astronomy/ai produced the result/Astronomy and Astrophysics 2025 · v2
Neural network stands in for slow spectral modelling to read 64 stars' chemistry
Astronomers built LRPayne, a method for pulling temperatures, gravities and 24 elemental abundances out of low-resolution starlight. A neural network trained on 70,000 simulated spectra replaces the physics code inside the fitting loop.
spectrum · one line per step, placed by what the step does · bright lines used AI
LRPayne: Stellar parameters and abundances from low-resolution spectra
Astronomy and Astrophysics, 2025
doi:10.1051/0004-6361/202556502 · record aix-00191 v2 · checked 2026-10-09
- AI was for
- Property prediction, Simulation surrogate
- Model family
- Multilayer perceptron
- Checked by
- Held-out64 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
When starlight is spread into a spectrum, dark lines appear where atoms in the star's outer layers have absorbed particular colours. The pattern of those lines encodes the star's surface temperature, its surface gravity, how much iron it holds relative to hydrogen, and the amounts of other elements. Reading that code is done by simulation: astronomers compute what the spectrum of a star with assumed properties would look like, compare it with the real one, adjust the assumptions and try again. The catch is that each simulated spectrum is slow to compute, and here there are thirty properties being varied at once. Searching that many dimensions one simulation at a time is impractical.
The researchers set out to build a faster route to the same answers, working at low spectral resolution, where lines are blurred together and individual elements are harder to separate. They then checked the method against 64 real stars whose properties have already been measured by other means: 23 benchmark stars of roughly sun-like type and 41 stars poor in heavy elements.
Where AI came in
A fully connected neural network, a stack of five layers of simple numerical units, was trained on a library of 70,000 synthetic spectra computed with standard stellar-atmosphere physics. Given a set of 30 stellar properties, it predicts the brightness at each wavelength. Trained this way, it stands in for the simulation code: inside the fitting loop, the network produces candidate spectra and a conventional least-squares routine adjusts the properties until the candidate matches the observation. Every temperature, gravity and abundance the work reports comes out of that loop, so the results are the network's outputs, filtered through the fit.
The network was also used in the checking. On 5,000 synthetic spectra held back from training, the gap between its prediction and the full simulation had a median below 0.13 per cent for 90 per cent of the sample. To put error bars on the results, ten networks were trained on differently shuffled versions of the same library and refitted. Compared with published values, the fits differed on average by 22±87 K in temperature, 0.19±0.23 dex in surface gravity and 0.01±0.17 dex in metallicity, with larger gaps for oxygen, aluminium and manganese in metal-rich giants.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
LRPayne is a method for measuring stellar properties from low-resolution optical spectra. A fully connected neural network was trained on 70,000 synthetic spectra, computed with Turbospectrum and 1D MARCS atmospheres inside iSpec, to predict normalised flux at each wavelength given 30 stellar labels; the trained network then stands in for the synthesis code inside a chi-squared fit to an observed spectrum. On 5000 held-out synthetic spectra the median interpolation error was below 0.13 per cent for 90 per cent of the sample. Comparing fits of 23 Gaia FGK benchmark stars and 41 metal-poor stars with literature values gave mean differences of 22±87 K in effective temperature, 0.19±0.23 dex in surface gravity and 0.01±0.17 dex in metallicity, with larger discrepancies for oxygen, aluminium, manganese in metal-rich giants, and surface gravity in hot metal-poor dwarfs.
How AI was used
A library of 70,000 synthetic spectra was generated with the Turbospectrum code and 1D MARCS LTE model atmospheres within iSpec, over 30 randomly drawn labels (Teff, log g, [Fe/H], microturbulence, macroturbulence, v sin i and 24 elemental abundances), with Teff and log g pairs checked against MIST evolutionary tracks, then degraded to R=5000 and resampled to four pixels per Angstrom. This library was split 80-20 and used to train, in TensorFlow, a five-layer fully connected network with three 200-neuron hidden layers, LeakyReLU activations between the first four layers and a sigmoid before the output, fitted with ADAM at an initial learning rate of 0.001, a scheduler halving the rate after ten steps of unchanged loss, he-initialised hidden weights and a mean absolute percentage error loss. The trained network acts as a spectral interpolator: given trial labels it constructs a normalised spectrum, and Scipy's curve_fit performs chi-squared minimisation against an observed spectrum to recover the labels. Observed spectra from HARPS, NARVAL and ESPaDOnS were homogenised with an iSpec-based pipeline (range trimming, cross-correlation radial velocity correction, co-addition, segmented spline continuum normalisation, Gaussian convolution to R=5000, resampling). Pixels deviating by more than 0.03 between the observed solar spectrum and the network's solar construction, plus the Balmer lines and the data gap, were masked before fitting. Uncertainties were obtained by training ten networks on different shuffles of the training set and running 2000 fits with varied initial guesses per network for each of ten stars.
The shape of the work
Structural · the record, drawn
no AI
Synthesise training spectral library
Numerical or physics simulation, including where a learned surrogate replaces it.
We computed 70 000 synthetic spectra with a total of 30 varying parameterswhere the paper describes this · verbatim
AI
Train ANN spectral interpolator
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
we settled with a five-layer model where each of the three hidden layers has 200 neuronswhere the paper describes this · verbatim
no AI
Pre-process and homogenise observed spectra
Cleaning, filtering, normalising or labelling data already obtained.
we created a pipeline to perform the pre-processing. This pipeline heavily relies on functions defined within iSpec toolwhere the paper describes this · verbatim
no AI
Mask unreliable pixels
Cleaning, filtering, normalising or labelling data already obtained.
masked any pixel that has a deviation larger than 0.03where the paper describes this · verbatim
AI
Fit observed spectra to derive labels
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
perform a simple χ2 minimisation between the constructed spectrum and the observed spectrum to determine the best-fit stellar parameterswhere the paper describes this · verbatim
AI
Internal accuracy tests on held-out synthetic spectra
Testing outputs against ground truth.
An additional 5000 synthetic spectra were synthesised to be used for the two internal accuracy testswhere the paper describes this · verbatim
no AI
Compare derived labels with literature values
Testing outputs against ground truth.
we compare the results obtained from LRPayne with those obtained from literature sourceswhere the paper describes this · verbatim
AI
Estimate label uncertainties by repeated retraining and refitting
Testing outputs against ground truth.
We train 10 different ANN models using different shuffling of the training set but keeping all the hyper-parameters the samewhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The neural network spectral interpolator produces every stellar parameter and abundance the paper reports; the results are the model's outputs.
utilising a fully connected artificial neural network (ANN), trained on a library of 70,000 synthetic stellar spectrawhere the paper describes this · verbatim
Validation on 64 real stars (23 Gaia FGK benchmark stars and 41 metal-poor stars) reveals robust performancewhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of LRPayne ANN spectral interpolatorWhich version of the model was used is not stated.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
- What step 8 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00191, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error