Andreas W. Neitzel
CV
ESO/S. Brunier, CC BY 4.0

Publications

Manifold learning for stellar populations

In a simulated Milky Way, a method that needs no labels told disc from halo stars and, in error-free data, found the debris of two swallowed galaxies.

Workflow of Neitzel et al. 2025A pipeline of shapes, left to right: a simulated galaxy seen face-on and edge-on with the Sun's viewpoint; four sample clouds, two sharp and two blurred by errors; five feature rulers feeding one bundle; UMAP collapsing the bundle into a two-dimensional cloud, then a trustworthiness gate; HDBSCAN drawing two hulls, MC1 and MC2, with persistence bars above a baseline; a second run splitting MC1 into two and MC2 into three; and a three-by-three confusion grid with its diagonal filled, where the galaxy's own record arrives along a dashed return line.THE SURVEYSunFOUR SAMPLESABCD75,000 stars76,168 giantswith errorsFIVE FEATURESUMAP2-D10,000 epochsT > 0.90HDBSCANMC1MC2baselinepersistenceSECOND RUNTHE CHECKinformedness0.97 to 0.9997.6 to 98.7 %accreted in MC2the galaxy's own recordCreator: Andreas W. Neitzel. © Andreas W. Neitzel. All rights reserved. https://andreaswneitzel.github.io

Dissecting stellar populations with manifold learning

Twelve scatter plots in four columns, one per test sample (A to D), and three rows: velocity, alpha-element abundance against iron abundance, and iron abundance against age. Stars are coloured by the group the method assigned; columns A and B show five groups, columns C and D three.
Groups the method found without labels in the four test samples. With measurement errors (C and D) the disc stays one group. From Neitzel et al. 2025, A&A 695, A243, Fig. 11, CC BY 4.0.

The question

The Milky Way grew from stars it formed itself and from smaller galaxies it swallowed. Stars from each episode differ in age, chemistry and motion, but the groups overlap and are hard to separate (Sect. 1). The paper tests whether manifold learning, an unsupervised machine-learning approach that assumes little about how many groups exist or what they look like, can find them.

What we did

The test bed was a simulated galaxy with a known history: m12f, from the FIRE-2 simulations, which went through two major mergers, 6.5 and 1.4 billion years ago (Sect. 2.1). A mock Gaia survey of it gave four samples (Sect. 2.2): 75,000 stars at all stages of evolution (A), and 76,168 red giants like those in the TESS catalogue of oscillating giants, with no errors (B), realistic errors (C) or optimistic errors (D).

UMAP turned five numbers per star (age, two chemical abundances, two measures of velocity) into a position on a two-dimensional map, and HDBSCAN found the groups on that map (Sect. 3).

What came out

The method separated disc from halo in every sample. The first pass scored an informedness of 0.966 to 0.989, where 1 is a perfect match, and put 97.6 to 98.7 per cent of the stars born in other galaxies into the halo group, although no input said where a star was born (Sect. 4.1, Table 2).

A second pass on each group found more structure. Samples A and B gave five groups: two discs with slightly different chemistry, the halo that formed in place, and one group per merger (Table 3, Fig. 11). With measurement errors added (C and D), the discs could no longer be told apart, and the halo split into two groups instead of three (Sect. 4.3).

My part

I am the first and corresponding author; the co-authors are my PhD supervisors. The paper reports the first project of my PhD: I built the pipeline, compared UMAP with PCA and t-SNE (Appendix D, on Zenodo) and ran the tests. Starfold, a Python package I develop, extends the method.