Toward model-free stellar chemical abundances - Potential applications in the search for chemically peculiar stars in large spectroscopic surveys

Home

All issues

Volume 704 (December 2025)

A&A, 704 (2025) A279

Full HTML

Open Access

Table B.1

Hyper-parameter configurations.

Parameter	Value
Architecture Parameters
Encoder convolutional layers size	(128, 256, 512)
Encoder convolutional filters size	(5, 11, 21)
Encoder fully connected layers size	(128, 64, 32)
Decoder layers size	(32, 64, 128)
Training Parameters
Optimizer	Adam (Kingma 2014)
Maximum learning rate	10⁻³
Learning rate scheduler	1Cycle (Smith & Topin 2019)
Training steps	4000
Batch size	512
Weighting Parameters (Eq. (7))
β	5 × 10⁻³
λ	${\begin{matrix} 0, & if step \leq 200 \\ 10^{- 2}, & if 200 < step \leq 400 \\ 10^{- 1}, & if 400 < step \leq 1000 \\ 1, & if step > 1000 \end{matrix}$ $\begin{cases}0, & \text { if step } \leq 200 \\ 10^{-2}, & \text { if } 200<\text { step } \leq 400 \\ 10^{-1}, & \text { if } 400<\text { step } \leq 1000 \\ 1, & \text { if step }>1000\end{cases}$

Notes. This step refers to the number of training epochs (full passes through the dataset), not individual gradient updates. The annealing schedule for λ is designed to allow the network to focus on reconstruction early on, before gradually introducing the disentanglement term.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.