Korg: Spectral Synthesis, Security & Music Tech

• Korg is a polysemous term used in research to denote distinct systems in astrophysics, formal methods, and embedded music technology.
• In astrophysics, Korg is a modern, open-source 1D LTE stellar spectral-synthesis package in Julia, optimized for speed, extensibility, and automatic differentiation.
• Beyond spectroscopy, KORG also refers to an automated attacker synthesis prototype in formal methods and a platform for digital audio effects in embedded music.

Korg is a name used in contemporary research for several distinct software and hardware contexts. In astrophysics, it denotes a modern, open-source 1D LTE stellar spectral-synthesis package written in Julia, developed for rapid generation and fitting of theoretical spectra and later extended from FGK to FGKM stars (Wheeler et al., 2022). In formal methods, KORG denotes a prototype for automated attacker synthesis for distributed protocols (Hippel et al., 2020). In embedded computational music, the name appears as the target platform for the “QubitCrusher” effect module for the Korg Nu:Tekt NTS-1 (Carney, 2023). Across these usages, the shared label refers not to a single unified system but to unrelated research artifacts operating in different technical domains.

1. Nomenclature and research domains

The cited literature uses the name in at least three technically distinct ways.

Usage of the name	Domain	Defining characterization
Korg	Stellar spectroscopy	A modern, open-source 1D LTE stellar spectral-synthesis package written in Julia (Wheeler et al., 2022)
KORG	Formal methods / security	A prototype tool for automated attacker synthesis for distributed protocols (Hippel et al., 2020)
Korg	Embedded music technology	The platform name in Korg Nu:Tekt NTS-1, which hosts the QubitCrusher logue-sdk effect module (Carney, 2023)

The astrophysical software is the most extensively documented of these uses in the cited corpus. It was introduced as a package for FGK stars that computes spectra from the near-ultraviolet to the near-infrared, supports both plane-parallel and spherical radiative transfer, and was designed explicitly for speed, extensibility, interoperability, and differentiability (Wheeler et al., 2022). A later update added built-in fitting functions, a validated model-atmosphere interpolation method, an extended chemical-equilibrium solver, and improved treatment of infrared Brackett lines, broadening its intended regime to FGKM stars (Wheeler et al., 2023).

2. Stellar spectral-synthesis package: aims and architecture

Korg was developed because existing stellar-synthesis codes, while scientifically valuable, were often harder to integrate programmatically, relied on custom file-I/O conventions, and were not designed for workflows such as automatic differentiation, repeated forward modeling, or large-scale survey analysis (Wheeler et al., 2022). Its design goal was therefore not only physical fidelity within the classical 1D LTE framework, but also software properties needed for modern inference pipelines.

In its original formulation, Korg assumes the stellar atmosphere is hydrostatic, one-dimensional, and in local thermodynamic equilibrium. The LTE source function is therefore Planckian,

$S_\lambda = B_\lambda(T),$

and elemental abundances are supplied on the standard logarithmic scale

$A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$

Given a model atmosphere, a line list, and a set of abundances, Korg solves the chemical equilibrium in each atmosphere layer, computes continuum and line opacities, and solves the radiative-transfer equation to obtain the emergent disk-integrated flux (Wheeler et al., 2022).

The transfer formalism is implemented for both supported geometries. In plane-parallel geometry, Korg uses

$$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$

while in spherical geometry it uses

$$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$

The returned quantity is the disk-averaged emergent intensity, or astrophysical flux,

$$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$

This architecture places Korg squarely within classical LTE stellar spectroscopy, but with an implementation optimized for repeated synthesis inside statistical and optimization loops (Wheeler et al., 2022).

3. Physical assumptions and numerical formulation

The original Korg paper recommends the code explicitly for FGK stars and identifies its principal limitations outside that regime (Wheeler et al., 2022). In the initial release, the equation of state includes all elements through uranium as neutral, singly ionized, and doubly ionized species, together with 247 diatomic molecules. Species more highly ionized than doubly ionized are neglected, as are ionic and triatomic/polyatomic molecules. This is one stated reason applicability outside FGK stars is limited, especially toward cool stars.

Continuum opacity includes a wide set of absorbers, including $\mathrm{H^-}$ bound-free and free-free, He$^-$ free-free, H I and He II bound-free and free-free, $\mathrm{H_2^+}$ bf/ff, metal free-free, several metal bound-free cross-sections, Rayleigh scattering by H, He, and $\mathrm{H_2}$, and electron scattering (Wheeler et al., 2022). A notable implementation choice is that all scattering is currently treated as absorption. The code paper identifies this as a known limitation, especially in blue/UV and hotter-star regimes.

For line formation, Korg uses LTE number densities and partition functions. For almost all non-hydrogen lines, it uses a Voigt profile. The Gaussian width is

$\sigma_D = \lambda_0 \sqrt{\frac{2kT}{m} + \xi^2},$

where $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$0 is the microturbulent velocity, and the Lorentz width in frequency units is

$A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$1

Hydrogen lines are treated separately using tabulated Stark-broadened hydrogen profiles from Stehlé & Hutcheon (1999), pre-convolved with Doppler broadening, with added self-broadening for $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$2, $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$3, and $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$4 (Wheeler et al., 2022).

The 2023 update revised several of these internals. Korg now solves for the electron density $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$5 self-consistently rather than assuming the model-atmosphere value, extends the chemical-equilibrium solver to include polyatomic molecules and positively charged molecules, and thereby extends applicability toward M stars (Wheeler et al., 2023). The added polyatomic set is based on ExoMol partition functions, and the added charged molecules use partition functions from Barklem & Collet (2016). The same update also applies the Mihalas–Hummer–Däppen (MHD) occupation-probability formalism to hydrogen bound-free opacity, which is especially relevant near the Balmer break (Wheeler et al., 2023).

4. Fitting, interpolation, differentiability, and infrared hydrogen lines

A major development after the initial release was the addition of built-in fitting interfaces (Wheeler et al., 2023). Korg now provides find_best_fit_params for fitting observed rectified spectra by direct synthesis and ews_to_abundances for inference from equivalent widths. Direct fitting uses the BFGS algorithm from Optim.jl, can operate on non-contiguous windows, and can fit parameters including $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$6, $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$7, metallicity, $A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$8-element abundance, individual elemental abundances, microturbulence, projected rotation, and a linear limb-darkening coefficient.

The EW-based solver is analogous in purpose to MOOG ABFIND, but its implementation differs. Rather than constructing a fictitious curve of growth, Korg synthesizes each line explicitly and infers the abundance from

$A(X) = \log_{10}(n_X/n_{\rm H}) + 12.$9

where $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$0 is the observed equivalent width and $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$1 is the synthetic equivalent width at the initial abundance $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$2 (Wheeler et al., 2023). This preserves local continuum and blend information that a precomputed curve of growth would not encode.

These fitting tools are tightly coupled to Korg’s compatibility with automatic differentiation. The update paper states that this can accelerate derivative-based calculations by an order of magnitude or more, which is important for optimization, uncertainty propagation, and gradient-based inference (Wheeler et al., 2023). The same paper emphasizes that fitting and synthesis are facilitated by a rigorously tested model-atmosphere interpolation procedure over the SDSS/MARCS grid of 579,150 model atmospheres, spanning $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$3, $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$4, $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$5, $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$6, and $$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$7. Its principal result is that interpolation error is negligible for stars with

$$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$8

while cooler regimes are complicated by molecular effects; for cool dwarfs a resampled cubic method reduces interpolation error to the subpercent level, whereas cool giants remain substantially more difficult (Wheeler et al., 2023).

The same update paper also identifies a common oversight in the treatment of infrared hydrogen lines. It argues that Brackett-line cores can be Stark-dominated because they form much deeper in the atmosphere than visible hydrogen cores. When broadening components are added rather than truly convolved, the wavelength-integrated cross section can be wrong by roughly a factor of

$$\frac{\mu}{\alpha_\lambda}\frac{d I_\lambda}{dz} = S_\lambda - I_\lambda,$$9

when Stark broadening dominates (Wheeler et al., 2023). Korg corrects this by using true numerical convolution of the broadening profiles and reports markedly improved Brackett-line agreement with observations.

5. Validation, benchmarks, and astronomical applications

The foundational validation of the stellar-synthesis package compares Korg against MOOG, Turbospectrum, and SME for four benchmark atmospheres and six spectral regions (Wheeler et al., 2022). The headline result is that Korg disagrees with the other codes no more than they disagree with one another, although overall code-to-code disagreement remains substantial, sometimes reaching the 10% flux level. Agreement is best in the infrared and worst in the blue and near-UV, especially near the Balmer jump and the Ca II K wing. The same paper’s most detailed discrepancy analysis concerns a solar C$$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$0 band at 5160–5165 Å, where differences among codes were traced largely to differing adopted molecular equilibrium constants and dissociation energies, rather than to numerical transport algorithms alone (Wheeler et al., 2022). On a single AMD Epyc 7702P core, Korg was reported to be 1–100× faster than the comparison codes in typical use, and for an APOGEE-like benchmark it could compute abundance gradients with a cost scaling of approximately

$$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$1

for an $$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$2-element gradient spectrum (Wheeler et al., 2022).

A separate update benchmark reanalyzed the 18 Sco equivalent-width dataset used by Meléndez et al. The mean differential abundances agreed with the past analysis, but the Fe line-to-line abundance scatter fell from

$$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$3

and most other elements also showed significantly smaller line-to-line scatter (Wheeler et al., 2023). This established Korg’s EW workflow as suitable for high-precision differential abundance work.

Korg has since been used as an independent fitting engine in several observational studies. In the GALAH DR4 cluster analysis, Korg was used to refit red-channel (CCD3) spectra in Melotte 22, Melotte 25, and NGC 2632, after degrading all spectra to

$$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$4

and masking the $$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$5 and Li regions. The study reported that “The iron abundance trend produced by Korg’s values is much flatter than any trend we get from SME’s values or in GALAH DR4,” and that a significant Korg iron decline appears only below

$$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$6

(Kos et al., 10 Jan 2025). The same paper also notes that some Korg fits did not converge and that $$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$7 could become unrealistically high in those failed cases.

In a study of low-metallicity subgiants, Korg was the abundance engine for high-resolution, high-S/N PEPSI spectra from the Large Binocular Telescope, using MARCS atmospheres and solar abundances from Asplund et al. (2021) (Griffith et al., 1 Dec 2025). The analysis produced homogeneous 1D-LTE abundances for 23 elements, including 11 heavy elements, and inferred intrinsic heavy-element scatter ranging from 0.11 dex (Zn) to 0.27 dex (Eu) after removing two strongly $$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$8-process-enhanced outliers.

Korg has also been applied to medium-resolution X-Shooter abundance work. In one halo-star analysis, it served as the main 1D LTE spectral synthesis tool for metallicities and abundances from 16 metal-poor stars, using fit_spectrum, adjust_continuum, ews_to_abundances, and synthesise, with MARCS atmospheres, VALD line data, and external post-hoc corrections where NLTE or 3D NLTE effects mattered (Lowe et al., 28 Oct 2025). In a solar-twin reanalysis, Korg was used in an equivalent-width, line-by-line differential framework with MARCS atmospheres for 79 nearby solar twins. That study reported typical uncertainties of 5.4 K in $$\frac{\mu}{\alpha_\lambda}\frac{\partial I_\lambda}{\partial r} + \frac{1-\mu^2}{\alpha_\lambda r}\frac{\partial I_\lambda}{\partial \mu} = S_\lambda - I_\lambda.$$9, 0.016 dex in $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$0, 0.01 km s$$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$1 in $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$2, 0.008 dex in $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$3, and an average abundance precision of 0.015 dex (3.5%) (Babatsikos et al., 2 Jul 2026).

Korg has also functioned as a reference code in external validation. The STARDIS introduction reports spectral comparisons against Korg with the same input atmospheric structures and finds agreement for solar models at the few percent level or better, with larger divergences in the ultraviolet and more extreme differences in cooler stars (Shields et al., 24 Apr 2025).

6. Other research systems and platform uses named Korg

In formal-methods research, KORG is an unrelated prototype for automated attacker synthesis for distributed protocols (Hippel et al., 2020). The name is stated not as an acronym but as a reference to the Korg microKORG synthesizer, which includes a dedicated “attack” control. KORG operates on a process-algebraic model in which a threat model is written as

$$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$4

where $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$5 is a target process, the $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$6 are vulnerable processes, and $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$7 is an LTL property such that the benign composition satisfies $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$8. It then synthesizes deterministic attacker processes $$\mathcal{F}_\lambda = 2\pi \int_0^1 \mu I_\lambda^{\rm top}(\mu)\, d\mu.$$9 that replace the $\mathrm{H^-}$0 and cause

$\mathrm{H^-}$1

The implemented algorithms solve the $\mathrm{H^-}$2ASP and R-$\mathrm{H^-}$3ASP problems by reducing attacker synthesis to model checking via the Daisy and RDaisy gadget constructions. The prototype is written in Python 3, uses Spin as the underlying model checker, and in the reported TCP case study could automatically generate well-known attacks within seconds or minutes (Hippel et al., 2020).

That formal-methods KORG was later extended for SCTP analysis. The SCTP paper modified Korg to support arbitrary finite packet types, to report attacks even if search-space exhaustion failed, to skip a redundant benign pre-check already discharged directly in Spin, and, most significantly, to support replay attackers with packet storage and replay (Ginesin et al., 2024). Using the extended tool, the authors synthesized 14 unique attacks: 1 Off-Path, 4 Evil-Server, 1 Replay, and 8 On-Path. They also used it to verify that the proposed RFC 9260 patch for CVE-2021-3772 eliminated the synthesized Off-Path vulnerability without introducing new attacks against their modeled properties (Ginesin et al., 2024).

In embedded computational music, the name Korg appears again in a different sense, as the platform manufacturer in Korg Nu:Tekt NTS-1. The paper “Quid Manumit -- Freeing the Qubit for Art” describes QubitCrusher, a Korg logue-sdk effect module that was compiled and loaded into the Korg Nu:Tekt NTS-1 (Carney, 2023). QubitCrusher implements the paper’s “quantum distortion” method by replacing user-defined bitcrusher cutoff parameters with values generated from a fixed two-qubit state model. The implementation uses a pre-computed statevector, updates it through the relevant $\mathrm{H^-}$4 and $\mathrm{H^-}$5 relations, commits the resulting values to the effect memory, and retains the paper’s measurement-sampling algorithm inside the effect chain (Carney, 2023). In that literature, “Korg” denotes the host hardware and SDK ecosystem rather than the quantum-processing algorithm itself.

Taken together, these usages show that “Korg” is not a single technical object but a polysemous label spanning stellar spectral synthesis, formal attack synthesis, and embedded music technology. Within astrophysics, however, Korg has developed into a substantial 1D-LTE analysis framework whose distinct contribution lies in combining classical stellar-spectroscopy assumptions with high speed, automatic differentiation, inference-oriented APIs, and a growing record of use in survey validation, precision abundance work, and external code benchmarking (Wheeler et al., 2022).