Leveraging Deep Learning for Time Series Extrinsic Regression in predicting photometric metallicity of Fundamental-mode RR Lyrae Stars

Lorenzo Monti, Tatiana Muraveva, Gisella Clementini, Alessia Garofalo·October 23, 2024

Summary

Deep learning models predict photometric metallicity of RR Lyrae stars using Gaia optical G-band light curves, achieving low mean absolute error and high R2 regression performance. This showcases deep learning's effectiveness in astronomical research with large datasets, emphasizing its importance in analyzing complex astronomical phenomena. The study focuses on 6002 stars from Gaia DR3, using phase folding and alignment techniques to analyze their periodic variability. Key parameters include pulsation period, G-band amplitude, number of epochs, and photometric metallicity. The dataset is divided into a training set of 4801 stars and a validation set of 1201 stars. The text discusses the importance of aligning observations for studying stars with irregular variability patterns, focusing on RRab type stars. A sawtooth-shaped light curve, asymmetric with a rapid rise and slow decline, is highlighted. The study discusses the distribution of 6002 RRab stars' metallicities, visualized through histograms and kernel density estimates. The methodology section explores deep learning model selection and optimization for predicting metallicity from photometric light curves, testing nine models. The text discusses the choice of neural network architecture for time-series data, emphasizing the unique functionalities and applications of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and hybrid models. The study utilizes a workstation with an NVIDIA GeForce RTX 4070 GPU and Python 3.10, TensorFlow 2.13.0, Keras 2.13.1, and CuDNN 11.5 libraries for training. The experiments evaluated the regression performance of pre-processed models using standard metrics on stratified k-fold validation datasets, indicating minimal generalization error. The best-performing model, featuring three GRU blocks, L1 and L2 regularizations, and dropout layers, was designed to balance complexity and performance, ensuring robust predictions and computational efficiency.

Key findings

7

Tables

1

Introduction
Background
Overview of RR Lyrae stars and their significance in astronomy
Importance of photometric metallicity in understanding stellar properties
Brief on Gaia mission and its role in astronomical research
Objective
Aim of the study: applying deep learning to predict photometric metallicity
Expected outcomes: low mean absolute error and high R2 regression performance
Method
Data Collection
Source of data: Gaia DR3
Description of the dataset: 6002 stars, including pulsation period, G-band amplitude, number of epochs, and photometric metallicity
Data Preprocessing
Techniques used: phase folding and alignment for irregular variability patterns
Focus on RRab type stars: sawtooth-shaped light curves with rapid rise and slow decline
Model Selection and Optimization
Exploration of deep learning models for predicting metallicity
Evaluation of nine models, considering neural network architectures for time-series data
Choice of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and hybrid models
Implementation
Hardware and Software
Workstation specifications: NVIDIA GeForce RTX 4070 GPU, Python 3.10, TensorFlow 2.13.0, Keras 2.13.1, CuDNN 11.5
Training Process
Training methodology: stratified k-fold validation for evaluating model performance
Metrics used: standard regression metrics to assess generalization error
Results
Model Performance
Best-performing model: three GRU blocks, L1 and L2 regularizations, and dropout layers
Balance between model complexity and performance
Computational efficiency and robust predictions
Conclusion
Summary of Findings
Deep learning's effectiveness in predicting photometric metallicity of RR Lyrae stars
Importance of alignment techniques for studying stars with irregular variability patterns
Potential implications for future astronomical research and deep learning applications
Basic info
papers
instrumentation and methods for astrophysics
artificial intelligence
Advanced features
Insights
What dataset was used in the study, and how was it divided for the analysis?
What hardware and software were utilized for training the deep learning models, and what metrics were used to evaluate their performance?
Which deep learning models were tested for predicting the photometric metallicity of RR Lyrae stars, and what were the key features of the best-performing model?