astronomy/ai produced the result/Monthly Notices of the Royal Astronomical Society 2026 · v2
Machine learning predicts galaxy 'relicness' from survey measurements, cutting spectroscopy demand
Astronomers trained regression models to reproduce the degree of relicness of ultra-compact massive galaxies from ordinary survey measurements. The models stood in for a slower analysis of detailed starlight spectra, and flagged candidate relics for follow-up.
spectrum · one line per step, placed by what the step does · bright lines used AI
E-INSPIRE – II. Finding relics from wide-sky multiband surveys: a proof-of-concept machine learning regression algorithm
Monthly Notices of the Royal Astronomical Society, 2026
doi:10.1093/mnras/stag1718 · record aix-00271 v2 · checked 2026-10-20
- AI was for
- Property prediction, Classification
- Model family
- Support vector machine, Random forest, Gradient-boosted trees, Linear model
- Checked by
- Held-out52 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Most galaxies grow over billions of years by merging with others and forming fresh stars. A few appear to have skipped that history. They formed their stars quickly, very early, and then stopped, surviving to the present almost untouched. Astronomers call these relics, and study them as surviving examples of how the first massive galaxies were assembled. The candidates are ultra-compact massive galaxies: objects that pack a large mass of stars into a small radius. Deciding how relic-like one of them really is normally requires a spectrum, meaning its starlight split into its component wavelengths, taken with enough signal to reconstruct when its stars formed. That is expensive telescope time, and it limits how many galaxies can be checked.
The degree of relicness is a single number summarising how much of a galaxy's stellar mass formed in that early, rapid burst. The team asked whether that number could instead be estimated from quantities already measurable in wide surveys of the sky, such as a galaxy's age, its chemical composition, how fast its stars move and its size, rather than from a dedicated high-quality spectrum.
Where AI came in
The researchers set the problem up as supervised regression: given a set of measured galaxy properties, predict a number. They first re-derived ages, chemical compositions and star formation histories for 430 galaxies in the E-INSPIRE catalogue using conventional spectral fitting software, which supplied both the input features and the target values the models were asked to reproduce. Four families of standard algorithms (ridge regression, random forest, XGBoost and support vector regression) were trained from scratch on those galaxies, across eight different combinations of input features. Support vector regression scored highest and was carried through the rest of the work.
The trained model was then applied to an independent set of 52 INSPIRE galaxies it had not seen, where its predictions were compared with the spectroscopically derived values. In effect the model stood in for the full spectral fitting of high-quality spectra, which the authors note it does not replace. Separately, classifiers were trained to sort galaxies as relic or not at various cut-off values, and the researchers measured which inputs mattered most by shuffling each one in turn. Redshift-normalised age came out most informative; stellar mass and size contributed almost nothing. Because predictions were squeezed towards middling values, a predicted cut at 0.6 was used to pick candidates for telescope follow-up.
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
The study trained regression models to reproduce the spectroscopically inferred degree of relicness (DoR) of ultra-compact massive galaxies from stellar population, kinematic and structural observables. Models were fitted on 430 E-INSPIRE galaxies and applied to an independent set of 52 INSPIRE galaxies; among the four algorithm families tested, Support Vector Regression gave the highest R2 on the test sample, where values spanned 0.783 to 0.824 across the eight feature configurations. Redshift-normalised age was the most informative feature in permutation-importance tests, with metallicity becoming more important at the highest DoR thresholds, while stellar mass and effective radius showed negligible importance. Predictions were compressed toward intermediate values, under-predicting DoR above 0.6, which the authors use to argue that a predicted DoR cut at 0.6 gives a conservative candidate selection.
How AI was used
Star formation histories, ages, metallicities and DoR values were re-derived for the E-INSPIRE sample with the pPXF full-spectral-fitting code using an unrestricted metallicity grid, giving a homogeneous parameter set. Eight input feature configurations were defined from these observables (redshift-normalised age, metallicity, [Mg/Fe], velocity dispersion, stellar mass, effective radius and the uncertainties on age and metallicity) and standardised to zero mean and unit variance with the scaler fitted on the training data only. Four regression families implemented in scikit-learn (Ridge Regression, Random Forest, XGBoost and Support Vector Regression with a degree-3 polynomial kernel) were fitted on the 430 E-INSPIRE galaxies with the spectroscopic DoR as target, using five-fold cross-validation and a grid search over C, gamma and epsilon; SVR was then carried through the rest of the analysis. The trained models predicted DoR for the 52 INSPIRE galaxies, scored by R2 and RMSE, with an additional test in which the two samples were pooled and randomly repartitioned. Separately, SVC classifiers were trained at binary DoR thresholds between 0.30 and 0.65 and permutation importance was computed to rank the features; a threshold on predicted DoR was used to define a candidate set for follow-up.
The shape of the work
Structural · the record, drawn
Assemble UCMG training and test catalogues
Obtaining raw data, whether by measurement, download or retrieval.
This dataset comprises a total of 430 galaxies in the redshift range 0.03<z<0.3where the paper describes this · verbatim
Re-derive star formation histories with unrestricted metallicity models
Cleaning, filtering, normalising or labelling data already obtained.
we re-derive all the SFHs for the E-INSPIRE I objects by running the code Penalised Pixel-fitting (pPXF;)where the paper describes this · verbatim
Define feature configurations and standardise inputs
Encoding data into features, descriptors, embeddings or graphs.
We therefore apply standardisation (zero mean, unit variance) to all input features, fitting the scaler on the training datawhere the paper describes this · verbatim
Fit and compare regression algorithms
Fitting model parameters, including fine-tuning an existing model.
Hyperparameters were optimised through a grid search over logarithmically spaced values of C, γ, and ϵwhere the paper describes this · verbatim
Predict DoR for the independent sample
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
We compare the predicted DoR from the SVR algorithm (y-axis) with the ’True DoR’where the paper describes this · verbatim
Evaluate against spectroscopic DoR
Testing outputs against ground truth.
Performance on the INSPIRE sample is reported as Rtest2.where the paper describes this · verbatim
Measure feature importance across DoR thresholds
Extracting understanding from model behaviour. The AI stood in for expert judgement.
we train an SVC classifier at a range of DoR thresholds (0.30 – 0.65) and measure permutation importance at each of themwhere the paper describes this · verbatim
Apply a predicted-DoR cut for candidate selection
Reducing a candidate set by filtering or ranking, in a single pass.
By selecting galaxies with predicted DoR ≥0.6, we obtain a conservative and high-purity sample of relic candidates.where 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 paper's result is the recoverability of the spectroscopic degree of relicness by regression; the regression output is the object of study, not an auxiliary analysis
we formulate its estimation as a supervised regression problemwhere the paper describes this · verbatim
When evaluated on an independent INSPIRE sample of 52 UCMGs, the predictive power remains robustwhere the paper describes this · verbatim
The E-INSPIRE catalogue is instead publicly available from the survey websitewhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of Support Vector Regression (polynomial kernel, degree 3)Which version of the model was used is not stated.
- Version of Ridge RegressionWhich version of the model was used is not stated.
- Version of Random ForestWhich version of the model was used is not stated.
- Version of XGBoostWhich version of the model was used is not stated.
- Version of Support Vector Classifier (SVC)Which version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00271, version 2, checked by a person on 2026-10-20. 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
Cite this record: AIxSci (2026). Machine learning predicts galaxy 'relicness' from survey measurements, cutting spectroscopy demand. Record aix-00271, version 2, checked 2026-10-20. https://aixsci.org/articles/aix-00271/