Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

IrisML: Neural Posterior Estimation for the Spectral Energy Distribution fitting

Mateusz Kapusta

TL;DR

IrisML addresses the slow, CPU-bound Bayesian inference used in spectral energy distribution fitting by introducing Neural Posterior Estimation that uses a set-invariant transformer to encode variable photometric inputs and a Masked Autoregressive Flow to learn the conditional posterior of stellar parameters. The method jointly infers $T$, $\log g$, [M/H], $A_V$, and $R_V$ from measurements spanning 25 filters, trained on BOSZ+$Fitzpatrick$ extinction syntheses. It achieves temperature estimates with RMSE ≈ $314$ K and metallicity with RMSE ≈ $0.31$ dex, while delivering a 3–4 order-of-magnitude speedup over CPU MCMC and enabling robust posterior coverage. The work highlights both the practical benefits for large-scale dust mapping and the remaining challenges in metallicity inference and OOD robustness for normalizing flows.

Abstract

Over the past 30 years, numerous large-scale photometric astronomical surveys have been conducted, including SDSS, Pan-STARRS, Gaia,2MASS, WISE, and others. These surveys provide extensive photometric measurements that can be used to infer a wide range of physical parameters of astronomical objects. Traditionally, Bayesian approaches, such as Markov Chain Monte Carlo (MCMC) sampling have been employed for such inference tasks. However, these methods tend to be computationally intensive and often require manual tuning or expert supervision. In this work, we propose a novel machine learning model designed to perform automatic and robust inference from photometric data, offering a scalable and efficient alternative to conventional techniques.

IrisML: Neural Posterior Estimation for the Spectral Energy Distribution fitting

TL;DR

IrisML addresses the slow, CPU-bound Bayesian inference used in spectral energy distribution fitting by introducing Neural Posterior Estimation that uses a set-invariant transformer to encode variable photometric inputs and a Masked Autoregressive Flow to learn the conditional posterior of stellar parameters. The method jointly infers , , [M/H], , and from measurements spanning 25 filters, trained on BOSZ+ extinction syntheses. It achieves temperature estimates with RMSE ≈ K and metallicity with RMSE ≈ dex, while delivering a 3–4 order-of-magnitude speedup over CPU MCMC and enabling robust posterior coverage. The work highlights both the practical benefits for large-scale dust mapping and the remaining challenges in metallicity inference and OOD robustness for normalizing flows.

Abstract

Over the past 30 years, numerous large-scale photometric astronomical surveys have been conducted, including SDSS, Pan-STARRS, Gaia,2MASS, WISE, and others. These surveys provide extensive photometric measurements that can be used to infer a wide range of physical parameters of astronomical objects. Traditionally, Bayesian approaches, such as Markov Chain Monte Carlo (MCMC) sampling have been employed for such inference tasks. However, these methods tend to be computationally intensive and often require manual tuning or expert supervision. In this work, we propose a novel machine learning model designed to perform automatic and robust inference from photometric data, offering a scalable and efficient alternative to conventional techniques.

Paper Structure

This paper contains 8 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Inference Task
  3. Methods
  4. Preprocessing
  5. Masked Autoregressive Flow
  6. Training and data generation
  7. Results
  8. Conclusions

Figures (4)

  • Figure 1: Schematic representation of the IrisML model. CLS stands for the classification token, which is learned during the training. Red and orange arrows indicate the movement of the data during the training and inference respectively.
  • Figure 2: Predicted and measured temperatures for stars in the Disc sample. Color represents the number of measurments used to infer the stellar properties.
  • Figure 3: Predicted and measured metallicities for stars in the Halo sample. Color represents the number of measurments used to infer the stellar properties.
  • Figure 4: The coverage plot for the APOGEE sample with the distinction between the Disc and the Halo samples.