Kurucz-a1: Dual Approaches in Stellar Atmospheres

• Kurucz-a1 represents two frameworks in stellar-atmosphere research: a physics-informed neural network for 1D LTE models and a photometric-hydrodynamic calibration for RR Lyrae stars.
• The PINN formulation employs a dual-encoder MLP architecture trained on ATLAS-12 data, yielding fast, differentiable predictions that adhere closely to hydrostatic equilibrium.
• The photometric-hydrodynamic approach derives phase-resolved stellar parameters from multicolour photometry using QSAA conditions and dynamical constraints to recover distance and mass.

Searching arXiv for papers mentioning “Kurucz-a1” and the cited 2025 PINN work. Kurucz-a1 is a label used in two related but technically distinct contexts in stellar-atmosphere research. In one usage, it denotes a physics-informed neural network (PINN) that emulates 1D stellar atmosphere models under Local Thermodynamic Equilibrium (LTE) and is designed as a differentiable atmospheric structure solver for stellar spectroscopy (Li et al., 8 Jul 2025). In another, earlier usage, it denotes a calibration based on static Kurucz model atmospheres that converts multicolour photometry of RR Lyrae stars into phase-resolved atmospheric and stellar parameters, including distance and mass, through a generalized quasi-static atmosphere approximation (QSAA) and hydrodynamic constraints (Barcza, 2010). The shared label reflects a common dependence on Kurucz atmospheric models, but the two constructions operate at different points in the methodological spectrum: one is a neural surrogate for atmospheric structure, the other a photometric-hydrodynamic inversion scheme.

1. Terminological scope and scientific setting

A common source of confusion is that “Kurucz-a1” does not denote a single historically fixed model. In the 2025 differentiable-spectroscopy literature, Kurucz-a1 is a dual-encoder PINN trained on ATLAS-12 atmospheres and regularized by hydrostatic equilibrium. In the earlier RR Lyrae literature summarized from Barcza’s work, the same label is used for a calibration procedure built from static Kurucz model atmospheres and broadband photometry. Both uses are centered on stellar atmospheres, effective temperature, gravity, composition, and radiative quantities, but they solve different inverse problems and rely on different numerical primitives (Li et al., 8 Jul 2025, Barcza, 2010).

The later Kurucz-a1 formulation addresses what is described as a critical bottleneck in differentiable stellar spectroscopy: the need for a fast, differentiable atmosphere model whose outputs remain physically consistent. The earlier calibration addresses a different bottleneck: the recovery of $T_{\rm e}$, $g_{\rm e}$, angular radius, reddening, metallicity, distance, mass, and atmospheric motion from $UBV(RI)_C$ photometry without radial-velocity data. A plausible implication is that the name now spans both a classical model-atmosphere calibration tradition and a modern differentiable-ML reimplementation of atmosphere modeling.

2. Neural surrogate formulation for 1D LTE atmospheres

In the 2025 usage, Kurucz-a1 is a physics-informed neural network that emulates 1D stellar atmosphere models under LTE. Its inputs are global stellar parameters—effective temperature $T_{\rm eff}$, surface gravity $\log g$, metallicity $[\mathrm{Fe}/\mathrm{H}]$, and $\alpha$-enhancement $[\alpha/\mathrm{Fe}]$—together with a local depth coordinate sampled at 80 points in Rosseland optical depth $\tau_{\rm Ross}$. The architecture is a dual-encoder design. A small MLP maps $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$ to $g_{\rm e}$0, and a second MLP maps each scalar $g_{\rm e}$1 to $g_{\rm e}$2 for $g_{\rm e}$3. The 512-dimensional stellar embedding is broadcast to all depths and concatenated with each 512-dimensional depth embedding, producing a 1024-dimensional representation per layer, which is then passed through a 3-layer MLP with widths $g_{\rm e}$4 and GeLU activations (Li et al., 8 Jul 2025).

The network predicts six depth-dependent outputs: column mass density $g_{\rm e}$5, temperature $g_{\rm e}$6, gas pressure $g_{\rm e}$7, electron number density $g_{\rm e}$8, Rosseland mean opacity $g_{\rm e}$9, and radiative acceleration ACCRAD. This design makes the atmospheric structure explicitly differentiable with respect to the global stellar parameters. The standard equation of state,

$UBV(RI)_C$0

is not imposed as a separate loss term; it is implicit in the ATLAS-12 training labels. Within the scope stated in the source, Kurucz-a1 is therefore a supervised emulator whose physical regularization is concentrated on hydrostatic equilibrium rather than a fully unconstrained generative atmosphere model.

3. Training objective, dataset, and numerical behavior

The training objective combines a data term and a physics term,

$UBV(RI)_C$1

with $UBV(RI)_C$2, chosen by grid search. The data loss is the mean-squared error over all six outputs and all depths. The physics loss enforces hydrostatic equilibrium in optical-depth coordinates through the relation

$UBV(RI)_C$3

with automatic differentiation used to compute $UBV(RI)_C$4. The training set consists of 104,269 atmospheric models from ATLAS-12, each sampled at 80 Rosseland optical depths and covering $UBV(RI)_C$5–50,000 K, $UBV(RI)_C$6–5.5, $UBV(RI)_C$7–$UBV(RI)_C$8, and $UBV(RI)_C$9–$T_{\rm eff}$0. The split is 90% train (93,842 models) and 10% validation (10,427 models) (Li et al., 8 Jul 2025).

Validation behavior is reported in several complementary ways. The median relative error across the validation set is $T_{\rm eff}$1 for temperature, $T_{\rm eff}$2 for pressure and density, and $T_{\rm eff}$3 for opacity; these values are stated to be within or below inter-model differences between ATLAS and MARCS. For hydrostatic equilibrium, Kurucz-a1’s $T_{\rm eff}$4 distribution is centered near zero and is nearly indistinguishable from ATLAS-12’s own discrete solution, with deviations of order $T_{\rm eff}$5–$T_{\rm eff}$6, whereas an MLP baseline without the physics loss shows order-of-magnitude larger violations. In a solar-spectrum test, RMS residuals to the observed solar spectrum are lower for Kurucz-a1 than for ATLAS-12 in many line wings, with overall synthetic flux residuals $T_{\rm eff}$7 in continuum regions and comparable residuals in line cores. Post-training inference is reported as $T_{\rm eff}$8 ms per model on Apple M1 Pro, corresponding to forward-evaluation speedups of $T_{\rm eff}$9–$\log g$0 relative to minutes-to-hours ATLAS-12 calls, while gradients with respect to stellar parameters are available in $\log g$1 ms rather than through finite-difference calculations taking minutes.

4. Differentiable spectroscopy pipeline and stated applications

Kurucz-a1 is intended to function inside an end-to-end differentiable spectral-analysis pipeline. The workflow described is: observed spectrum $\log g$2 differentiable radiative transfer, such as Korg or PySME in JAX or PyTorch, $\log g$3 residuals, with gradient flow back through the radiative-transfer module to atmospheric-structure outputs and then back through the Kurucz-a1 PINN to the global parameters $\log g$4. The source explicitly states that automatic differentiation computes both

$\log g$5

which enables direct gradient-based fitting of stellar labels and even atomic physics parameters (Li et al., 8 Jul 2025).

The applications proposed in the source are correspondingly broad but still tightly specified. They include data-driven calibration of atomic line parameters such as oscillator strengths by fitting large surveys end-to-end; global optimization of convection, opacity tables, and equation-of-state parameters using diverse stellar samples; extension to higher-dimensional abundance spaces, including CNO, Li, He, and individual elements, by expanding the global input dimension; and incorporation of additional physics constraints such as energy balance, non-LTE effects, and 3D corrections via augmented PINN losses. The same source also presents practical large-scale spectral analyses without sparse-grid interpolation as a target use case, with the stated benefit of eliminating associated systematic errors. Within that framework, Kurucz-a1 is best understood as a building block for fully differentiable, data-driven stellar spectroscopy rather than as an isolated atmospheric emulator.

5. Photometric-hydrodynamic “Kurucz-a1” calibration for RR Lyrae stars

In the earlier usage summarized from Barcza’s RR Lyrae work, Kurucz-a1 is a calibration of static Kurucz model atmospheres used to derive fundamental stellar parameters from multicolour photometry. Starting from $\log g$6 observations, one forms ten basic colour indices,

$\log g$7

Any pair $\log g$8 containing at least one $\log g$9 band constrains $[\mathrm{Fe}/\mathrm{H}]$0 via the Kurucz grid. For assumed metallicity $[\mathrm{Fe}/\mathrm{H}]$1 and interstellar reddening $[\mathrm{Fe}/\mathrm{H}]$2, each pair defines two model surfaces,

$[\mathrm{Fe}/\mathrm{H}]$3

whose intersection yields one solution $[\mathrm{Fe}/\mathrm{H}]$4 per pair. Because there are 30 such pairs, there are 30 estimates at each pulsation phase $[\mathrm{Fe}/\mathrm{H}]$5, which are averaged to obtain $[\mathrm{Fe}/\mathrm{H}]$6 and $[\mathrm{Fe}/\mathrm{H}]$7 (Barcza, 2010).

The photometric QSAA criterion is designated Condition I. It requires that the scatter $[\mathrm{Fe}/\mathrm{H}]$8 and $[\mathrm{Fe}/\mathrm{H}]$9 be consistent with photometric errors, specifically $\alpha$0 K in $\alpha$1 and $\alpha$2 dex in $\alpha$3; phases satisfying this criterion are described as “shock-free.” Once $\alpha$4 are known, the angular radius is obtained from model monochromatic emergent fluxes $\alpha$5 and the observed dereddened flux $\alpha$6 through

$\alpha$7

so that

$\alpha$8

The best values of $\alpha$9 and $[\alpha/\mathrm{Fe}]$0 are obtained by varying both and minimizing the phase-to-phase scatter of $[\alpha/\mathrm{Fe}]$1 and $[\alpha/\mathrm{Fe}]$2 over the “good” phases satisfying Condition I.

6. Generalized QSAA, dynamical constraints, and recovery of distance and mass

The same RR Lyrae formulation augments the static photometric inversion with a hydrodynamic treatment. The full radial momentum equation is given as

$[\alpha/\mathrm{Fe}]$3

where $[\alpha/\mathrm{Fe}]$4 and $[\alpha/\mathrm{Fe}]$5 are the total static-plus-dynamic pressure and density. Writing $[\alpha/\mathrm{Fe}]$6 and $[\alpha/\mathrm{Fe}]$7, the dynamical acceleration correction is defined as

$[\alpha/\mathrm{Fe}]$8

In the static QSAA limit, $[\alpha/\mathrm{Fe}]$9, so the net acceleration driving pulsation is

$\tau_{\rm Ross}$0

Using the mass-continuity equation together with the static exponential stratification

$\tau_{\rm Ross}$1

where $\tau_{\rm Ross}$2, $\tau_{\rm Ross}$3, and $\tau_{\rm Ross}$4, the method admits the analytic 1-D velocity profile

$\tau_{\rm Ross}$5

Condition II, the hydrodynamic QSAA criterion, requires at $\tau_{\rm Ross}$6 that $\tau_{\rm Ross}$7, equivalently $\tau_{\rm Ross}$8 to within the photometric error $\tau_{\rm Ross}$9 (Barcza, 2010).

Distance and stellar mass are then recovered by combining the photometric radius $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$0 with the Euler equation at the photosphere. Defining the instantaneous dynamical mass,

$$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$1

one notes that $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$2 is a known function of $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$3 at each phase because $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$4, $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$5, and $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$6 scale with $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$7. Enforcing $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$8 at two or more shock-free phases yields the distance $$\{T_{\rm eff},\log g,[\mathrm{Fe}/\mathrm{H}],[\alpha/\mathrm{Fe}]\}$$9, and substitution back into $g_{\rm e}$00 gives the stellar mass. The calibration itself is tied to the Kurucz (1997) model grid, described as a tabulation of emergent fluxes $g_{\rm e}$01 and bolometric corrections $g_{\rm e}$02 for $g_{\rm e}$03, $g_{\rm e}$04, and a range of $g_{\rm e}$05, with interpolation in $g_{\rm e}$06 at fixed $g_{\rm e}$07 and $g_{\rm e}$08 and photometric zero-points $g_{\rm e}$09 used to convert magnitudes to physical fluxes. In this usage, Kurucz-a1 is therefore a photometric-hydrodynamic calibration framework rather than a neural atmospheric solver.

7. Relationship between the two usages

The two meanings of Kurucz-a1 are methodologically different but conceptually aligned. Both are built around atmospheric structure as an intermediate representation linking observables to stellar parameters. In the RR Lyrae calibration, the intermediate variables are phase-resolved $g_{\rm e}$10 inferred from multicolour photometry and filtered by photometric and hydrodynamic QSAA conditions. In the PINN formulation, the intermediate variables are depth-dependent fields such as $g_{\rm e}$11, $g_{\rm e}$12, $g_{\rm e}$13, $g_{\rm e}$14, and $g_{\rm e}$15 inferred from global stellar labels and constrained by hydrostatic equilibrium (Li et al., 8 Jul 2025, Barcza, 2010).

A second important distinction concerns the physical regime explicitly represented. The PINN emulates 1D LTE stellar atmosphere models and states future incorporation of energy balance, non-LTE effects, and 3D corrections as extensions rather than present features. The RR Lyrae calibration, by contrast, is grounded in static Kurucz atmospheres but supplements them with a generalized QSAA and explicit momentum and continuity equations to recover atmospheric motion, distance, and mass. This suggests that the common label should be interpreted contextually: in modern differentiable spectroscopy it refers to a fast physics-informed emulator of ATLAS-12-like atmospheres, whereas in RR Lyrae photometric analysis it refers to a calibration pipeline derived from static Kurucz atmosphere grids.