Synthetic Defects: Control & Applications

• Synthetic defects are deliberately engineered or computationally generated configurations that serve as controlled proxies in materials science, quantum systems, and machine learning.
• They enable precise analysis of defect phenomena in physical systems and enhance training data by simulating realistic anomalies for robust model evaluation.
• Applications span solid-state materials, topological quantum systems, industrial inspection, and software quality, each leveraging specific synthetic strategies and methodologies.

Searching arXiv for the cited papers and closely related work to ground the article in current arXiv metadata. arXiv search: "Synthetic defect generation review industrial anomaly detection defect GAN" Synthetic defects are deliberately engineered or computationally generated defect configurations introduced for control, study, simulation, or data augmentation. In the literature represented here, the term spans at least two major uses. In materials and quantum systems, it denotes intentionally created defect states whose orientation, binding configuration, topology, or localization is part of the design objective, as in nitrogen-vacancy centers in diamond, luminescent $sp^3$-defects in single-walled carbon nanotubes, finite-area defects in graphene, and synthetic twist defects in the surface code (Lesik et al., 2014, Settele et al., 2021, Cockayne, 2011, Kairys et al., 11 May 2026). In data-centric machine learning and inspection, it denotes synthetic defect samples generated procedurally, physically, or generatively in order to train or evaluate models when real defects are scarce, as in industrial anomaly detection, photolithographic SEM inspection, infrastructure imagery, visual restoration, and logging-statement repair (Mou et al., 2022, Krassnig et al., 6 Mar 2025, Shinde et al., 15 May 2025, Mandati et al., 2024, Sinha et al., 28 May 2025, Zhong et al., 2024).

1. Scope and domain structure

Across these works, synthetic defects are not a single object class but a family of controllable defect representations whose purpose depends on the host system and downstream task. In physical systems, the defect itself is the target of synthesis. In machine-learning pipelines, the defect image, mask, mesh, or text mutation is synthesized primarily as supervision, augmentation, or simulation input.

Domain	Synthetic-defect construct	Representative papers
Solid-state materials	NV orientation, $sp^3$-binding configuration, graphitic defect loops	(Lesik et al., 2014, Settele et al., 2021, Cockayne, 2011)
Topological quantum systems	Local-perturbation-induced twist defects, defect-classifying symbols $\mathcal{H}(\mathbf{k},\mathbf{r})$	(Kairys et al., 11 May 2026, Goft et al., 2023)
Industrial vision and NDT	Procedural images, pasted defects, 3D rendered defects, Voronoi mesh defects	(Krassnig et al., 6 Mar 2025, Mandati et al., 2024, Jeziorski et al., 5 Feb 2026)
Generative augmentation	GAN-based, Cut&Paste-based, MLLM-based defect synthesis	(Zhang et al., 2021, Wang et al., 2023, Mou et al., 2024, Wang et al., 9 Mar 2026)
Restoration and software quality	Synthetic greening defects, synthetic defective logs	(Sinha et al., 28 May 2025, Zhong et al., 2024)

A recurring distinction is between defect synthesis for functionality and defect synthesis for data. The first aims at optical, electronic, or topological behavior; the second aims at better classifiers, segmenters, restorers, or recommenders. This suggests that “synthetic defect” is best understood as a design paradigm rather than a single methodology.

2. Bottom-up engineering of physical defects

In synthetic diamond, the most explicit demonstration of bottom-up defect control is the preferential orientation of nitrogen-vacancy defects during microwave-plasma CVD growth on a $(111)$-oriented substrate. The reported process used single-crystal diamond plates cleaved and polished from HPHT octahedral synthetic crystals, a $2\!-\!5^\circ$ offcut along $[111]$, $250$ mbar total pressure, $3500$ W microwave power, $T\approx 1000 \pm 30^\circ\mathrm{C}$, and $1.5\%$ $sp^3$0 in $sp^3$1. An $sp^3$2 epilayer was grown in $sp^3$3 h, corresponding to $sp^3$4, and the spontaneously generated NV defects were oriented with a $sp^3$5 probability along the $sp^3$6 axis (Lesik et al., 2014). The mechanistic account is step-flow growth on the hydrogen-rich $sp^3$7 surface: terrace nucleation is slow, incorporation at the step edge is favored, and the N–V pair is built in along the surface normal. In the rate form reported there,

$sp^3$8

with $sp^3$9. ESR splitting and polarization-dependent photoluminescence were used jointly to identify orientation, and saturated PL count rate was $\mathcal{H}(\mathbf{k},\mathbf{r})$0 higher for $\mathcal{H}(\mathbf{k},\mathbf{r})$1 NVs (Lesik et al., 2014).

In polymer-wrapped $\mathcal{H}(\mathbf{k},\mathbf{r})$2 single-walled carbon nanotubes, synthetic control targets the binding configuration of luminescent $\mathcal{H}(\mathbf{k},\mathbf{r})$3-defects rather than their mere presence. The reported protocol used CoMoCAT nanotubes fractionated by PFO-BPy in toluene, 2-haloanilines such as 2-iodoaniline at $\mathcal{H}(\mathbf{k},\mathbf{r})$4 mM, KO$\mathcal{H}(\mathbf{k},\mathbf{r})$5Bu at $\mathcal{H}(\mathbf{k},\mathbf{r})$6 eq., and a toluene/DMSO/THF solvent mixture of $\mathcal{H}(\mathbf{k},\mathbf{r})$7 vol\%. Under dark conditions, the reaction yielded exclusively deep traps $\mathcal{H}(\mathbf{k},\mathbf{r})$8, whereas UV illumination at $\mathcal{H}(\mathbf{k},\mathbf{r})$9 nm produced a mixture of $(111)$0 and $(111)$1. For $(111)$2 nanotubes, $(111)$3 is reported at $(111)$4 nm with $(111)$5 meV and $(111)$6 at $(111)$7 nm with $(111)$8 meV; the deeper trap has longer photoluminescence lifetime, with $(111)$9 ps, and the maximum total PLQY is $2\!-\!5^\circ$0. Single-photon emission at room temperature was observed with $2\!-\!5^\circ$1 (Settele et al., 2021). The proposed reaction mechanism is nucleophilic addition, and the central lesson is that synthetic control can be exerted at the level of binding motif, with direct spectroscopic consequences.

Graphene provides a different notion of synthesis: finite-area topological defects are constructed by replacing a simply connected honeycomb patch $2\!-\!5^\circ$2 by a patch $2\!-\!5^\circ$3 with the same number of dangling bonds. In the dual triangular lattice, this becomes a cut-and-graft operation on a triangulated patch with equal perimeter, after which the modified dual graph is mapped back to graphene (Cockayne, 2011). The identity

$2\!-\!5^\circ$4

constrains admissible grafts, and a prominent subset consists of defects bounded by alternating 5- and 7-membered rings. The paper gives explicit examples including the isolated divacancy, the $2\!-\!5^\circ$5 and $2\!-\!5^\circ$6 reconstructions, divacancy clusters, and the flower defect, with quoted formation energies such as $2\!-\!5^\circ$7 eV for the isolated divacancy and $2\!-\!5^\circ$8 eV for the flower defect (Cockayne, 2011). Here, “synthetic defect” design is algorithmic and crystallographic rather than statistical.

3. Topological formulations and emergent synthetic defects

A broader topological formalism appears in the graphene classification based on the Weyl symbol $2\!-\!5^\circ$9, which extends the Bloch Hamiltonian to systems with spatial defects or textures (Goft et al., 2023). In that framework, the defect is assigned a dimension $[111]$0, the momentum space contributes spatial dimension $[111]$1, and the total homotopy space is $[111]$2. The paper states that a stable $[111]$3-topology occurs when

$[111]$4

with $[111]$5 and $[111]$6. Vacancy and Kekulé defects in graphene are treated as nontrivial cases with protected zero-mode physics, whereas a local adatom mass alone is not assigned a $[111]$7 invariant in the same way (Goft et al., 2023). This formulation shifts the discussion from local ring statistics to defect topology encoded in $[111]$8.

In the surface-code setting, synthetic defects are not lattice dislocations but effective twists induced by a local perturbation on a cut $[111]$9. The unperturbed Wen-plaquette Hamiltonian is supplemented by a transverse field $250$0, yielding

$250$1

The construction admits both a spin-language description with emergent virtual symmetries and a Majorana-language mapping to the open Kitaev chain (Kairys et al., 11 May 2026). For a cut of length $250$2, exact diagonalization shows a minimum in the spectral gap near $250$3, and the low-energy splitting closes exponentially with cut length, consistent with edge-mode splitting in a Kitaev chain. In the synthetic-defect phase, the two edge-localized Majorana zero modes constitute the twist defects and encode one non-local qubit (Kairys et al., 11 May 2026). The paper explicitly emphasizes that the defects emerge under local perturbation rather than as pre-imposed static features of the lattice.

These two topological lines of work address different levels of description. The graphene-symbol formalism classifies defect textures by symmetry and dimension (Goft et al., 2023), whereas the surface-code study derives an emergent defect phase from a concrete Hamiltonian deformation and follows its spectral properties numerically (Kairys et al., 11 May 2026). Together they show that synthetic defects can be either constructed objects in the lattice graph or emergent excitations generated by controlled perturbations.

4. Procedural and physically grounded defect synthesis for data generation

In industrial inspection, synthetic defects often begin with explicit procedural models. A survey focused on display front-of-screen inspection organizes the field into data augmentation, rule-based or procedural insertion, physically based simulation or rendering, and deep generative models (Mou et al., 2022). That taxonomy is borne out by later pipelines.

ISP-AD uses the defect-synthesis engine of Haselmann and Gruber (2017), where a defect contour is drawn by a random walk with momentum and noise,

$250$4

and similarly in $250$5. Binary masks can then be dilated, disks can be sampled with $250$6, and intensity is added by

$250$7

The reported patch size is $250$8 px, $250$9 of patches contain a synthetic defect, the random-walk length is uniform in $3500$0, $3500$1 is uniform in $3500$2, and $3500$3 is uniform in $3500$4 gray levels (Krassnig et al., 6 Mar 2025). This is a canonical rule-based generator: explicit, parameterized, and tunable.

Other pipelines move from 2D masks to scene- or geometry-level synthesis. In utility-asset inspection, defective cross-arms are modeled in Maya, coordinated through Omniverse Nucleus, procedurally assembled in Houdini, and rendered via Omniverse or Unreal Engine, with 10,000 images generated for each of the three defect classes split, break, and decay, each accompanied by per-pixel semantic maps, bounding boxes, and instance-segmentation masks (Mandati et al., 2024). In cast-metal inspection, defect geometries are modeled as 3D meshes derived from 2D Voronoi tessellations; elongated defects are created by shortest paths on the Voronoi graph, scab-like defects by coarse-to-fine tessellation unions, and the resulting meshes are embedded in the object and rendered through Monte Carlo simulation of the testing modality. Pixel-perfect annotation is produced by rendering an additional ID pass in Blender (Jeziorski et al., 5 Feb 2026). These works explicitly couple synthetic defects to digital-twin or physics-based NDT workflows.

The photolithography case is even more tightly specified. Synthetic SEM line-space images of size $3500$5 are built from alternating photoresist and space stripes with sampled widths and pitch, and Bridge and Break defects are inserted as rectangular intrusions. The training corpus contains 10,000 images with $3500$6 train/validation/test split and approximately balanced classes, with nearly 10,425 Break and 10,390 Bridge defects in training (Shinde et al., 15 May 2025). Because the generator writes bounding boxes directly, annotation is autonomous. A different but equally explicit simulator appears in autochrome restoration, where synthetic greening defects are built from sampled point-shaped or large-region corruptions, radial intensity maps

$3500$7

per-channel corruption vectors $3500$8, Gaussian smoothing, and channel-wise intensity change

$3500$9

to create paired clean/corrupted training data (Sinha et al., 28 May 2025).

The common feature of these procedural systems is not photorealism alone but parameter transparency. Defect class, size, position, contrast, mesh depth, renderer, and annotation are all available as first-class controls. This suggests why such methods remain important even when GANs or MLLMs are available.

5. Generative, selection-based, and optimization-based learning with synthetic defects

Modern learning pipelines increasingly treat synthetic defect generation as a trainable or selectively filtered component rather than a fixed preprocessor. Defect-GAN is an unpaired image-to-image translation system with defacement and restoration processes, a compositional layer-wise architecture, spatial and categorical control via SPADE, adaptive noise injection, and explicit foreground-mask composition

$T\approx 1000 \pm 30^\circ\mathrm{C}$0

On CODEBRIM, the reported FID is $T\approx 1000 \pm 30^\circ\mathrm{C}$1, compared with $T\approx 1000 \pm 30^\circ\mathrm{C}$2 for CycleGAN and $T\approx 1000 \pm 30^\circ\mathrm{C}$3 for StarGAN+SPADE, and mixing 50,000 synthesized samples into training raises ResNet-34 classification accuracy from $T\approx 1000 \pm 30^\circ\mathrm{C}$4 to $T\approx 1000 \pm 30^\circ\mathrm{C}$5 (Zhang et al., 2021). DT-GAN extends StarGAN v2 by disentangling foreground defects from background products and defect shape from style. Its overall FID is reported as $T\approx 1000 \pm 30^\circ\mathrm{C}$6 versus $T\approx 1000 \pm 30^\circ\mathrm{C}$7 for StyleGAN2, and downstream classification error is reduced to $T\approx 1000 \pm 30^\circ\mathrm{C}$8, $T\approx 1000 \pm 30^\circ\mathrm{C}$9, and $1.5\%$0 on three products, corresponding to reductions of $1.5\%$1, $1.5\%$2, and $1.5\%$3 relative to the no-augmentation setting (Wang et al., 2023).

Other approaches optimize the synthetic process against downstream validation performance. Synth4Seg formulates defect synthesis and segmentation as a bi-level optimization problem, with learnable augmentation-source weights and learnable pasting-location heatmaps. The reported gains are up to $1.5\%$4 over random pasting when learning locations and up to $1.5\%$5 when learning source importance weights rather than assigning equal importance to all sources (Mou et al., 2024). In a related but non-adversarial direction, synth-dacl combines semi-synthetic balancing by pasting real defect polygons onto synthetic concrete textures with fully synthetic crack and cavity datasets. On the raw dacl10k test split, mIoU rises from $1.5\%$6 to $1.5\%$7, and on the perturbed test set it rises from $1.5\%$8 to $1.5\%$9; the paper also states a $sp^3$00 increase in mean IoU, F1 score, Recall, and Precision on perturbed test sets when all synthetic extensions are used (Flotzinger et al., 17 Jun 2025).

MLLM-based synthesis replaces dedicated image generators with prompt-conditioned multimodal systems. In ceramic insulator inspection, Gemini 3 Pro Image is used as a training-free generator with dual-reference conditioning, iterative prompt refinement, human verification, and embedding-based selection. Class centroids are computed from real training embeddings,

$sp^3$01

synthetic samples are ranked by Euclidean distance $sp^3$02, and the top $sp^3$03 per class are retained. With $sp^3$04 real training images, augmenting the set with $sp^3$05 selected synthetic images improves test F1 from $sp^3$06 to $sp^3$07, corresponding to an estimated $sp^3$08 data-efficiency gain (Wang et al., 9 Mar 2026).

A consistent empirical pattern is that selection and mixing matter as much as generation quality. ISP-AD reports that pure synthetic training on the LSM-1 modality gives MCC $sp^3$09, Recall $sp^3$10, and FPR $sp^3$11, whereas injecting only 10 real defects at $sp^3$12 yields MCC $sp^3$13, Recall $sp^3$14, and FPR $sp^3$15; using all 172 available real defects gives MCC $sp^3$16, Recall $sp^3$17, and FPR $sp^3$18 (Krassnig et al., 6 Mar 2025). This directly contradicts the common simplification that synthetic defects are either a complete substitute for real data or useful only in large volumes.

6. Extensions beyond visual inspection, and recurrent limitations

The synthetic-defect paradigm is not limited to images of physical surfaces. In software engineering, LogUpdater synthesizes four defect types in logging statements—Statement–Code Inconsistency, Static–Dynamic Inconsistency, Temporal Inconsistency, and Readability Issues—by heuristic mutation and GPT-4o-based semantic rewriting. The resulting dataset contains 20,000 synthetic negative examples and 20,000 positive examples, with an 80/10/10 split. A UniXcoder-based classifier trained on this synthetic corpus reaches Precision $sp^3$19, Recall $sp^3$20, and F1 $sp^3$21 on held-out synthetic data, and Precision $sp^3$22, Recall $sp^3$23, and F1 $sp^3$24 on 516 real log-centric commits. In the online repair pipeline, type-aware prompts then drive LLM-based log updates, and human evaluation on 100,000 fresh log sites finds $sp^3$25 correct detect-and-fix behavior (Zhong et al., 2024). Here the defect is textual and semantic rather than physical, but the synthetic-data logic is structurally identical.

Several limitations recur across domains. The utility-asset inspection study reports a $sp^3$26 MAP lift in its best synthetic-data experiment but notes remaining synthetic-real domain shift, a precision gap between healthy and defective classes, and the absence of statistical-significance tests or evidence on new geographies or sensors (Mandati et al., 2024). The MLLM insulator pipeline is evaluated on a single public UAV dataset with two defect classes, and cross-dataset or cross-component generalization remains to be tested (Wang et al., 9 Mar 2026). The autochrome restoration system attains strong synthetic-ground-truth metrics and removes real greening defects convincingly, but its training pairs are still purely synthetic, so real degradation factors beyond the modeled greening process remain outside the supervised distribution (Sinha et al., 28 May 2025). In synth-dacl, crack performance is influenced by annotation mismatch: the paper notes that synthetic fine cracks can generalize better to refined finecrack masks than to coarse polygon labels (Flotzinger et al., 17 Jun 2025).

A recurring misconception is that any synthetic defect generator is interchangeable with any other. The cited literature does not support that view. Rule-based random walks, Voronoi tessellations, physics-based rendering, cut-and-paste, GAN composition, bi-level augmentation learning, MLLM prompting, and chemically selective defect synthesis each encode different priors and different failure modes (Krassnig et al., 6 Mar 2025, Jeziorski et al., 5 Feb 2026, Zhang et al., 2021, Mou et al., 2024, Wang et al., 9 Mar 2026). Another misconception is that “synthetic” necessarily implies reduced physical relevance. In the diamond, nanotube, graphene, and surface-code studies, the defect is synthetic precisely because its orientation, binding configuration, graph embedding, or emergent zero mode is under experimental or theoretical control (Lesik et al., 2014, Settele et al., 2021, Cockayne, 2011, Kairys et al., 11 May 2026).

Taken together, these works define synthetic defects as a cross-disciplinary strategy for making defects controllable. In materials science, that control is exerted through growth kinetics, chemical selectivity, or lattice surgery. In topological matter, it is exerted through symbolic classification or local perturbations that induce new quasiparticle structure. In machine learning, it is exerted through procedural simulation, rendering, generative modeling, filtering, and validation-aware optimization. This suggests that the unifying scientific question is not whether a defect is “real” or “synthetic,” but which invariants of the defect distribution are preserved under synthesis and which downstream objectives those preserved invariants are sufficient to support.