Cluster-to-cluster differences in variability of YSOs identified with deep learning methods

We employed the Vizier SED API to collect the most complete spectral energy distributions for nearly 1 million young stellar object (YSO) candidates based on existing YSO catalogues from the literature. Using various deep learning methods we re-classified and selected more than 160 thousand high probability YSO candidates with reliable distance values. With a 3D minimum spanning tree, over 300 young clusters with at least 20 members were identified. We have also collected the available light curve statistics for our YSOs from ZTF and ASAS-SN and found significant differences in the variability between members of different clusters. For a smaller sample of sources, we used up-to-date Gaia alert light curves and developed a Long Short-Term Memory based autoencoder for their characterisation. We then combined the latent space generated by the autoencoder with statistical features for k-means clustering, effectively clustering the sources. Subsequently, we employed t-distributed stochastic neighbour embedding to downproject this high-dimensional data to a 2-dimensional space, enabling the identification of distinct patterns in YSO behaviour. With this method, that we will extend to the Gaia DR3 and DR4 data, we are able to identify stochastic variability with different amplitudes, single or multiple outliers, like flares, dimming and outbursting events on different timescales, and monotonic brightening or fading. Beyond details of the work, in this talk I will also discuss the differences of the variability seen in members of different young clusters, and correlate the time dependent features to the cluster density, age, mass and environment.