# SUPPNet: Neural network for stellar spectrum normalisation

**Authors:** Tomasz Różański, Ewa Niemczura, Jakub Lemiesz, Natalia Posiłek, Paweł Różański  
**arXiv:** [2111.15052](https://arxiv.org/abs/2111.15052)  
**Date:** 2021-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.15052v1/

## Key points

- Stellar spectroscopic analyses require precise continuum normalization of merged échelle spectra
- Automatic methods for continuum normalization have limited effectiveness
- Manual normalization is time-consuming
- Researchers have developed a fully automated normalization tool that works with order-merged spectra and offers flexible manual fine-tuning if necessary
- The proposed method uses a novel fully convolutional deep neural network (SUPP Network) trained to predict a pseudo-continuum, followed by post-processing using smoothing splines that provide access to regressed knots useful for optional manual corrections
- Active learning technique was applied to deal with possible biases that may arise from training with synthetic spectra and extend the applicability of the proposed method to features absent in this kind of spectra
- The developed normalization method was tested on high-resolution spectra of stars ranging from spectral types O to G and gave root mean squared (RMS) errors over the set of test stars equaling 0.0128 in the spectral range from 3900 Å to 7000 Å and 0.0081 in the range from 4200 Å to 7000 Å.
- The proposed method gives results comparable to careful manual normalization and can significantly reduce time-consuming manual efforts while providing accurate results comparable to those obtained through careful manual normalization.
- This approach is general and can be used in other fields of astronomy where background modeling or trend removal is part of data processing.
- The algorithm is available online at https://git.io/JqJhf for easy access by users worldwide.

## Summary

Stellar spectroscopic analyses require precise continuum normalization of merged échelle spectra, which is a demanding task due to the variety of features present in star spectra. While automatic methods have limited effectiveness, manual normalization is time-consuming. To address this issue, researchers have developed a fully automated normalization tool that works with order-merged spectra and offers flexible manual fine-tuning if necessary. The proposed method uses a novel fully convolutional deep neural network (SUPP Network) trained to predict a pseudo-continuum, followed by post-processing using smoothing splines that provide access to regressed knots useful for optional manual corrections. The active learning technique was applied to deal with possible biases that may arise from training with synthetic spectra and extend the applicability of the proposed method to features absent in this kind of spectra. The developed normalization method was tested on high-resolution spectra of stars ranging from spectral types O to G and gave root mean squared (RMS) errors over the set of test stars equaling 0.0128 in the spectral range from 3900 Å to 7000 Å and 0.0081 in the range from 4200 Å to 7000 Å. Experiments with synthetic spectra give RMS errors of approximately 0.0050. The proposed method gives results comparable to careful manual normalization and can significantly reduce time-consuming manual efforts while providing accurate results comparable to those obtained through careful manual normalization. Furthermore, this approach is general and can be used in other fields of astronomy where background modeling or trend removal is part of data processing. The algorithm is available online at https://git.io/JqJhf for easy access by users worldwide. The researchers also discuss how they overcame limitations associated with previous methods by using active learning techniques during training and evaluation stages as well as exploring different architectures for their neural network model during the second step of exploration. This work presents an innovative solution for precise continuum normalization that can be widely used across various astronomical fields requiring background modeling or trend removal processes as part of their data processing tasks.