Exploring substructures in the Milky Way halo

Context. The identification of stellar structures in the Galactic halo, including stellar streams and merger remnants, often relies on the dynamics of their constituent stars. However, this approach has limitations due to the complex dynamical interactions between these structures and their environment. Perturbations such as tidal forces exerted by the Milky Way, the potential escape of stars, and passages through the Galactic plane can result in the loss of dynamical coherence of stars in these structures. Consequently, relying solely on dynamics may be insufficient for detecting such disrupted or dispersed remnants. Aims. We combine chemistry and dynamics, integrated through a system of neural networks, to develop a clustering method for identifying accreted structures in the Galactic halo. Methods. We developed an integrated approach combining Siamese neural networks (SNNs), graph neural networks (GNNs), autoencoders, and the O<ns0:sc>PTICS</ns0:sc> algorithm to create a comprehensive procedure named C<ns0:sc>REEK</ns0:sc>. This method is designed to uncover stellar structures in the Galactic halo. Initially, C<ns0:sc>REEK</ns0:sc> was trained on known globular clusters (GCs) and then applied to the dataset to identify stellar streams. Results. C<ns0:sc>REEK</ns0:sc> successfully recovered 80% of the GCs present in the APOGEE dataset, re-identified several known stellar streams, and identified a potential new stream. Additionally, within highly populated stellar structures, C<ns0:sc>REEK</ns0:sc> can identify substructures that exhibit distinct chemical compositions and orbital energies. This approach provides an objective data-driven method for selecting stars associated with streams and stellar structures in general.