Work

Enhancing photometric redshift estimation for LSST with a hybrid LSTM-mixture density network

Year: 2026

Type: article

Abstract: Context . Accurate photometric redshift (photo- z ) estimation and robust uncertainty quantification are essential for the Legacy Survey of Space and Time (LSST) to achieve its precision cosmology goals. Traditional machine learning algorithms are largely restricted to point estimates, struggling to characterize the multimodal nature of redshift probability density functions (PDFs) and the degeneracies within the color–redshift space. Aims . We validated the LSTM-MDNz architecture, which integrates sequential feature extraction with flexible probability density modeling. Our goal was to enhance both prediction accuracy and uncertainty calibration across a broad redshift range, meeting the stringent data quality requirements necessitated by next-generation cosmological analysis. Methods . The LSTM-MDNz framework treats multiband photometry as wavelength-ordered sequences, utilizing long short-term memory (LSTM) networks to capture nonlinear evolutionary correlations across the spectral energy distribution. A mixture density network (MDN) was then employed to explicitly model posterior PDFs via Gaussian mixture models. Performance was evaluated on the Hyper Suprime-Cam GalaxiesML dataset – which serve as a small-scale proxy for next-generation surveys such as LSST – and benchmarked against an established Bayesian neural network (BNN) architecture. Results . The proposed model consistently outperforms the BNN baseline, achieving a ∼10% improvement in point-estimation accuracy (specifically across RMSE, MAE, scatter, and σ NMAD ) and a ∼20% reduction in both outlier and catastrophic outlier rates. A uniform probability integral transform distribution confirms well-calibrated probabilistic outputs. Furthermore, the PDF-based confidence metric z conf enables high-purity catalog construction: excluding just approximately 4% of extremely low-confidence ( z conf < 0.05) samples reduces the overall outlier rate by ∼48%. Conclusions . The LSTM-MDNz framework offers a scalable and high-precision approach for photometric redshift estimation in LSST-scale surveys. Its robust uncertainty quantification capability provides important support for precision cosmology, particularly in advancing weak lensing and large-scale structure studies.

Source: Astronomy and Astrophysics

DOI: 10.1051/0004-6361/202660374