Convolutional neural network for Lyman break galaxies classification and redshift regression in DESI (Dark Energy Spectroscopic Instrument)

Overview

• DESI is an international project to create a 3D map of the sky by observing over 40 million quasars and galaxies
• The goal is to use this map to study various aspects of cosmology, from dark energy to neutrino mass
• This paper focuses on using a convolutional neural network (CNN) to classify Lyman Break Galaxies (LBGs) and determine their redshift from their spectra

Plain English Explanation

The DESI project is an ambitious international effort to create a detailed 3D map of the universe by observing and cataloging over 40 million galaxies and quasars over a 5-year period. This 3D map will enable researchers to study a wide range of cosmological phenomena, from the mysterious "dark energy" that is driving the accelerating expansion of the universe to the tiny mass of neutrinos.

In this paper, the researchers are focusing on a specific type of galaxy observed by DESI called Lyman Break Galaxies (LBGs). LBGs appear in the distant universe due to the way their light is "redshifted" as it travels towards us over billions of years. By analyzing the spectra (light signatures) of these LBGs, the researchers aim to both confirm their identity and determine their exact distance from Earth using this redshift effect.

To do this, the researchers developed a convolutional neural network (CNN) model inspired by a previous system called QuasarNET. This CNN can simultaneously classify whether a galaxy is an LBG or not, and also predict its redshift (distance) value. To train the model, the researchers used data augmentation techniques like shifting the galaxy spectra and adding synthetic data to expand their training dataset from 3,019 examples to over 66,000.

After further refining the CNN architecture and optimizing its hyperparameters, the researchers were able to achieve significant improvements in the model's performance - up to a 26% gain on key evaluation metrics, particularly for LBGs at low and high redshifts. The best model ultimately reached an average score of 94%, a substantial improvement over the initial 75% baseline.

Technical Explanation

The researchers developed a convolutional neural network (CNN) inspired by QuasarNET to simultaneously classify Lyman Break Galaxies (LBGs) and predict their redshift from their spectra. This was done in the context of the larger DESI project, which aims to create a detailed 3D map of the observable universe by studying over 40 million galaxies and quasars.

Initially, the researchers had access to a dataset of 3,019 LBG spectra. To increase the size and diversity of the training data, they applied several data augmentation techniques, including shifting the spectra in wavelength, adding noise, and generating synthetic spectra. This expanded the dataset to over 66,000 examples.

In a second phase, the researchers made modifications to the QuasarNET architecture, incorporating transfer learning and using Bayesian optimization to tune the hyperparameters. These changes led to substantial performance gains, with improvements of up to 26% on the Purity/Efficiency curve - a key metric for evaluating model performance, particularly in regions of interest at low (around 2) and high (around 4) redshifts.

The best model ultimately achieved an average score of 94%, a significant improvement over the initial 75% baseline. This demonstrates the power of combining data augmentation, transfer learning, and hyperparameter optimization to enhance the performance of deep learning models for challenging astronomical tasks like classifying radio galaxies or analyzing galaxy images.

Critical Analysis

The researchers acknowledge several limitations and areas for further research in this work. For example, they note that the data augmentation techniques used, while effective, may not fully capture the true diversity of LBG spectra observed in the real universe. Additionally, the model was trained and evaluated on a relatively small dataset, and its performance on the full DESI dataset remains to be seen.

Furthermore, the researchers do not delve deeply into the potential sources of error or uncertainty in the redshift predictions made by the model. Accurate redshift estimation is crucial for properly mapping the 3D structure of the universe, so a more thorough analysis of the model's redshift prediction accuracy and its implications would be valuable.

Finally, while the researchers demonstrate significant performance gains compared to the initial QuasarNET model, it would be helpful to have a more detailed comparison to other state-of-the-art approaches for LBG classification and redshift estimation. This would provide a better understanding of how the proposed model compares to alternative techniques in the field.

Overall, this research represents an important step forward in leveraging deep learning to support the ambitious goals of the DESI project. However, further work is needed to fully understand the capabilities and limitations of the proposed approach, and to ensure its robust performance on the large-scale datasets that DESI will ultimately produce.

Conclusion

This paper presents a convolutional neural network (CNN) model that can simultaneously classify Lyman Break Galaxies (LBGs) and predict their redshift (distance) from their spectral data. This work is part of the larger DESI project, which aims to create a detailed 3D map of the observable universe by studying over 40 million galaxies and quasars.

The researchers used data augmentation techniques to expand their initial dataset of 3,019 LBG spectra to over 66,000 examples, and then made architectural modifications and performed hyperparameter tuning to improve the model's performance. They were able to achieve gains of up to 26% on key evaluation metrics, particularly for LBGs at low and high redshifts, with the best model reaching an average score of 94%.

While this research represents an important step forward, the researchers acknowledge several limitations and areas for further work, such as the need for a more thorough analysis of the model's redshift prediction accuracy and a more detailed comparison to other state-of-the-art approaches. Nonetheless, this work demonstrates the power of deep learning techniques to support the ambitious goals of large-scale astronomical projects like DESI, and lays the groundwork for continued advancements in this field.