| Name: Yogesh Wadadekar |
| Affiliation: National Centre for Radio Astrophysics |
| Conference ID: ASI2026_642 |
| Title: Ensemble Machine Learning Photometric Redshifts for the Rubin/LSST Era: Performace and Readiness for Precision Cosmology |
| Abstract Type: Invited |
| Abstract Category: Facilities, Technologies and Data science |
| Author(s) and Co-Author(s) with Affiliation: Yogesh Wadadekar(NCRA-TIFR, 411007, India), Swagata Biswas(TCS Research, 700135, India), Shubhrangshu Ghosh(TCS Research, 700135, India), Avyarthana Ghosh(TCS Research, 700135, India), Arijit Mukherjee(TCs Research, 700135, India), Shailesh Deshpande(TCS Research, 411028, India), Arpan Pal(TCS Research, 700135, India) |
| Abstract: The imminent start of the Rubin Observatory’s Legacy Survey of Space and Time (LSST), with its unprecedented depth, cadence, and sky coverage over the next decade, marks a transformative phase for observational cosmology. Alongside earlier wide-area surveys such as the Sloan Digital Sky Survey and Euclid, LSST will fundamentally reshape our ability to probe galaxy evolution, large-scale structure, and dark energy. A key prerequisite for fully exploiting LSST data is the accurate and robust estimation of photometric redshifts for billions of faint galaxies using multi-band optical photometry alone.
In this work, we present a new ensemble-based machine learning framework designed specifically for LSST-like data, targeting faint galaxies and extending reliably to higher redshifts. The model uses only optical "grizy" photometry, aligning closely with LSST observing constraints. Our scaled ensemble architecture combines multiple learning algorithms—including gradient boosting, extreme gradient boosting, k-nearest neighbors, and artificial neural networks—trained on bagged input datasets to enhance robustness and predictive accuracy. Compared to individual models, the ensemble delivers consistently improved performance across all redshift ranges, maintaining strong accuracy up to z ∼ 4.
We validate the framework using publicly available Subaru telescope data from the Hyper Suprime-Cam Strategic Survey Program, which provides an excellent precursor dataset for LSST. The resulting photometric redshifts show significant improvements in precision, bias control, and catastrophic outlier rates. Importantly, the achieved performance meets—and in several cases exceeds—almost all the benchmarks defined in the LSST Science Requirements Document. Our work demonstrates the readiness of ensemble ML approaches for photometric redshift estimation in the LSST era and highlights their potential for enabling precision cosmology with the upcoming Rubin Observatory data. |
