Accelerating Stellar Photometric Distance Estimates with Neural Networks
Building on the Bayesian approach to estimating stellar distances from broadband photometry, we show that the computation can be accelerated by about an order of magnitude by using neural networks. Focusing on the case of the ugrizy filter complement for Rubin’s Legacy Survey of Space and Time (LSST...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IOP Publishing
2025-01-01
|
| Series: | The Astronomical Journal |
| Subjects: | |
| Online Access: | https://doi.org/10.3847/1538-3881/addf51 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Building on the Bayesian approach to estimating stellar distances from broadband photometry, we show that the computation can be accelerated by about an order of magnitude by using neural networks. Focusing on the case of the ugrizy filter complement for Rubin’s Legacy Survey of Space and Time (LSST), we show that the Bayesian approach is equivalent to mapping from a 10-dimensional space of five measured colors and their uncertainties to a three-dimensional space of absolute magnitude, metallicity, and interstellar dust extinction along the line of sight. Once the neural network is trained, this mapping is faster by more than an order of magnitude compared to the Bayesian approach, for both optimized grid search and Markov chain Monte Carlo implementation methods. We have developed and tested a pipeline that achieves significant acceleration by first running the Bayesian method on 5%–10% of the sample, then using it to train a neural network, and finally processing the entire sample with the resulting neural network. This computation is done in patches of about 10 deg ^2 due to the variation of Bayesian priors across the sky. We present an analysis of pipeline performance, including speed and biases as functions of input stellar parameters and signal-to-noise ratio, using TRILEGAL-based simulated LSST catalogs by P. Dal Tio et al. We intend to run this pipeline on LSST data releases and make its outputs publicly available. |
|---|---|
| ISSN: | 1538-3881 |