Graded Damage: Models & Applications

• Graded damage is a continuum concept that models partial material degradation using scalar, tensorial, or non-local variables to capture spatial distribution and intensity.
• It regularizes softening in simulations by incorporating gradient regularization and internal length scales, thereby controlling localization and ensuring mesh independence.
• Beyond mechanics, graded damage is applied in disaster assessment and AI safety, grading damage severity with ordinal or pixel-wise measures for improved evaluation.

Searching arXiv for recent and foundational papers on graded damage across mechanics and damage grading. Graded damage denotes representations of degradation that preserve degree, spatial distribution, or severity rather than collapsing failure into an intact/failed or success/failure bit. In continuum mechanics, it appears as scalar, tensorial, gradient-enhanced, or non-local internal variables that encode partial stiffness loss, anisotropy, and finite-width process zones; in adaptive multi-scale methods, it also denotes transfer procedures that preserve accumulated damage during refinement. In disaster assessment and AI safety evaluation, the same idea appears as ordinal damage grades, pixel-wise severity fields, or action-graded harm scales that retain information about extent, directionality, reversibility, and cross-scope consequence that binary labels suppress (Valoroso, 2023, Müller et al., 2023, Zhang et al., 3 Jul 2025, Owiredu-Ashley, 8 Jul 2026).

1. Constitutive meaning and state-variable conventions

In mechanics, graded damage is not tied to a single state variable convention. One strand uses a scalar damage variable with $0 \leq \alpha \leq 1$, where $\alpha = 0$ denotes undamaged material and $\alpha = 1$ fully damaged material; this convention is explicit in the gradient damage model for slender rods (Bonnetier et al., 2 Jan 2026). Another strand uses a scalar field $\chi \in [0,1]$ representing the local proportion of undamaged bonds, so that $\chi = 1$ is undamaged and $\chi = 0$ fully damaged; this convention is used in the rate-independent model coupling damage with plasticity via structured strains (Bonetti et al., 2015). A related reversible-damage formulation employs a scalar variable $\zeta \in [0,1]$, with both elastic moduli and plastic yield stress depending on $\zeta$, and with healing admitted through the dissipation potential (Roubíček et al., 2015). The literature therefore uses both “damage fraction” and “undamaged fraction” conventions.

Graded damage also appears in explicitly anisotropic form. In the damage-preserving transformation for materials with microstructure, the coarse-scale damage variable is a symmetric second-order tensor $\mathbf{D}$, entering the constitutive law as

$$\boldsymbol{\sigma} = (\mathbf{I} - \mathbf{D}) : \mathbf{C} : \boldsymbol{\varepsilon},$$

with $\alpha = 0$0 and $\alpha = 0$1 given by the maximum principal tensile strains in the $\alpha = 0$2 and $\alpha = 0$3 directions. Each component follows a power-law relation, $\alpha = 0$4 and $\alpha = 0$5, with a simple coupling for biaxial loading (Müller et al., 2023). This formulation is explicitly orthotropic, not merely scalar or isotropic tensorial.

A different constitutive route derives graded damage from probability. The CDF-generated damage laws define the stored-energy density by

$\alpha = 0$6

with $\alpha = 0$7, where $\alpha = 0$8 is the tensile part of the strain energy density and only the tensile contribution is degraded. The choice of $\alpha = 0$9 spans exponential, Cauchy, logistic, half-normal, Gudermannian, hypergeometric, radical, rational, piece-wise, and rapid-decay cumulative-distribution functions (Ren, 7 Jun 2025). In this class, graded damage is interpreted as the probability that the local tensile energy exceeds a random microstructural threshold.

The one-dimensional “Graded damage” model introduces yet another formulation: instead of a quadratic gradient penalty, it imposes an explicit bound on the damage gradient,

$\alpha = 1$0

thereby prescribing the steepest admissible damage profile directly (Valoroso, 2023). This shifts regularization from an energetic penalty to a hard kinematic constraint on spatial variation.

2. Gradient regularization, non-locality, and internal length scales

A central reason graded damage appears in modern constitutive models is the need to regularize softening and avoid pathological localization. In the slender-rod reduction of a three-dimensional gradient damage model, the total energy contains the damage gradient term

$\alpha = 1$1

and, after nondimensionalization and the limit $\alpha = 1$2, the $\alpha = 1$3-limit is a one-dimensional functional in which only $\alpha = 1$4 survives. The resulting damage field becomes independent of transverse coordinates and remains smoothly graded only along the rod axis (Bonnetier et al., 2 Jan 2026). The paper states that cross-sectional variations become energetically prohibitive because the transverse gradient terms scale with $\alpha = 1$5.

Equivalent regularization appears in finite-strain formulations. The mixed finite element for gradient damage is based on the energy potential

$\alpha = 1$6

and is designed to yield robust, efficient, and mesh-independent simulations without numerical stabilization or penalty parameters (Riesselmann et al., 2022). The large-deformation gradient-enhanced damage model uses

$\alpha = 1$7

with $\alpha = 1$8; the paper reports mesh-objective global and local responses and states that the extra cost of damage computation is negligible relative to elasticity in the proposed numerical treatment (Junker et al., 2021).

In healing models for soft tissues, non-locality is expressed through an auxiliary field $\alpha = 1$9. Both proposed constitutive laws include

$\chi \in [0,1]$0

so that $\chi \in [0,1]$1 controls the width of damaged and healing regions. The paper explicitly states that increasing $\chi \in [0,1]$2 broadens the predicted zones and that the gradient terms introduce intrinsic tissue length scales and overcome mesh sensitivity (He et al., 2019).

These models share a common structural role for the internal length scale: it fixes the width of graded damage or healing zones and prevents collapse to zero-width bands. A plausible implication is that the term “graded” in this literature is as much about admissible spatial structure as about partial loss of stiffness.

3. Admissibility, localization control, and the transition to fracture

The analytical status of graded damage laws is addressed directly in the CDF-based framework. For any CDF-based degradation map, the paper proves monotonicity, boundedness, and dissipativity, with

$\chi \in [0,1]$3

and concludes that the associated hyperelastic material is thermodynamically admissible (Ren, 7 Jun 2025). The same work establishes compactness and $\chi \in [0,1]$4-convergence in $\chi \in [0,1]$5 to a sharp-interface Griffith functional and proves the existence of rate-independent quasi-static evolutions via global energetic solutions satisfying stability and energy balance. Within this framework, graded damage is a regularized precursor that rigorously converges to brittle fracture.

A persistent difficulty, however, is that some gradient or non-local damage laws broaden the damage band as softening proceeds. The modified non-local damage model identifies two causes of this pathology: a non-vanishing thermodynamic driving force at full damage and a non-decaying forcing term in the non-local equation. The unified MNLD modification changes both the stress degradation function and the forcing term so that the driving force $\chi \in [0,1]$6 vanishes as $\chi \in [0,1]$7 and the non-local source decays through

$\chi \in [0,1]$8

thereby producing fixed-width damage bands in the reported benchmarks (Saji et al., 30 Jun 2025). The paper explicitly contrasts this with conventional gradient damage models that exhibit unrealistic widening.

A related issue arises in elastomers. The chain stretch-based gradient-enhanced damage model states that it allows fracture to localize while also capturing the development of a physically diffuse damage zone, in contrast to the phase-field paradigm in which a sharp crack is numerically approximated in a diffuse manner. Its non-local chain-stretch equation includes both a bounded driving force and a relaxation function,

$\chi \in [0,1]$9

specifically introduced to prevent spurious broadening and to shut off non-local communication in fully damaged regions (Mousavi et al., 5 Feb 2025). The same paper states that fracture toughness is realized as an output rather than an input.

The one-dimensional graded-damage formulation offers a further alternative: instead of modifying the driving force or free energy, it prescribes the spatial damage profile through the bound $\chi = 1$0. The paper derives analytical solutions for a tensile rod and a mode-I delamination problem, with the cohesive law formulated starting from the graded-damage concept by prescribing the shape of damage distribution within the cohesive process zone (Valoroso, 2023). This is a direct example of graded damage being used to define, rather than merely regularize, process-zone structure.

4. Coupling with plasticity, healing, and multi-scale transfer

Graded damage often functions as one component of a larger inelastic system. In the rate-independent model coupling damage with plasticity via structured strains, the symmetric strain is decomposed as

$\chi = 1$1

with the specific choice $\chi = 1$2. This enforces that inelastic strain is active only where the material is damaged; when $\chi = 1$3, $\chi = 1$4, and when $\chi = 1$5, $\chi = 1$6 (Bonetti et al., 2015). The damage evolution is rate-independent and irreversible through the dissipation term $\chi = 1$7, and existence is established in the energetic framework of Mielke and coworkers.

The small-strain perfect-plasticity model with damage and healing extends this logic by allowing reversible damage. Its free energy includes

$\chi = 1$8

and both elastic moduli and yield stress depend on $\chi = 1$9. The damage dissipation potential is convex and symmetric in positive and negative rates, so healing is admitted, and the paper proves existence of weak solutions by a fractional-step time discretization with numerical stability and convergence (Roubíček et al., 2015). The computational implementation splits elastoplastic and damage subproblems, solving the latter by quadratic programming.

Healing as a graded, non-local process is developed further in soft tissues. One model combines gradient-enhanced damage with a temporally homogenized growth-and-remodeling law in which the growth direction is determined according to local principal stress directions; the second is based on a gradient-enhanced healing model with continuously recoverable damage (He et al., 2019). In both, healing changes the damage field and, in one formulation, the geometry in the media layer during balloon angioplasty simulations.

At the scale-transfer level, graded damage becomes a consistency condition between coarse and fine representations. The damage-preserving transformation for materials with microstructure calibrates the continuum damage law not to experimental macroscopic material data but to homogenized degradation obtained from explicit microstructural simulations. Upon refinement, the continuum state variable $\chi = 0$0 is mapped to fractions of failed microstructural elements $\chi = 0$1, and pre-damaged microstructures are reconstructed stochastically with probabilities depending on element direction and strength (Müller et al., 2023). The paper states that the generated fine-scale damage patterns are overall consistent with explicitly simulated damage patterns, with minor discrepancies that subsequently vanish when explicit damage evolution continues under increased load.

5. Numerical formulations and reduction strategies

The numerical treatment of graded damage is itself a significant research topic because regularization must be enforced without compromising robustness or computational cost. The Lagrange-multiplier mixed finite element for gradient damage enforces irreversibility through the KKT conditions

$\chi = 0$2

while using a $\chi = 0$3 interpolation for displacement, a $\chi = 0$4 plus bubble interpolation for damage, and an elementwise constant $\chi = 0$5 interpolation for the Lagrange multiplier (Riesselmann et al., 2022). The paper emphasizes that the multiplier and bubble degrees of freedom are statically condensed at element level, so no additional global degrees of freedom arise.

The large-deformation model uses a neighbored-element method in which displacement is solved by standard finite elements and damage is updated on an element-centered grid using finite-difference approximations of the Laplacian. The damage evolution satisfies KKT conditions of the form

$\chi = 0$6

and the paper augments this with an element erosion technique for severely damaged material (Junker et al., 2021). Once an element exceeds a critical damage threshold, its residual force is set to zero, stiffness is reduced to a negligible value, and damage evolution in that element is frozen.

Dimension reduction provides a different computational simplification. The slender-rod analysis shows that minimizers of the three-dimensional model converge to minimizers of a one-dimensional energy defined on fields independent of transverse coordinates, with the limiting strain approaching a diagonal form indicative of uniaxial deformation (Bonnetier et al., 2 Jan 2026). This furnishes a rigorous route from three-dimensional graded-damage mechanics to one-dimensional rod models.

Adaptive multi-scale simulation provides still another strategy. The damage-preserving transformation is explicitly designed for domain-decomposed multi-scale methods with dynamical refinement, so that regions refined from continuum to lattice or beam-truss representations do not lose accumulated damage history (Müller et al., 2023). This suggests that computational economy and constitutive fidelity are treated jointly: refinement is not enough unless damage itself is transferred consistently.

6. Damage grading in sensing, disaster assessment, and safety evaluation

Outside constitutive mechanics, graded damage is operationalized as an ordinal or pixel-wise severity representation. In earthquake building assessment using ensemble machine learning and deep learning, the target variable is the building damage grade encoded as three classes $\chi = 0$7. The workflow includes label encoding, anomaly removal with Isolation Forest, feature selection by SelectKBest with ANOVA F-test, balancing by SMOTE and RandomUnderSampler, and multi-class classifiers including Logistic Regression, Decision Tree, Random Forest, GBM, AdaBoost, LightGBM, and XGBoost, together with Voting, Bagging, Stacking, FFN, and KAN models (Panda et al., 27 Jun 2025). The paper presents damage grades as a multi-class classification problem rather than a binary damaged/undamaged task.

A more explicitly spatial approach is proposed for post-earthquake social-media imagery. The segmentation model classifies each pixel into Undamaged Structure, Damaged Structure, Debris, or Background, using a SegFormer fine-tuned on 547 expert-labeled images, and then computes a depth-corrected damage severity score with DSweight $\chi = 0$8 using depth predicted by DPT (Zhang et al., 3 Jul 2025). The paper argues that image-level classification is subjective and incapable of accounting for varying extents of damage within an image; the proposed score instead aggregates area, severity class, and relative depth.

In photogrammetric point-cloud assessment, graded damage is defined by four structural grades—No damage, Heavy damage, Extreme damage, and Destruction—derived from an expert damage catalogue aligned with EMS-98 and limited to damage patterns visible in UAV-borne 3D data. A random forest trained on virtual laser scanning data is then transferred to real dense image matching point clouds, yielding overall accuracies of $\chi = 0$9 to $\zeta \in [0,1]$0 (Zahs et al., 2023). The paper states that region-specific real training data improves performance only slightly.

Flood damage assessment introduces a different ordinal structure. Flood-DamageSense predicts four building-damage states—No Damage $\zeta \in [0,1]$1, Minor Damage $\zeta \in [0,1]$2, Moderate Damage $\zeta \in [0,1]$3, and Major Damage $\zeta \in [0,1]$4—from pre- and post-event SAR/InSAR, very-high-resolution optical basemaps, and a historical flood-risk layer, using a multimodal Mamba backbone with semi-Siamese encoding, FFSS fusion, and task-specific decoders for building damage, floodwater extent, and building localization (Ho et al., 7 Jun 2025). The paper reports a mean F1 improvement of up to 19 percentage points over strong baselines and identifies the inherent flood-risk feature as the single most significant contributor.

The most abstract severity scale in the supplied literature is the action-graded harm rubric for tool-using AI agents. It grades executed actions on a seven-level ordinal scale $\zeta \in [0,1]$5 to $\zeta \in [0,1]$6 according to reversibility, scope crossing, and privilege expansion, using either a deterministic oracle over tool-call trajectories or a panel of three frontier language-model judges (Owiredu-Ashley, 8 Jul 2026). The paper reports that the judge panel reproduces the oracle with high ordinal agreement, $\zeta \in [0,1]$7, but shares systematic blind spots, especially failure to recognize escalation chains. In this context, graded damage no longer refers to material degradation but to severity-preserving evaluation of harmful outcomes.

The cross-domain commonality is exact rather than metaphorical: each of these systems replaces a binary decision with an ordered or spatially resolved representation of damage. This suggests that graded damage has become a general strategy for retaining consequential information that coarse labels erase.