Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Deep Multimodal Multi--Head Neural Network for Joint Estimation of Stellar Age, Lifetime, and Evolutionary Stage

Jing Rou Puah, Sasa Arsovski

TL;DR

This work tackles joint estimation of stellar age, lifetime, and evolutionary stage from SDSS DR17 multimodal data using a hybrid deep architecture that fuses an MLP for photometry with a spectral Transformer. It conducts a rigorous three-phase evaluation across eight architectures, uncovering a surprising advantage for data-level undersampling over weighted losses and highlighting a key precision-versus-realism trade-off when balancing regression and classification objectives. The final model achieves an Age RMSE of $0.093$ in $\log(\mathrm{yrs})$ space with calibrated uncertainties via Monte Carlo Dropout, while ensuring physically plausible predictions, thereby establishing a benchmark and methodological blueprint for multimodal, multi-task stellar parameter estimation. The findings provide practical guidance for scientific AI in astronomy and point toward future work that integrates additional surveys and develops specialized, constrained architectures.

Abstract

Accurate estimation of stellar parameters -- stellar age, lifetime, and evolutionary stage -- remains a fundamental challenge in astrophysics. We introduce a hybrid deep learning architecture combining multimodal spectroscopic and photometric data from SDSS DR17. The model comprises a Multi-Layer Perceptron for numerical features and a CNN with a Vision Transformer for spectra, with three output heads for age, lifetime, and evolutionary stage prediction. Training labels are derived from MIST v1.2 isochrones, with evolutionary stage binned into five classes (Hot, Medium, Cool, Subgiant, Red Giant). We conduct multi-phase evaluation: Phase I explores model architectures and data balancing strategies, Phase II tunes architectural complexity, and Phase III optimizes the multi-task loss composition. The final model achieves a balance between precision (Age RMSE 0.093 in $\log(\mathrm{yrs})$) and physical realism. Monte Carlo Dropout confirms well-calibrated uncertainties, enabling meaningful astrophysical interpretation and establishing a new benchmark for multimodal stellar parameter estimation.

A Deep Multimodal Multi--Head Neural Network for Joint Estimation of Stellar Age, Lifetime, and Evolutionary Stage

TL;DR

This work tackles joint estimation of stellar age, lifetime, and evolutionary stage from SDSS DR17 multimodal data using a hybrid deep architecture that fuses an MLP for photometry with a spectral Transformer. It conducts a rigorous three-phase evaluation across eight architectures, uncovering a surprising advantage for data-level undersampling over weighted losses and highlighting a key precision-versus-realism trade-off when balancing regression and classification objectives. The final model achieves an Age RMSE of in space with calibrated uncertainties via Monte Carlo Dropout, while ensuring physically plausible predictions, thereby establishing a benchmark and methodological blueprint for multimodal, multi-task stellar parameter estimation. The findings provide practical guidance for scientific AI in astronomy and point toward future work that integrates additional surveys and develops specialized, constrained architectures.

Abstract

Accurate estimation of stellar parameters -- stellar age, lifetime, and evolutionary stage -- remains a fundamental challenge in astrophysics. We introduce a hybrid deep learning architecture combining multimodal spectroscopic and photometric data from SDSS DR17. The model comprises a Multi-Layer Perceptron for numerical features and a CNN with a Vision Transformer for spectra, with three output heads for age, lifetime, and evolutionary stage prediction. Training labels are derived from MIST v1.2 isochrones, with evolutionary stage binned into five classes (Hot, Medium, Cool, Subgiant, Red Giant). We conduct multi-phase evaluation: Phase I explores model architectures and data balancing strategies, Phase II tunes architectural complexity, and Phase III optimizes the multi-task loss composition. The final model achieves a balance between precision (Age RMSE 0.093 in ) and physical realism. Monte Carlo Dropout confirms well-calibrated uncertainties, enabling meaningful astrophysical interpretation and establishing a new benchmark for multimodal stellar parameter estimation.

Paper Structure

This paper contains 27 sections, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Stellar Parameter Estimation
  4. Deep Learning in Stellar Astrophysics
  5. Multimodal and Multi-Task Learning
  6. Methodology
  7. Data and Data Acquisition
  8. Data Preprocessing Pipeline
  9. Data Filtering
  10. Label Generation and Engineering
  11. Photometric Processing
  12. Spectroscopic Processing
  13. Experimental Design
  14. Phase I: Foundational Experiments
  15. Phase II: Architectural Fine-Tuning
  16. ...and 12 more sections

Figures (4)

  • Figure 1: The overall three-phase research workflow.
  • Figure 2: General multimodal framework diagram.
  • Figure 3: Distribution of age prediction errors in the logarithmic space ($\log(\mathrm{yrs})$) for the final three model configurations.
  • Figure 4: Uncertainty quantification results for the final best-performing model (Setup A).