Unmasking the Cosmos: Discovering Hidden Symbiotic Stars with AI

Prompt

Title: Unveiling Hidden Symbiotic Stars: A Variational Autoencoder Approach to Anomaly Detection Introduction Symbiotic stars (SySts) are interacting binary systems in which a cool, evolved red giant star accretes onto a white dwarf companion. Conventional methods for identifying SySts often rely on direct detection of X-rays or rapid optical brightness variations from an accretion disk. However, these signatures are often masked by the luminous giant and/or the bright wind nebula that often surrounds these binary systems. Lucy et al. (2023) proposed a novel approach using data from the SkyMapper Southern Sky Survey to identify outliers from among hundreds of thousands of red giant stars that are actually symbiotics, revealing 12 new SySts and additional candidates. This proposal builds on their work by leveraging data from the latest SkyMapper data-Data Release 4 (DR4)--and employing a variational autoencoder (VAE) framework to identify symbiotic stars, which remain elusive in traditional surveys. Data Background Lucy et al. analyzed SkyMapper DR2 data to create a starting dataset of 366,721 red giant stars. With DR4’s expanded dataset, over 1,000,000 red giants are now available, presenting an opportunity for us to uncover a greater number of wide interacting binary stars. However, changes in parameter definitions, such as the removal of the nch_max flag and the introduction of new object merging techniques, require us to undertake careful preprocessing. Using TOPCAT’s TAP Access Point, we extracted updated photometric data for our analysis. Methods and Approaches 1. Supervised Learning (Benchmarking Past Approaches) Akras et al. (2019) applied machine learning to identify SySts using two other data sets (2MASS and WISE) data. Their approach involved: Diagnostic color-color diagrams for distinguishing SySts from mimicking objects. Machine learning classifiers (classification trees, LDA, and KNN) to select color indices. A classification tree model that achieved high accuracy (85–96%) in recovering known SySts and identifying new candidates. While effective, these methods are limited by their reliance on predefined feature sets and supervised learning constraints. 2. Unsupervised Learning for Anomaly Detection To address limitations in supervised methods, we propose an unsupervised learning approach to identify SySts lacking accretion disks or nuclear burning. Key techniques include: Principal Component Analysis (PCA): Reducing feature dimensionality to uncover inherent structures in the dataset. K-Means Clustering: Exploring hyperparameter variations to cluster potential SySt candidates. Variational Autoencoder (VAE): Learning a latent representation of normal (non-anomalous) objects and identifying anomalies through reconstruction error. SySts without nuclear burning would manifest as outliers due to their deviation from the learned distribution of normal objects. Expected Outcomes and Impact Newly Identified Symbiotic Candidates: Using our VAE approach, we aim to detect SySts that elude current detection methods. Improved Understanding of Wide Binaries: Analysis of DR4 data will refine our knowledge of binary system formation and evolution. Pioneering Machine Learning Techniques in Astronomy: Demonstrating the effectiveness of unsupervised learning and foundation models in stellar classification. References Akras, S. et al. (2019). Machine Learning Approach for Identification and Classification of Symbiotic Stars Using 2MASS and WISE. Monthly Notices of the Royal Astronomical Society. Link Birk, J., Hallin, A., & Kasiecz, G. K. (2024). OmniJet-α: The First Cross-Task Foundation Model for Particle Physics. Lanusse, F., et al. (2024). AstroCLIP: Cross-Modal Contrastive Learning for Astronomical Foundation Models. Conclusion This research leverages cutting-edge machine learning techniques to uncover hidden symbiotic stars, expanding our astrophysical knowledge and pioneering new methodologies in anomaly detection. Presenting this work at Columbia University’s Data Science Day will facilitate interdisciplinary collaboration and propel the field forward

Image Details