The identification of exoplanets within habitable zones remains a central objective in modern astrophysics, particularly with the availability of large-scale photometric datasets from space-based missions such as the Transiting Exoplanet Survey Satellite (TESS). This study investigates the effectiveness of unsupervised machine learning techniques–specifically k-means and k-medians clustering–for analyzing and classifying light curves derived from galactic stellar populations. By extracting both basic and extended statistical features, dimensionality reduction methods including t-distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP) are employed to project high-dimensional data into interpretable low-dimensional spaces. To evaluate the relevance of the identified clusters, the results are systematically compared with the TESS Objects of Interest (TOI) catalog, incorporating information on confirmed planets and candidate signals. This comparison reveals that clusters containing known TOIs often include additional unlabeled objects, suggesting the presence of potentially undiscovered exoplanet candidates. Moreover, the clustering framework effectively distinguishes between transit-like signals and noise-dominated light curves, even in sectors with few or no known TOIs. These findings highlight the capability of unsupervised learning to recover known exoplanetary signals while simultaneously identifying new candidate-rich regions within the data. The proposed framework offers a scalable and data-driven approach for prioritizing targets in large survey datasets, contributing to the advancement of automated exoplanet detection pipelines.
Adhikary et al. (2026) studied this question.