Clad: Multidisciplinary Perspectives

• Clad is a polysemous term that denotes engineered boundaries in optics (as cladding for light confinement), beads in additive manufacturing, safety barriers in nuclear reactors, and even a computational tool in automatic differentiation.
• In optics, cladding is engineered to enhance light guidance and pump efficiency, illustrated by designs achieving modal isolation above 60 dB and minimal excess loss over broad wavelength ranges.
• In computing, 'Clad' refers to a Clang-based automatic differentiation tool that generates efficient derivative code for C/C++ and CUDA, significantly accelerating gradient-based computations in scientific applications.

“Clad” is a polysemous technical term whose meaning depends strongly on disciplinary context. In optics, it most often denotes cladding: the lower-index region surrounding a guided core, with roles that range from total-internal-reflection confinement to pump guidance, mode selection, broadband signal collection, and thermal management (Becerra-Deana et al., 2024). In directed energy deposition, a clad is a single deposited bead whose geometry and dilution govern bonding, defect formation, and downstream part quality (Tayebati et al., 2023). In pressurized water reactors, Zircaloy clad is the first safety barrier around the fuel pellets and a central object of corrosion and chemo-mechanical analysis (Minne et al., 2013). In computational science, “Clad” also names a compiler-assisted automatic differentiation system for C/C++ and CUDA, while uppercase variants such as CLAD and CLaD are reused as acronyms for unrelated benchmarks and models in machine learning, robotics, and security (Ifrim et al., 2022).

1. Terminological scope and disciplinary distinctions

In optical-fiber usage, cladding is a structural and refractive-index layer. The standard definition in the cited photonics literature is explicit: the cladding is the lower-index glass surrounding a higher-index core that guides light by total internal reflection, and more elaborate profiles such as double-clad fibers use two concentric claddings with $$n_{\text{core}} > n_{\text{clad}1} > n_{\text{clad}2}$$ to control guidance across multiple boundaries (Becerra-Deana et al., 2024).

In metal additive manufacturing, the same word shifts from a surrounding layer to a deposited object. In directed energy deposition, a clad is a single track formed when a laser creates a melt pool and injected powder or wire solidifies into a bead. Its width $w$, height $h_c$, and penetration depth $d_p$ are treated as primary process outputs, and its quality is operationalized through dilution rather than through optical confinement (Tayebati et al., 2023).

In nuclear engineering, “clad” refers to fuel-rod cladding. Zircaloy cladding surrounds the $\mathrm{UO_2}$ pellets, retains fission products, and develops distinct inner and outer corrosion morphologies under very different thermo-mechanical and chemical conditions (Minne et al., 2013).

A separate line of usage is nominal rather than material. In software and AI papers, “Clad,” “CLAD,” or “CLaD” may be an acronym unrelated to layered media. This is true both for the Clang-based automatic differentiation tool “Clad” and for a growing family of ML systems such as CLAD for continual learning, audio deepfake detection, compressed-domain log anomaly detection, federated intrusion detection, or cross-modal robotic planning (Ifrim et al., 2022).

A common misconception is that “clad” always denotes a surrounding sheath. The literature does not support that restriction: in manufacturing it denotes the deposited bead itself, and in AI it is often only an acronym.

2. Optical cladding as an engineered functional region

In photonics, cladding is rarely a passive boundary. It is a controllable part of the mode structure, pump geometry, and device robustness. The standard formulas recur across the cited work: the numerical aperture is $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$, and step-index modal behavior is often parameterized through the $V$-number $V = (2\pi a/\lambda)\,NA$ (Becerra-Deana et al., 2024).

Mode-selective photonic lanterns built from double-clad fibers show this explicitly. In the reported design, the cores remain single-mode and SMF-28-compatible, while the first cladding becomes the guiding region during tapering. Mode selectivity is obtained by using three custom-pulled double-clad fibers with first-cladding diameters of $19.6\,\mu\text{m}$, $26.8\,\mu\text{m}$, and $w$0. With a fluoride-doped capillary, the resulting lanterns exhibit modal isolation above $w$1 and excess loss lower than $w$2 over more than $w$3; with synthetic fused silica capillaries, the same architecture still reaches modal isolation above $w$4 and excess loss lower than $w$5 over the same broad range (Becerra-Deana et al., 2024). The cladding here is both an optical and a fabrication variable: it mediates symmetry breaking, permits steeper adiabatic tapers, and improves mechanical robustness.

A different use appears in double-clad antiresonant hollow-core fibers for multiphoton micro-endoscopy. There, the inner hollow core delivers ultrashort pulses, while a second large-area collection cladding bounded by a low-index polymer supports total-internal-reflection guidance of backscattered nonlinear signals. The reported device has a $w$6 core, a $w$7 inner cladding, a $w$8 outer cladding, and a collection-cladding $w$9. Core loss remains below $h_c$0 from $h_c$1 to $h_c$2, pulse-width increase at $h_c$3 is below $h_c$4 per meter, and only the double-clad antiresonant fiber supports de-scanned detection in the compared endoscopy configurations (Szwaj et al., 2023).

Core/clad refractive-index engineering also remains central in mid-infrared fibers. A Pr$h_c$5-doped Ge–Ga–S step-index fiber was fabricated with $h_c$6, $h_c$7, and $h_c$8, giving $h_c$9. The resulting $d_p$0-core fiber emitted from $d_p$1 to $d_p$2 under $d_p$3 pumping (Carcreff et al., 2022). In a related theoretical study, As–Se–Te/As–S core/clad step-index fibers with strong index contrast were predicted to support supercontinuum extension from about $d_p$4 to beyond $d_p$5 with a $d_p$6 core and pump energy of order $d_p$7 from a $d_p$8 femtosecond all-fiber source (Anashkina et al., 2017).

Pump-guiding claddings are equally consequential in rare-earth fiber lasers. A holmium-doped triple-clad fiber reduced the pump-guiding inner cladding from $d_p$9 in a previous generation to $\mathrm{UO_2}$0, with an outer fluorinated-silica cladding diameter of $\mathrm{UO_2}$1 and $\mathrm{UO_2}$2. In a laser oscillator at $\mathrm{UO_2}$3, this geometry yielded a slope efficiency of $\mathrm{UO_2}$4 with respect to absorbed pump power, $\mathrm{UO_2}$5 with respect to coupled pump power, and $\mathrm{UO_2}$6 output at $\mathrm{UO_2}$7 (Beaumont et al., 2022). At the opposite end of the architecture space, a novel single-clad Ho-doped fiber with $\mathrm{UO_2}$8 core/cladding and $\mathrm{UO_2}$9 reached $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$0 optical-optical slope efficiency at about $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$1 under $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$2 core pumping, linking single-clad core pumping to nearly unity pump-core overlap and short absorption lengths (Tench et al., 1 Apr 2026).

Metal-clad resonators extend the same concept to integrated photonics. In InP nanodisk whispering-gallery lasers bonded to Si, Au sidewall cladding enabled room-temperature lasing down to $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$3 diameter, whereas purely photonic counterparts lased only down to $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$4. Thermal simulations indicated an improved heat-sinking capability corresponding to a reduction in device temperature of $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$5 for the metal-clad nanodisk compared with the non-Au device (Tiwari et al., 2020). In this setting, “clad” refers not to a low-index dielectric but to a metal boundary used simultaneously for confinement and heat spreading.

A plausible implication is that optical cladding is best understood not as a single material category but as an interface design space: refractive, geometric, modal, thermal, and fabrication constraints are all encoded there.

3. Clad as the elementary deposited bead in metal additive manufacturing

In directed energy deposition, a clad is the elementary volumetric unit from which multi-track walls and layered structures are assembled. The reported framework parameterizes single-track clad geometry by three cross-sectional features: width $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$6, height $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$7, and penetration depth $NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$8. Clad quality is assessed through dilution,

$NA = \sqrt{n_{\text{core}}^2 - n_{\text{clad}}^2}$9

with $V$0 treated as acceptable, lower values associated with lack of fusion and balling, and higher values associated with keyhole porosity and excessive dilution (Tayebati et al., 2023).

The cited study addresses prediction of clad geometry and quality by combining experiments with calibrated multiphysics CFD. Its dataset contains $V$1 samples: $V$2 experimental single clads in 316L stainless steel and $V$3 CFD-generated clads spanning a wider processing window. Two feature sets are used: machine-setting parameters $V$4 and physics-aware parameters, including volumetric energy density

$V$5

and linear mass density $V$6. Regression targets are $V$7, $V$8, and $V$9; classification labels are “desirable” and “undesirable” clads derived from the dilution criterion (Tayebati et al., 2023).

The best reported regressors are Gradient Boosting Regression models using the full four-feature set. Their $V = (2\pi a/\lambda)\,NA$0 and MAE values are $V = (2\pi a/\lambda)\,NA$1 and $V = (2\pi a/\lambda)\,NA$2 for width, $V = (2\pi a/\lambda)\,NA$3 and $V = (2\pi a/\lambda)\,NA$4 for height, and $V = (2\pi a/\lambda)\,NA$5 and $V = (2\pi a/\lambda)\,NA$6 for penetration depth. For quality classification, a neural network reaches accuracy $V = (2\pi a/\lambda)\,NA$7 and AUC $V = (2\pi a/\lambda)\,NA$8, outperforming KNN and logistic regression on the same desirable-versus-undesirable task (Tayebati et al., 2023).

The paper’s formulation makes clear that “clad characteristics” are not merely geometric descriptors. They are the process-facing observables through which energy input, powder supply, melt-pool hydrodynamics, and defect formation become controllable. This suggests that in manufacturing the term “clad” functions as a bridge concept between process physics and quality assurance rather than as a static material layer.

4. Nuclear clad as a safety barrier and corrosion system

In pressurized water reactors, Zircaloy clad is the first safety barrier of the fuel rod. It surrounds the $V = (2\pi a/\lambda)\,NA$9 or MOX pellets, contains radioactive fission products, and must simultaneously withstand irradiation, thermal gradients, contact stresses, and both internal and external oxidation environments (Minne et al., 2013).

The cited analysis focuses on internal corrosion after pellet–clad gap closure. At inner-surface temperatures around $19.6\,\mu\text{m}$0, a dense oxidized layer of $19.6\,\mu\text{m}$1 to $19.6\,\mu\text{m}$2 forms on the inner surface. Unlike waterside corrosion, this internal oxide remains fully tetragonal and dense; the paper attributes its stabilization to high stresses, stated as greater than about $19.6\,\mu\text{m}$3, together with irradiation and fission-induced defects. Growth is described in three stages: a rapid initial stage over about the first $19.6\,\mu\text{m}$4, a slower diffusion-limited stage up to about $19.6\,\mu\text{m}$5–$19.6\,\mu\text{m}$6, and eventual saturation rarely exceeding $19.6\,\mu\text{m}$7. The diffusion-limited regime is represented by

$19.6\,\mu\text{m}$8

with $19.6\,\mu\text{m}$9 at about $26.8\,\mu\text{m}$0 (Minne et al., 2013).

By contrast, outer or waterside corrosion develops a protective pre-transition black oxide, then transitions into a cracked and porous layer beyond about $26.8\,\mu\text{m}$1–$26.8\,\mu\text{m}$2. The paper associates this with a tetragonal-to-monoclinic transformation, pseudo-linear post-transition kinetics, and the large Pilling–Bedworth ratio of about $26.8\,\mu\text{m}$3, which drives volume-change stresses and cracking (Minne et al., 2013).

The most technically distinctive part of the study is the explicit stress–diffusion coupling in oxygen transport through $26.8\,\mu\text{m}$4-Zr(O). In Larché–Cahn form,

$26.8\,\mu\text{m}$5

and in the authors’ constitutive model the stress dependence of diffusivity is written

$26.8\,\mu\text{m}$6

Finite-element simulations on a 2D axisymmetric pellet/oxide/clad geometry indicate that increasing compressive stress near the inner surface lowers effective oxygen diffusivity deeper in the metal while reshaping the near-surface oxygen profile. The simulations report bulk compressive stresses reaching the gigapascal range, with values around $26.8\,\mu\text{m}$7 for thicker oxides, and use an oxygen solubility limit at the metal/oxide interface of $26.8\,\mu\text{m}$8 (Minne et al., 2013).

In this literature, “clad” is therefore not only a containment shell. It is a coupled chemo-mechanical system whose oxide phase stability, diffusion kinetics, and load-transfer behavior directly influence pellet–clad mechanical interaction and fuel performance margins.

5. Clad as compiler-assisted automatic differentiation infrastructure

In computing, “Clad” names a source-transformation automatic differentiation tool implemented as a plugin to the Clang compiler. It operates on the compiler’s abstract syntax tree, generates derivative code as ordinary C++, supports forward and reverse accumulation modes, integrates with Cling and xeus-cling, and is available in ROOT (Ifrim et al., 2022).

This design is especially relevant for high-energy physics workflows, where statistical models are often written in highly abstract C++ frameworks. For RooFit and HistFactory, the key difficulty is that the original object graph contains virtual function calls, stateful nodes, and framework-specific overhead that obstruct direct differentiation. The reported solution is to generate a single stateless C++ function for the full RooFit model, compile it with Cling, and then differentiate that generated function with Clad. The resulting derivative code is free of virtual dispatch and amenable to whole-function compiler optimization (Singh et al., 2023).

For binned likelihoods with observed counts $26.8\,\mu\text{m}$9 and expectations $w$00,

$w$01

and the gradient is

$w$02

This is the regime where reverse-mode AD is most attractive: scalar outputs, many parameters, and repeated calls from minimizers. On a HistFactory template with about $w$03 free parameters, the code-generation-plus-Clad strategy reports a $w$04 speedup over RooFit’s standard numerical differentiation (Singh et al., 2023).

The CUDA extension pushes the same toolchain into heterogeneous execution. Clad can preserve CUDA host/device attributes in differentiated functions, allowing automatically generated gradients of well-behaved C++ functions to execute on the GPU. The cited GPU study reports performance gains from offloading, including about $w$05 in ROOT histogram fitting, and emphasizes that AD combines machine-precision derivatives with a low constant-factor overhead relative to the original function (Ifrim et al., 2022).

Here the term “Clad” has no connection to material cladding. It denotes compiler-based differential analysis, and its significance lies in reducing the cost and fragility of derivative computation in large C++ scientific codes.

6. CLAD and CLaD as acronymic research systems

Uppercase or mixed-case forms such as CLAD, CLAd, and CLaD are frequently acronyms rather than references to cladding. Their meaning is field-specific.

Acronym and domain	Expansion or function	arXiv id
CLAD in autonomous driving	Continual Learning benchmark for Autonomous Driving	(Verwimp et al., 2022)
CLAD in audio forensics	Contrastive Learning-based Audio deepfake Detector	(Wu et al., 2024)
CLAD in instructional video planning	Constrained Latent Action Diffusion	(Shi et al., 9 Mar 2025)
CLAD in systems security	Log anomaly detection directly on compressed representations	(Tang et al., 14 Apr 2026)
CLAD in visual robustness	Contrastive Learning based Approach for Background Debiasing	(Wang et al., 2022)
CLAD in federated intrusion detection	Clustered Label-Agnostic Federated Learning for joint anomaly detection and attack classification	(Ofeidis et al., 7 May 2026)
CLaD in robotics	Planning with Grounded Foresight via Cross-Modal Latent Dynamics	(Jeong et al., 31 Mar 2026)
CLAD in diffusion LLMs	Cluster-Level Attention-Guided Decoding	(Qi et al., 28 May 2026)
CLAd-VR in immersive training	Cognitive Load-based Adaptive Training for machining tasks in virtual reality	(Matam et al., 6 Oct 2025)

These acronymic uses are technically unrelated to optical, metallurgical, or nuclear cladding. They range from benchmark design and representation learning to anomaly detection, robotic planning, and adaptive VR training. The term must therefore be resolved from local expansion, capitalization, and disciplinary venue rather than from the word alone.

Across these literatures, “clad” functions less as a single concept than as a family of domain-specific terms. In optics and materials, it often names an engineered boundary or deposited layer whose geometry and composition determine transport, confinement, or integrity. In software and AI, it increasingly serves as a compact acronym. The shared lesson is lexical rather than ontological: accurate interpretation of “clad” requires the technical frame in which it appears.