A value-added catalogue of neural network-based europium abundances for GALAH DR4

TL;DR

This work develops a CNN-based label-transfer approach to extend GALAH DR3 $[Eu/H]$ abundances onto GALAH DR4 spectra, producing a public catalogue of $[Eu/H]$ for DR4 targets with a high-confidence giant golden sample and a metal-poor synthesis-based companion. The model is trained on a DR3/DR4 cross-match with high-SNR data and validated against DR3, achieving RMSE ≈ 0.074 dex and $R^2 ≈ 0.88$, while a Korg synthesis-based secondary catalogue addresses metal-poor regimes. The catalogue reproduces key Galactic chemical evolution trends, including elevated Eu in accreted stars and the knee near $[Fe/H] \\sim -1$, demonstrating the practical utility of data-driven Eu abundances for r-process studies and chemo-dynamical analyses. Overall, the study delivers over 100k high-quality $[Eu/H]$ abundances on a homogeneous scale, enabling robust investigations of Eu enrichment, accretion history, and the Galactic r-process landscape.

Abstract

The rapid neutron-capture ($r$-process) element europium (Eu) is a valuable tracer of neutron star mergers and other rare nucleosynthetic events. The stellar spectroscopic survey GALAH's unique wavelength range and setup include the Eu absorption feature at $\sim6645$ Å for almost a million stars in the most recent Data Release 4 (DR4). However, DR4 also saw a decreased precision in reported Eu measurements compared to previous data releases. In this work, we use a convolutional neural network (CNN) to perform label transfer, wherein we use the GALAH DR4 spectra and stellar parameters to infer DR3 [Eu/H] abundances. This CNN is then applied to DR4 spectra without corresponding DR3 Eu abundances to develop a new, publicly available catalogue of [Eu/H] values for high signal-to-noise targets. We include [Eu/H] predictions for $118\,946$ stars, out of which $54\,068$ giants constitute our "golden sample" of high-confidence predictions, which pass stricter quality cuts and have a reported precision $\lesssim0.1$. To overcome the scarcity of training data in the low metallicity regime, we provide an additional catalogue of [Eu/H] abundances for metal poor ($\mathrm{[Fe/H]}<-1$) stars derived from synthesis of the Eu feature. Our "golden sample" can be combined with [Eu/H] values from GALAH DR3 to create a catalogue of over $100\,000$ vetted, high-quality abundances on a homogeneous scale. Moreover, we are able to reproduce known science results, including the elevated Eu abundances of accreted stars and previously observed Galactic chemical evolution trends. This catalogue represents one of the largest available samples of [Eu/H] abundances for high signal-to-noise targets.

Figures (17)

• Figure 1: The GALAH DR4 versus GALAH DR3 [Eu/Fe] measurements for a subsample of stars with GALAH best practice cuts (https://www.galah-survey.org/dr3/using_the_data/), in log-scaled density of the number of stars. The black dashed diagonal line marks the 1:1 relation for illustrative purposes; the sharp cut at [Eu/Fe] = 1 is a feature of the GALAH DR4 pipeline. The $R^2$ score is listed in the lower left corner, demonstrating that GALAH DR4 [Eu/Fe] is not a good predictor of GALAH DR3 [Eu/Fe].
• Figure 2: A set of five example spectra from GALAH DR4 in the wavelength range used by the neural network in this work ($\sim6640$ Å to $\sim6650$ Å), with varying [Eu/H] values from GALAH DR3 indicated by the colorbar. A vertical offset of 0.1 is applied to the normalized flux for easier visualization. Each of these stars have similar stellar parameters in GALAH DR3, with $T_\textrm{eff}\approx4450$ K, $\log g\approx1.65$, and [Fe/H] $\approx-1.1$. The spectra are all interpolated onto our uniform grid of wavelengths. for input with the neural network. The Eu line as well as the Ni and two Fe lines within the spectra are marked.
• Figure 3: The three subplots illustrate the same sample of $52\,147$ stars used for the training and validation datasets ($52\,302$ including the $T_{\rm eff} > 6600$ K stars seen left of the dashed grey vertical line and 3 stars with corrupted Eu lines, which were eventually removed from the datasets), coloured by density (left) or chemical information (center/right). The dotted black line shows a boundary between GALAH DR3 dwarfs and giants from Borisov2022 used in the train-test split. Left: A 2D hexagonal binning of the stars on the Kiel diagram, coloured by density (the number of stars per bin). Center: As left, but coloured by average [Fe/H] within bins. Right: As left, but coloured by [Eu/H] values derived from GALAH DR3.
• Figure 4: The architecture of our two-branch neural network. Layers are represented by solid boxes, with dashed boxes denoting where dropout is applied and arrows marking the movement of input/output tensors between layers. 1D convolutional layers are labeled as "Conv1D," and linear/fully connected layers are denoted by the "Fully Connected" label.
• Figure 5: Performance of our model on the validation dataset. A typical (median $\sigma_{\text{y, pred}}$) errorbar for a single prediction is shown in the lower right corner in both subplots for reference. Top: A 2D histogram of predicted [Eu/H] values vs. labels (GALAH DR3 [Eu/H] values). Performance metrics RMSE and $R^2$ are listed in the upper left corner. Bottom: A 2D histogram of predicted - label residuals. The solid magenta line represents the mean residual value within 10 bins covering the full extent of the label values; the shaded magenta band around it represents 1$\sigma$ intervals. The predicted values show good agreement with labels, although some regression to the mean is still visible in the extreme values.