when , and a discontinuity in the tangential traction may occur across when , because the corresponding derivatives of the active potential need not vanish on these axes. The partitioned form avoids these discontinuities. After the tangent correction, each contribution has zero slope at the origin. Consequently, the composed potential is globally convex and with respect to .
Monotone Neural Network
We use an adaptation of the scalable monotonic neural network (SMNN) of Kim and Lee [26] to parametrize the inverse resistance function . We refer to the resulting architectures as the MNN and full-MNN.
For the MNN, is treated as the monotone input, whereas remains unrestricted. Following the partially connected structure of Kim and Lee [26], each hidden layer contains three streams: an exponentiated stream, a confluence stream, and an unrestricted nonlinear stream, denoted by the subscripts , , and , respectively. The first-layer inputs are
For , the streams are evaluated as
The streams are coupled between consecutive layers according to
where denotes vector concatenation. The scalar output is
The exponentiated weights are strictly positive. Together with the strictly increasing tanh activation, this construction makes strictly increasing with respect to , while its dependence on remains unrestricted. Here, tanh is adopted to obtain a smooth and strictly monotone parametrization. The architecture of the MNN in this work is illustrated in Figure 3.
Figure 3. Schematic of the MNN used to evaluate .
The raw network output has an arbitrary offset and is not directly tied to the damage scale. We therefore define the shifted output
where is a learnable scalar. Since
the trial damage satisfies for every mode mixity. When the shifted network output crosses zero, damage initiates at the mode-dependent threshold satisfying
Thus, the offset introduces a strictly positive damage-onset threshold without restricting the dependence of that threshold on .
We also consider a more restrictive variant, referred to as the full-MNN, for interfaces whose resistance increases with mode mixity. In this case, is required to increase with and decrease with . Both variables can therefore be passed through the monotone stream after replacing by :
Since no unrestricted input remains, the confluence and nonlinear streams are omitted. We define
The reference value in Eq. (40) is chosen to be independent of so that the imposed monotonicity with respect to mode mixity is preserved. Since is increasing in , this choice also ensures and hence for all . The resulting onset threshold is non-decreasing with by construction.
The learnable parameters of the MNN are
whereas the full-MNN omits the confluence and nonlinear streams, so
Remark 3.2. When mode mixity is absent, such as in a one-dimensional cohesive problem, the unrestricted streams are unnecessary. The architecture then reduces to a univariate monotone neural network with positive weights and non-decreasing activation functions [12, 30].
Dataset and Identification Objective
Because damage evolves irreversibly, the separation–traction measurements cannot be treated as independent samples. The calibration dataset is therefore organized into ordered loading–unloading paths. Each path is represented as
and the complete dataset is
Each path is assumed to start from a pristine interface, . The damage states are not treated as independent calibration variables; they are reconstructed sequentially from the separation history using the constitutive update. Consequently, the predicted traction depends on the preceding loading history.
Let collect the parameters of the active potential and inverse resistance. They are identified by minimizing
where and the expectation is taken uniformly over the loading paths. In the present implementation, each optimization step evaluates all paths over their complete ordered histories before backpropagation. The loss is minimized using the Adam optimizer.
Remark 4.1. The decomposition of the constitutive response into the active potential and the damage resistance is generally non-unique: distinct pairs may reproduce the same traction histories, as illustrated in Appendix § 13. Consequently, traction–separation data do not, in general, uniquely identify the two latent constitutive functions individually.
Numerical Examples
This section presents numerical examples that assess the representational and identification properties of the proposed cohesive formulation. The first example considers Mode I loading and examines whether the formulation can represent cohesive laws with sharp changes in the traction–separation response. It further tests the identified model under several successive unload–reload cycles with nonlinear unloading behavior to assess irreversible damage evolution and non-negative dissipation. The second example considers mixed-mode identification when calibration data are available only for the two pure modes and investigates how the predicted intermediate response depends on the constitutive structure imposed on the damage resistance. It also examines the sensitivity of the identification and resulting predictions to network size and random initialization. The third example considers several classical mixed-mode cohesive laws with distinct analytical forms and evaluates prediction at held-out mode mixities. The identified model is further tested under non-proportional loading to examine its response when the mode mixity evolves along the loading path.
All neural-network models and training procedures are implemented in PyTorch [44]. The network architectures, optimization hyperparameters, and implementation details are summarized in Appendix § 12.
Pure Mode I Identification
This example considers several synthetic Mode I traction–separation responses with different loading and unloading behaviors to assess the flexibility and thermodynamic consistency of the proposed formulation. Shear deformation is not considered, so the problem reduces to a one-dimensional cohesive response.
We first synthesize data from two classical Mode I cohesive laws — a trapezoidal law with a constant-strength plateau and an exponential law with a long softening tail — both exhibiting a linear increase in traction up to the peak. The plateau and extended softening tail make these responses challenging to represent accurately using fixed-form classical models, including versatile families such as PPR. The trained model, as shown in Figure 4, reproduces the target responses closely. The sharp corners and the smooth tails are automatically captured by the learned inverse resistance combined with the ICNN elastic potential.
Figure 4. Mode I traction–separation responses for (a) trapezoidal and (b) exponential-tail cohesive laws.
We further consider a nonlinear loading envelope with a concave-to-convex S-shaped unloading response. The model is trained on the monotonic loading envelope, two pre-peak unload–reload cycles that retrace the loading response, and one post-peak unload–reload cycle. The training and evaluation separation trajectories are shown in Figure 5(a), with the corresponding traction–separation responses in panel (b). The evaluation history contains five unload–reload cycles with progressively increasing maximum separation, each unloaded to the origin.
Figure 5. Mode I response under five unload–reload cycles followed by complete separation. (a) Prescribed evaluation history (solid blue) together with the loading segments used for training (dashed black). (b) Predicted traction–separation response (solid blue) compared with the training data (black dots). (c) Cumulative mechanical dissipation (solid blue, left axis) and damage (dashed orange, right axis) predicted by the learned model.
The first two evaluation cycles originate from the pre-peak branch and retrace the elastic loading response. The remaining three originate from the post-peak softening branch and exhibit nonlinear S-shaped unloading responses. Since only one post-peak cycle is included in the training data, the other post-peak branches evaluate the response predicted by the identified model at different damage states.
The cumulative mechanical dissipation is obtained from the intrinsic dissipation in Eq. (17) as
As shown in Figure 5c, both and remain constant during unloading and reloading below the previously attained maximum separation. They increase only when the loading path reaches a new maximum separation. Thus, the learned response satisfies damage irreversibility and non-negative mechanical dissipation throughout the loading history.
Mixed-Mode Identification from Pure-Mode Data
This example examines whether the proposed formulation can identify a mixed-mode cohesive response when calibration data are available only for the two pure modes. We compare the MNN and full-MNN parametrizations of the inverse resistance in terms of pure-mode accuracy, sensitivity to network capacity and initialization, and interpolation across unobserved mode mixities. Both models are trained on a synthetic irregular Mode I response and a half-sine Mode II response, with no intermediate mixed-mode data included during calibration. The two pure-mode responses are intentionally chosen to be markedly different, providing a nontrivial test of mixed-mode interpolation.
For each inverse-resistance architecture, we consider three network sizes and train each configuration from five random initializations. In every case, the active-potential ICNN and the inverse-resistance network are trained jointly using the combined Mode I and Mode II loading data. The ICNN architecture is kept identical across all configurations so that the comparison isolates the effect of the MNN and full-MNN parametrizations of the inverse resistance. For corresponding hidden-layer widths, the full-MNN contains fewer parameters because the unrestricted streams are omitted. The resulting pure-mode predictions are compared below.
The resulting Mode I predictions are shown in Figure 6. At the smaller network capacities, the predictions exhibit noticeable sensitivity to initialization, and some realizations do not reproduce the irregular features of the target response. Increasing the resistance-network capacity reduces the variability across initializations and improves agreement with the data. At the largest configuration, the realizations for both the MNN and full-MNN nearly collapse onto the target Mode I response.
Figure 6. Mode I traction–separation predictions obtained with (a) MNN and (b) full-MNN inverse-resistance parametrizations. Three resistance-network sizes are considered, with five random initializations for each size. Panel titles give the number of resistance-network parameters. The solid blue curve denotes the lowest-loss realization, the faint curves denote the remaining initializations, and the black dots denote the training data.
The corresponding Mode II results are shown in Figure 7. The smaller configurations again exhibit substantial sensitivity to initialization, with several realizations producing poor or irregular responses. This variability decreases markedly as the network capacity increases. At the largest configuration, both architectures closely reproduce the Mode II data with little dependence on initialization. For comparable hidden-layer widths, the full-MNN uses substantially fewer resistance-network parameters than the MNN; at the largest configuration, it uses 106 parameters compared with 354 for the MNN while attaining comparable pure-mode accuracy and reproducibility.
Figure 7. Mode II traction–separation predictions obtained with (a) MNN and (b) full-MNN inverse-resistance parametrizations. Three resistance-network sizes are considered, with five random initializations for each size. Panel titles give the number of resistance-network parameters. The solid blue curve denotes the lowest-loss realization, the faint curves denote the remaining initializations, and the black dots denote the training data.
The more consequential difference between the two parametrizations appears at intermediate mode mixities, which are not represented in the calibration data. We evaluate the highest-capacity models under proportional mixed-mode loading for eight values of . The traction and separation magnitudes are defined as and , respectively. The predicted traction–separation responses and corresponding fracture energies are shown in Figure 8.
Figure 8. Mixed-mode responses predicted by (a) the MNN and (b) the full-MNN after calibration using only pure Mode I and Mode II data. The left column shows traction magnitude versus separation magnitude for eight mode mixities . The right column shows the corresponding fracture energy .
Both models recover the prescribed pure-mode responses at and , but their predictions between these limits differ substantially. Because the dependence of the MNN on is unrestricted, the intermediate responses are weakly constrained by the pure-mode data alone. Consequently, the predicted traction–separation curves may vary irregularly and non-monotonically with mode mixity, as also reflected in the non-monotone . These responses remain thermodynamically admissible under the constitutive assumptions of the model, but thermodynamic admissibility alone does not ensure that an interpolation is physically representative where no data are available. The result therefore illustrates both the flexibility of the MNN and the limited identifiability of mixed-mode behavior from pure-mode data alone.
The full-MNN introduces the additional assumption that the resistance increases with . This inductive bias constrains the otherwise underdetermined interpolation and guarantees a non-decreasing fracture energy, but it represents prior constitutive information rather than behavior inferred from intermediate-mixity data. Reliable identification of the mixed-mode response therefore requires either suitable prior structure or calibration data spanning the relevant range of mode mixities. The latter setting is examined in the next example, where data at several intermediate mixities are included in the calibration and predictions are evaluated at held-out mixities. Additional results for the present example, including the learned constitutive functions and training histories, are provided in Appendix § 13.
Mixed-Mode Identification Across Multiple Mode Mixities
This example evaluates the proposed formulation for mixed-mode cohesive laws using calibration data from multiple mode mixities. The models are calibrated at selected mode mixities and evaluated at held-out intermediate mixities to assess interpolation across the mode-mixity range. We further examine one identified model under non-proportional loading, for which the mode mixity evolves along the loading path, to assess the resulting traction and dissipation histories beyond the proportional trajectories used for calibration.
Data are synthesized from three mixed-mode cohesive laws: an exponential law (CZM-Exp) [55], a bilinear law with a power-law mixed-mode criterion (CZM-BL) [10], and the polynomial potential PPR law [41]. The corresponding traction–separation relations and parameter values are summarized in Appendix § 14. These laws provide three distinct response classes: smooth exponential softening, a bilinear law with a sharp strength transition, and a mode-dependent polynomial response. Because these models are based on distinct analytical cohesive constructions, they provide a direct test of whether the proposed framework can represent different cohesive-law families within the same constitutive structure.
For each law, data are generated along thirteen evenly spaced mode mixities . Ten mode mixities are used for calibration and three are withheld for testing. Each configuration is trained from four random initializations for CZM-Exp and CZM-BL and from eight initializations for PPR; representative results are reported. The inverse resistance is represented by the MNN.
Figure 9 show the normal traction, tangential traction, and fracture energy for the three cohesive laws. The identified models closely reproduce the calibration responses and accurately predict the held-out mode mixities, as quantified in Table 1.
Figure 9. Mixed-mode identification results for (a) CZM-Exp, (b) CZM-BL, and (c) PPR. The left and center columns show the normal and tangential traction–separation responses, respectively, with color indicating the mode mixity . Filled markers denote calibration data and open markers denote the held-out mixities ; the corresponding model predictions are shown by the curves. The right column compares the predicted fracture energy with the analytical reference.
| CZM-Exp | CZM-BL | PPR | |
| Train | |||
| Test |
Table 1. Training and held-out test errors for the three models shown in Figure 9. Each entry is the normalized MSE averaged over the 10 training mixities or the 3 held-out test mixities .
The trajectories considered above are proportional, so the mode mixity remains constant along each loading path. We further examine the model identified from CZM-Exp under non-proportional loading, although no non-proportional
trajectories are included in the calibration data. Figure 10 considers two sequential paths in the – separation plane. The NT path first opens the interface in Mode I to µm and then increases at fixed . The TN path applies the same two loading stages in the opposite order. Both paths terminate at µm. The responses are plotted against the arc length traversed along each path.
Figure 10. Non-proportional response of the model identified from CZM-Exp data (Figure 9a). Two loading paths connect the origin to the same final separation: N→T (blue) first opens the interface in Mode I and then increases the tangential separation at fixed , whereas T→N (red) applies the same stages in the opposite order. (a) Loading paths in the separation plane. (b) Cumulative mechanical dissipation (solid, left axis) and damage (dashed, right axis). (c,d) Normal (solid, left axis) and tangential (dashed, right axis) tractions along the N→T and T→N paths, respectively. The horizontal axis in panels (b)–(d) is the arc length along the loading path. The two paths reach the same final separation, damage, and traction state but accumulate different amounts of mechanical dissipation.
The traction histories depend strongly on the loading sequence. Along NT, the first stage corresponds to Mode I loading, with while rises and subsequently decreases. During the second stage, the tangential traction develops as increases, while the normal traction continues to decrease despite the fixed . The TN path exhibits the corresponding interaction when normal opening is applied after the initial shearing stage. These changes reflect the combined effects of mixed-mode coupling and damage evolution. In particular, because a single scalar damage variable degrades the entire active potential, damage generated by either loading component contributes to degradation of both active traction components.
Despite their different histories, the two paths reach the same final damage in the present test. Because they also terminate at the same separation, the constitutive state is identical at the endpoint, and the predicted terminal traction vectors coincide. The cumulative dissipation, however, differs substantially between the two paths, as shown in Figure 10b. From Eq. (46), , so the dissipated energy depends on the values of the driving force at which the successive damage increments occur. Since during active damage growth, changing the order of normal and tangential loading changes the sequence of mode mixities and driving forces encountered during damage evolution. Consequently, the same final can be reached with different cumulative dissipation. Since both paths attain the same final damage, the larger value of for TN indicates that its damage increments occur at a larger damage-weighted average driving force.
Remark 5.1. Within the present scalar-damage formulation, a mode-dependent resistance generally leads to path-dependent cumulative dissipation under non-proportional loading. Suppose instead that the cumulative dissipation were a state function, . Variations of at fixed are non-dissipative, so along such admissible variations and can depend only on . During active damage growth, Eq. (46) and the consistency condition give
and therefore
If the same damage state can be attained by proportional loading at different mode mixities, the left-hand side is independent of , so the resistance must also be independent of . It would then follow that
is independent of mode mixity. Thus, within the constitutive structure adopted here, a mode-dependent fracture energy represented through generally entails path-dependent dissipation when evolves during damage growth.
Conclusions
This work introduces a data-driven cohesive zone formulation within the generalized standard material framework. Rather than prescribing the traction–separation relation and unloading rules directly, the cohesive response is defined through a learned active potential and a mode-dependent damage resistance. The active potential is represented by an input-convex neural network, while the inverse damage resistance is represented by a monotone neural network. The required convexity, monotonicity, normalization, and damage-irreversibility properties are incorporated directly into the constitutive representation. Direct parametrization of the inverse resistance also provides an explicit damage update and avoids a local nonlinear inversion of the resistance relation during constitutive evaluation.
The material-point studies demonstrate that the formulation can represent a broad range of loading, unloading, and softening responses while maintaining irreversible damage evolution and non-negative dissipation. A single identified model can accommodate distinct Mode I and Mode II responses as well as mixed-mode behavior. When calibration data are restricted to the two pure modes, the intermediate mixed-mode response remains underdetermined. The MNN allows unrestricted dependence on mode mixity, whereas the full-MNN introduces an additional inductive bias by constraining the resistance to increase with mode mixity. When calibration data span several mode mixities, the formulation accurately reproduces the calibration responses and predicts held-out intermediate mixities for several classical mixed-mode cohesive laws.
Overall, the proposed formulation provides a common thermodynamic structure for learning cohesive responses, without prescribing a particular traction–separation law or phenomenological unloading relation.
Limitations and Future Work. Several modeling assumptions delimit the present scope. First, the multiplicative degradation in Eq. (6) implies that, at fixed damage, the active unloading response is governed by the same potential . The present model therefore does not represent residual separation, frictional or plastic hysteresis, damage-dependent changes in the unloading shape, or fatigue degradation. Such behavior would require additional internal variables or evolution mechanisms. Second, a single scalar damage variable degrades the entire active potential and therefore cannot represent independent degradation of the normal and tangential responses. Extensions with multiple damage variables would permit more general anisotropic degradation. In three dimensions, dependence on the tangential separation only through imposes in-plane isotropy and does not distinguish Mode II from Mode III.
A further limitation concerns identifiability. Traction–separation histories do not, in general, uniquely determine the active potential and damage resistance individually, as illustrated in Appendix § 13. Moreover, all identifications considered here use synthetic, noise-free material-point data. The effects of measurement noise, incomplete loading histories, and experimental uncertainty remain to be investigated. The present formulation is restricted to small-deformation, isothermal, and rate-independent behavior and is evaluated at the material-point level. Natural extensions include finite-element implementation for homogeneous and heterogeneous interfaces and inverse identification from structural-scale measurements rather than directly observed traction–separation data. Extensions to rate- and temperature-dependent behavior, cyclic degradation, fatigue, and more general three-dimensional interface anisotropy are also of interest.
Data Availability
All data used in this work are either publicly available or can be reproduced from the information provided and are available from the corresponding author upon reasonable request.
Acknowledgment
B.B. acknowledges support from the startup fund provided by Northwestern University.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
During the preparation of this work the authors used ChatGPT and Claude in order to assist with language improvement. After using the services, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
Elements of Convex Analysis
For completeness we recall, informally, the convex-analysis notions used in the constitutive construction; rigorous statements may be found in standard references [46]. Throughout, is an extended-real-valued function and a convex set.
Convex function. is convex if for all and ; equivalently, its epigraph is a convex set. Convexity underlies both the stability of the stored energy and the non-negativity of the dissipation.
Composition rule. If are convex and is jointly convex and non-decreasing in every argument whose is not affine, then is convex [8 Section 3.2.4].
Lower semicontinuity. is lower semicontinuous if at every , equivalently if all sublevel sets are closed. Intuitively the function may jump up but not down in a limit; this mild requirement ensures that minima are attained and that a potential taking the value on a constraint (as in Eq. (14)) is admissible.
Proper and closed convex function. A convex is proper if it is finite for at least one argument and never equals , and closed if, in addition, it is lower semicontinuous (its epigraph is closed). These are the functions for which the notions below are well behaved.
Subdifferential. For a convex , the subdifferential at is the set of slopes of supporting hyperplanes,
It generalizes the gradient to non-smooth points: where is differentiable it reduces to , whereas at a kink it is a whole set. The evolution law is stated with this notion.
Positive homogeneity of degree one. is positively homogeneous of degree one if for every .
Indicator function. The indicator of is for and otherwise. It is convex, proper, and closed exactly when is convex and closed, and encodes a constraint as an infinite energy penalty — as with the irreversibility constraint .
Normal cone. For , the normal cone is , the set of outward directions at . It equals in the interior and widens on the boundary; it is precisely the subdifferential of the indicator function of .
Convexity Conditions for Composed Cohesive Potentials
Consider a composed function , where is the inner map and is the outer function. We first recall the affine case.
Lemma B.1 (affine map). Let , and let be convex on the range of . Then is convex in . If is invertible, the converse also holds.
Proof. For any , and ,
where the inequality follows from convexity of . Thus, affine precomposition preserves convexity. If is invertible, then
so convexity of also implies convexity of .
Non-affine inner maps require an additional condition. We next consider the case in which the positive-part operator acts on the first component,
Lemma B.2 (positive-part component). Let , where is defined on . Then is convex in if and only if is jointly convex and non-decreasing in its first argument.
Proof. Suppose first that is jointly convex and non-decreasing in its first argument. Since
while the remaining components are affine, monotonicity followed by convexity of gives
Conversely, assume that is convex. On the half-space , the inner map is the identity, so is convex on its domain. Fixing gives the convex function
Since is constant for , convexity implies that it is non-decreasing for . Hence, is non-decreasing in its first argument.
The same argument applies when the positive-part operator acts on several components. In that case, the outer function must be jointly convex and non-decreasing in each rectified argument.
A second case relevant to cohesive laws arises when one component enters through its magnitude.
Lemma B.3 (positive part and a magnitude). Let . Then is convex in if and only if is jointly convex and non-decreasing in its first two arguments.
Proof. Sufficiency follows from the convexity of and , together with joint convexity and monotonicity of in the corresponding arguments.
Conversely, restrict to and , where the inner map is the identity. This gives joint convexity of . Monotonicity in the first argument follows from Lemma B.2. For fixed values of the remaining arguments,
is convex and even. It is therefore non-decreasing for , which gives monotonicity of in its second argument.
We now specialize these results to the interface separation. The coordinate map is affine and therefore does not introduce any additional condition.
If the two sliding directions are distinguished, the active energy has the form . By Lemma B.2, it is convex in if and only if is jointly convex and non-decreasing in its first argument.
More commonly, symmetry with respect to the sliding direction is imposed: . In this case, Lemma B.3 requires to be jointly convex and non-decreasing in both arguments.
Similarly, is convex in if and only if is convex and non-decreasing.
Hyperparameter Settings and Implementation Details
All neural-network models and training procedures are implemented in Python using PyTorch [44].
The active potential is represented by an ICNN with softplus-based activations. To maintain a non-vanishing activation slope, we use . The added linear term preserves convexity and monotonicity while reducing near-saturation of the activation. The separation inputs and energy output of the ICNN are nondimensionalized using characteristic scales so that their numerical magnitudes remain of order one. The separation is normalized by the separation corresponding to the peak traction, while the energy is normalized by the area under the corresponding traction–separation curve.
The inverse resistance is represented by the MNN. In mixed-mode problems, the confluence and nonlinear streams use the same widths. The full-MNN is considered only in the pure-mode interpolation study of § 5.2; in that case, the unrestricted streams are omitted and the architecture is specified by the exponential-stream widths alone. The driving force is passed directly to the MNN. A small linear term is added to the MNN output to prevent the slope from becoming arbitrarily small when the bounded activation functions approach saturation. We use .
The restriction of the learned inverse resistance to the admissible damage range is imposed pointwise at each constitutive update. To improve numerical smoothness near damage onset during training, we replace the positive-part operator by the scaled softplus approximation
where controls the width of the smoothed transition. The onset corner is therefore regularized, while the upper bound remains a hard saturation so that and complete damage is attained exactly. As , , and Eq. (47) recovers the hard clamp in the limiting case.
All models are trained using Adam [27] with an initial learning rate of and cosine annealing with warm restarts [34]. Unless otherwise stated, the learning rate is annealed to within each cycle. The first cycle spans one fifth of the total epoch budget, and the cycle length is doubled after each restart. For the examples in § 5.3, we instead use a single restart after the first fifth of the epoch budget. The second cycle then spans the remaining epochs and anneals to . Gradient norms are clipped at 50 [43].
Problem-specific details are provided below.
Section 5.1 The study of Figure 4 uses an ICNN of three hidden layers of eight units each and an MNN of two hidden layers of four units each; the unload–reload study of Figure 5 uses a larger ICNN of three hidden layers of sixteen units each and an MNN of two hidden layers of eight units each, since both the loading and the unloading responses are more complex. Both are trained for 40,000 epochs with .
Section 5.2 The ICNN is held at three hidden layers of eight units each in every case, so the comparison isolates the inverse-resistance representation. Three resistance capacities are considered for each architecture — two hidden layers of two, four, or eight units — giving 42, 114, and 354 parameters for the MNN and 16, 38, and 106 for the full-MNN. The confluence and nonlinear streams use the same widths. Each configuration is trained for 30,000 epochs with . An regularization penalty with coefficient is applied to the MNN weights and biases to reduce overfitting.
Section 5.3 The ICNN and MNN each use two hidden layers with eight units per layer. Training is performed for 50,000 epochs with .
No systematic hyperparameter search or architecture optimization was performed; the network sizes were selected empirically to provide sufficient representational capacity for each example. The reported configurations should therefore not be interpreted as optimal architectures.
Additional Mixed-Mode Results
This appendix provides additional results for the mixed-mode example of § 5.2. We report the optimization histories for the different MNN and full-MNN capacities and compare the active potentials and damage resistances identified by the highest-capacity models.
Figure 11 shows the training loss for the three network capacities considered for each inverse-resistance representation, with five random initializations per configuration. The optimization exhibits greater sensitivity to initialization for the smaller networks, while the highest-capacity models attain the lowest terminal losses. The transient increases in the loss coincide with the learning-rate restarts described in Appendix § 12.
Figure 11. Training-loss histories for (a) the MNN and (b) the full-MNN. Three resistance-network capacities are considered, with five random initializations for each configuration. Panel titles give the number of resistance-network parameters. The solid blue curve denotes the lowest-loss realization and the faint curves denote the remaining initializations.
Although the MNN and full-MNN reproduce the observed pure-mode traction–separation responses with comparable accuracy, the corresponding latent constitutive functions need not be unique. Figure 12 and Figure 13 show the active potential and damage resistance identified by the highest-capacity model of each architecture. The two models recover noticeably different and while producing similar responses on the calibration paths. This illustrates that traction–separation data constrain the observable constitutive response more directly than the individual latent functions.
Figure 12. Active potential identified by (a) the MNN and (b) the full-MNN models, evaluated along proportional loading directions as a function of separation magnitude and mode mixity .
Figure 13. Damage resistance identified by (a) the MNN and (b) the full-MNN models.
The difference between the resistance surfaces also reflects the distinct assumptions imposed by the two architectures. The MNN leaves the dependence on unrestricted, whereas the full-MNN constrains the resistance to increase with mode mixity. Consequently, the full-MNN produces the monotonic fracture-energy trend reported in Figure 8, while the unconstrained MNN admits the non-monotonic interpolation observed there.
Reference Cohesive Laws for the Classical-Model Study
The three reference cohesive laws used in § 5.3 are summarized below. We consider the modified exponential law of Van den Bosch et al. [55], a bilinear damage law based on Camanho et al. [10], and the PPR potential of Park et al. [41]. The parameters are chosen as
For the exponential and bilinear laws, we set , while the PPR law uses . Synthetic data are generated along proportional loading paths with fixed mode mixity , as defined in Eq. (4). Only positive normal and tangential separations are used in the reference loading paths.
Remark E.1. In the following, some symbols are reused locally to avoid an unnecessary proliferation of notation; their meaning will be clear from context.
Exponential law (CZM-Exp). We use the modified exponential mixed-mode law of Van den Bosch et al. [55],
where
and
with . Here and denote the normalized normal and tangential separations, respectively. These definitions give the prescribed pure-mode peak tractions and and fracture energies and .
Bilinear law (CZM-BL). We use a bilinear damage law based on Camanho et al. [10], combined here with a power-law mixed-mode fracture criterion. A common penalty stiffness
is used in the normal and tangential directions. The pure-mode damage-onset separations are
Defining the effective separation and loading-direction ratio as
the mixed-mode onset separation is
Its pure Mode II limit is .
The mixed-mode fracture energy is prescribed by
where
The corresponding complete-failure separation is
Let
denote the maximum effective separation attained along the loading history. The damage variable is
The tractions are then
Park–Paulino–Roesler law (PPR). The PPR potential of Park et al. [41] is used with the shape parameters and , where denotes the tangential PPR exponent and is distinct from the mode-mixity variable . The initial-slope parameters are and . The corresponding exponents are
Following Park et al. [41], and denote the PPR shape exponents.
The energy constants are
and
For the present choice , these reduce to
The final normal and tangential separations are determined from the prescribed pure-mode cohesive strengths,
Introducing
the normal traction is
and the tangential traction is
These expressions are evaluated over and for the proportional loading paths considered here. The corresponding traction component is set to zero once its terminal separation is exceeded.
References
- [1]Giulio Alfano and Michael A Crisfield. Finite element interface models for the delamination analysis of laminated composites: mechanical and computational issues. International Journal for Numerical Methods in Engineering, 50(7):1701–1736, 2001.DOI
- [2]A. Amiri-Hezaveh and A. B. Tepole. A physics-augmented machine learning constitutive model for damage in solids. arXiv preprint arXiv:2508.05638, 2025.
- [3]Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. In International Conference on Machine Learning, pages 146–155. PMLR, 2017.arxiv.org/abs/1609.07152
- [4]Mohammad A Ansari, Rujing Zha, Rui Huang, and Kenneth M Liechti. Rotation control with dual actuators: A new device to extract mixed-mode traction–separation relations. Engineering Fracture Mechanics, 326:111318, 2025.
- [5]Faisal As’ad, Philip Avery, and Charbel Farhat. A mechanics-informed artificial neural network approach in data-driven constitutive modeling. International Journal for Numerical Methods in Engineering, 123(12):2738–2759, 2022.
- [6]G Bao and Z Suo. Remarks on crack-bridging concepts. Applied Mechanics Reviews, 45(8):355–366, 1992.DOI
- [7]Grigory I. Barenblatt. The mathematical theory of equilibrium cracks in brittle fracture. Advances in Applied Mechanics, 7:55–129, 1962.DOI
- [8]Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
- [9]Godofredo T Camacho and Michael Ortiz. Computational modelling of impact damage in brittle materials. International Journal of Solids and Structures, 33(20-22):2899–2938, 1996.DOI
- [10]Pedro Ponces Camanho, Carlos G Dávila, and Mário F de Moura. Numerical simulation of mixed-mode progressive delamination in composite materials. Journal of Composite Materials, 37(16):1415–1438, 2003.
- [11]HC Cao and AG Evans. An experimental study of the fracture resistance of bimaterial interfaces. Mechanics of Materials, 7(4):295–304, 1989.
- [12]Hennie Daniels and Marina Velikova. Monotone and partially monotone neural networks. IEEE Transactions on Neural Networks, 21(6):906–917, 2010.DOI
- [13]NV De Carvalho, MW Czabaj, and JG Ratcliffe. Piecewise-linear generalizable cohesive element approach for simulating mixed-mode delamination. Engineering Fracture Mechanics, 242:107484, 2021.DOI
- [14]Rossan Dimitri, Marco Trullo, Laura De Lorenzis, and Giorgio Zavarise. Coupled cohesive zone models for mixed-mode fracture: A comparative study. Engineering Fracture Mechanics, 148:145–179, 2015.DOI
- [15]Donald S. Dugdale. Yielding of steel sheets containing slits. Journal of the Mechanics and Physics of Solids, 8 (2):100–104, 1960.DOI
- [16]M. Flaschel, P. Steinmann, L. De Lorenzis, and E. Kuhl. Convex neural networks learn generalized standard material models. Journal of the Mechanics and Physics of Solids, 200:106103, 2025.
- [17]Konrad Friedrichs, Franz Dammäß, Karl A Kalina, and Markus Kästner. Precise, efficient and flexible modeling of crystallizing elastomers based on physics-augmented neural networks. Computer Methods in Applied Mechanics and Engineering, 455:118852, 2026.arxiv.org/abs/2511.14553
- [18]Jan N Fuhg, Govinda Anantha Padmanabha, Nikolaos Bouklas, Bahador Bahmani, WaiChing Sun, Nikolaos N Vlassis, Moritz Flaschel, Pietro Carrara, and Laura De Lorenzis. A review on data-driven constitutive laws for solids. Archives of Computational Methods in Engineering, 32(3):1841–1883, 2025.arxiv.org/abs/2405.03658
- [19]Philippe H Geubelle and Jeffrey S Baylor. Impact-induced delamination of composites: a 2D simulation. Composites Part B: Engineering, 29(5):589–602, 1998.DOI
- [20]Jamshid Ghaboussi, JH Garrett Jr, and Xiping Wu. Knowledge-based modeling of material behavior with neural networks. Journal of Engineering Mechanics, 117(1):132–153, 1991.DOI
- [21]Bernard Halphen and Quoc Son Nguyen. Sur les matériaux standard généralisés. Journal de mécanique, 14(1):39–63, 1975.
- [22]Arne Hillerborg, Mats Modéer, and P-E Petersson. Analysis of crack formation and crack growth in concrete by means of fracture mechanics and finite elements. Cement and Concrete Research, 6(6):773–781, 1976.DOI
- [23]Dirk A Hordijk and Hans W Reinhardt. Growth of discrete cracks in concrete under fatigue loading. In Toughening Mechanisms in Quasi-Brittle Materials, pages 541–554. Springer, 1991.DOI
- [24]J. W. Hutchinson and Z. Suo. Mixed mode cracking in layered materials. Advances in Applied Mechanics, 29:63–191, 1991.DOI
- [25]Noboru Kikuchi and John Tinsley Oden. Contact Problems in Elasticity: A Study of Variational Inequalities and Finite Element Methods. SIAM Studies in Applied Mathematics. Society for Industrial and Applied Mathematics, Philadelphia, 1988.
- [26]Hyunho Kim and Jong-Seok Lee. Scalable monotonic neural networks. In The Twelfth International Conference on Learning Representations, 2024.
- [27]Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learning Representations, 2015.arxiv.org/abs/1412.6980
- [28]Dominik K Klein, Mauricio Fernández, Robert J Martin, Patrizio Neff, and Oliver Weeger. Polyconvex anisotropic hyperelasticity with neural networks. Journal of the Mechanics and Physics of Solids, 159:104703, 2022.DOI
- [29]Meinhard Kuna and Stephan Roth. General remarks on cyclic cohesive zone models. International Journal of Fracture, 196(1):147–167, 2015.DOI
- [30]Bernhard Lang. Monotonic multi-layer perceptron networks as universal approximators. In International Conference on Artificial Neural Networks, pages 31–37. Springer, 2005.DOI
- [31]M. Leuschner, F. Fritzen, J. A. W. van Dommelen, and J. P. M. Hoefnagels. Potential-based constitutive models for cohesive interfaces: Theory, implementation and examples. Composites Part B: Engineering, 68:38–50, 2015.DOI
- [32]K. M. Liechti and Y. S. Chai. Asymmetric shielding in interfacial fracture under in-plane shear. Journal of Applied Mechanics, 59(2):295–304, June 1992. doi:10.1115/1.2899520.DOI
- [33]Shengduo Liu, Kaushik Bhattacharya, and Nadia Lapusta. Learning a potential formulation for rate-and-state friction. Mechanics of Materials, 212:105540, 2026.arxiv.org/abs/2507.09796
- [34]Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv:1608.03983, 2016.arxiv.org/abs/1608.03983
- [35]J Patrick McGarry, Éamonn Ó Máirtín, Guillaume Parry, and Glenn E Beltz. Potential-based and non-potential-based cohesive zone formulations under mixed-mode separation and over-closure. Part I: Theoretical analysis. Journal of the Mechanics and Physics of Solids, 63:336–362, 2014.DOI
- [36]J Mosler and Ingo Scheider. A thermodynamically and variationally consistent class of damage-type cohesive models. Journal of the Mechanics and Physics of Solids, 59(8):1647–1668, 2011.DOI
- [37]Alan Needleman. A continuum model for void nucleation by inclusion debonding. Journal of Applied Mechanics, 54(3):525–531, 1987.DOI
- [38]O Nguyen, EA Repetto, Michael Ortiz, and RA Radovitzky. A cohesive model of fatigue crack growth. International Journal of Fracture, 110(4):351–369, 2001.DOI
- [39]Michael Ortiz and Anna Pandolfi. Finite-deformation irreversible cohesive elements for three-dimensional crack-propagation analysis. International Journal for Numerical Methods in Engineering, 44(9):1267–1282, 1999.
- [40]K. Park and G. H. Paulino. Computational implementation of the PPR potential-based cohesive model in ABAQUS: Educational perspective. Engineering Fracture Mechanics, 93:239–262, 2012.DOI
- [41]K. Park, G. H. Paulino, and J. R. Roesler. A unified potential-based cohesive model of mixed-mode fracture. Journal of the Mechanics and Physics of Solids, 57(6):891–908, 2009.DOI
- [42]Kyoungsoo Park and Glaucio H Paulino. Cohesive zone models: a critical review of traction-separation relationships across fracture surfaces. Applied Mechanics Reviews, 64(6):060802, 2011.DOI
- [43]Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In International conference on machine learning, pages 1310–1318. PMLR, 2013.arxiv.org/abs/1211.5063
- [44]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.arxiv.org/abs/1912.01703
- [45]Michel Raous, Laurent Cang𝑒mi, and Marius Cocu. A consistent model coupling adhesion, friction, and unilateral contact. Computer Methods in Applied Mechanics and Engineering, 177(3-4):383–399, 1999.DOI
- [46]R Tyrrell Rockafellar. Convex analysis, volume 28. Princeton University Press, 1997.
- [47]T. Roubíček, C. G. Panagiotopoulos, and V. Mantič. A quasistatic mixed-mode delamination model. Discrete and Continuous Dynamical Systems-Series S, 6(2):591–610, 2013.DOI
- [48]Bent F Sørensen and Torben K Jacobsen. Large-scale bridging in composites: R-curves and bridging laws. Composites Part A: Applied Science and Manufacturing, 29(11):1443–1451, 1998.DOI
- [49]Bent F Sørensen and Torben K Jacobsen. Determination of cohesive laws by the J integral approach. Engineering Fracture Mechanics, 70(14):1841–1858, 2003.DOI
- [50]Chongcong Tao, Chao Zhang, Hongli Ji, and Jinhao Qiu. Neural cohesive model for delamination simulation in composite laminates under cyclic loadings. International Journal of Fatigue, 164:107123, 2022.DOI
- [51]Chongcong Tao, Chao Zhang, Hongli Ji, and Jinhao Qiu. Reconstruction and prediction of mode-i cohesive law using artificial neural network. Composites Science and Technology, 256:110755, 2024.DOI
- [52]Prakash Thakolkaran, Akshay Joshi, Yiwen Zheng, Moritz Flaschel, Laura De Lorenzis, and Siddhant Kumar. NN-EUCLID: Deep-learning hyperelasticity without stress data. Journal of the Mechanics and Physics of Solids, 169:105076, 2022.DOI
- [53]A Turon, Pedro Ponces Camanho, J Costa, and CG Dávila. A damage model for the simulation of delamination in advanced composites under variable-mode loading. Mechanics of Materials, 38(11):1072–1089, 2006.DOI
- [54]Viggo Tvergaard and John W Hutchinson. The relation between crack growth resistance and fracture process parameters in elastic-plastic solids. Journal of the Mechanics and Physics of Solids, 40(6):1377–1397, 1992.DOI
- [55]MJ Van den Bosch, PJG Schreurs, and MGD Geers. An improved description of the exponential Xu and Needleman cohesive zone law for mixed-mode decohesion. Engineering Fracture Mechanics, 73(9):1220–1234, 2006.
- [56]Nikolaos N. Vlassis and WaiChing Sun. Sobolev training of thermodynamic-informed neural networks for interpretable elasto-plasticity models with level set hardening. Computer Methods in Applied Mechanics and Engineering, 377:113695, 2021.
- [57]R. Vodička and V. Mantič. An energy based formulation of a quasi-static interface damage model with a multilinear cohesive law. Discrete and Continuous Dynamical Systems-Series S, 10(6):1539–1561, 2017.DOI
- [58]Kun Wang and WaiChing Sun. Meta-modeling game for deriving theory-consistent, microstructure-based traction–separation laws via deep reinforcement learning. Computer Methods in Applied Mechanics and Engineering, 346:216–241, 2019.DOI
- [59]Congjie Wei, Jiaxin Zhang, Kenneth M Liechti, and Chenglin Wu. Data driven modeling of interfacial traction–separation relations using a thermodynamically consistent neural network. Computer Methods in Applied Mechanics and Engineering, 404:115826, 2023.
- [60]Congjie Wei, Jiaxin Zhang, Kenneth M Liechti, and Chenglin Wu. Extracting interfacial interactions using elastic field decomposition and a deep complex neural network. International Journal of Solids and Structures, 334:113980, 2026.
- [61]K Willam, I Rhee, and B Shing. Interface damage model for thermomechanical degradation of heterogeneous materials. Computer Methods in Applied Mechanics and Engineering, 193(30-32):3327–3350, 2004.DOI
- [62]H. Xu, M. Flaschel, and L. De Lorenzis. Discovering non-associated pressure-sensitive plasticity models with EUCLID. Advanced Modeling and Simulation in Engineering Sciences, 12(1), 2025.DOI
- [63]X. P. Xu and A. Needleman. Numerical simulations of fast crack growth in brittle solids. Journal of the Mechanics and Physics of Solids, 42:1397–1434, 1994.DOI
- [64]X-P Xu and Alan Needleman. Void nucleation by inclusion debonding in a crystal matrix. Modelling and Simulation in Materials Science and Engineering, 1(2):111–132, 1993.DOI
- [65]B Yang and K Ravi-Chandar. A single-domain dual-boundary-element formulation incorporating a cohesive zone model for elastostatic cracks. International Journal of Fracture, 93(1):115–144, 1998.DOI











