TY - BOOK

T1 - Classification of Supernovae and Stars in the Era of Big Data and Artificial Intelligence

AU - Carrick, Jon

PY - 2022

Y1 - 2022

N2 - In recent years, artificial intelligence (AI) has been applied in manyfields of research. It is particularly well suited to astronomy, in whichvery large datasets from sky surveys cover a wide range of observations.The upcoming Legacy Survey of Space and Time (LSST) presentsunprecedented big data challenges, requiring state-of-the-art methodsto produce, process and analyse information. Observations of Type Iasupernovae help constrain cosmological parameters such as the darkenergy equation of state, and AI will be instrumental in the nextgeneration of cosmological measurements due to limited spectroscopicresources. AI also has the ability to improve our astrophysical understandingby perceiving patterns in data which may not be obvious tohumans.In this thesis we investigate how advanced AI methods can be usedin classification tasks: to identify Type Ia supernovae for cosmologyfrom photometry using supervised learning; by determining alow-dimensional representation of stellar spectra, and inferring astrophysicalconcepts through unsupervised learning.In preparation for photometric classification of transients from LSST werun tests with different training samples. Using estimates of the depthto which the 4-metre Multi-Object Spectroscopic Telescope (4MOST)Time-Domain Extragalactic Survey (TiDES) can classify transients, wesimulate a magnitude-limited training sample reaching rAB = 22.5 mag.We run our simulations with the software snmachine, a photometricclassification pipeline using machine learning. The machine-learningalgorithms struggle to classify supernovae when the training sampleis magnitude-limited as its features are not representative of the testset. In contrast, representative training samples perform very well,particularly when redshift information is included. Classificationperformance noticeably improves when we combine the magnitude-limitedtraining sample with a simulated realistic sample of faint,high-redshift supernovae observed from larger spectroscopic facilities;the algorithms' range of average area under ROC curve (AUC) scoresover 10 runs increases from 0.547-0.628 to 0.946-0.969 and purity ofthe classified sample reaches 95% in all runs for 2 of the 4 algorithms.By creating new, artificial light curves using the augmentation softwareavocado, we achieve a purity in our classified sample of 95% in all10 runs performed for all machine-learning algorithms considered. Wealso reach a highest average AUC score of 0.986 with the artificialneural network algorithm. Having real faint supernovae to complementour magnitude-limited sample is a crucial requirement in optimisationof a 4MOST spectroscopic sample. However, our results are a proofof concept that augmentation is also necessary to achieve the bestclassification results.During our investigation into an optimised training sample, we assumedthat every training object has the correct class label. Spectroscopy is areliable method to confirm object classification and is used to define ourtraining sample. However, it is not necessarily perfect and we thereforeconsider the impact of potential misclassifications of training objects.Taking the predicted error rates in spectroscopic classification from theliterature, we apply contamination to a TiDES training sample usingsimulated LSST data. With the recurrent neural network from thesoftware SuperNNova, we determine appropriate hyperparametersusing a perfect, uncontaminated TiDES training sample and thentrain a model on its contaminated counterpart to study its effectson photometric classification. We find that a contaminated trainingsample produces very little difference in classification performance, evenwhen increasing contamination to 5%. Contamination causes moreobjects of both Type Ia and non-Ia to be classified as Ia, increasingefficiency, but decreasing purity, with changes of less than 1% onaverage. Similarly, we see a decrease of 0.1% in average accuracy, andno clear difference in AUC score, only varying at the fourth significantfigure. These results are promising for photometric classification.Contaminated training appears to have little impact and propagationto cosmological measurements is expected to be minimal.In a separate study, we apply deep learning to data in the EuropeanSouthern Observatory (ESO) archive using an autoencoder neuralnetwork with the aim of improving similarity-based searches using thenetwork's own interpretation of the data. We train the network toreconstruct stellar spectra by passing them through an informationbottleneck, creating a low-dimensional representation of the data. Wefind that this representation includes several informative dimensionsand, comparing to known astrophysical labels, see clear correlationsfor two key nodes; the network learns concepts of radial velocity andeffective temperature, completely unsupervised. The interpretationof the other informative nodes appears ambiguous, leaving room forfuture investigation.The results presented in this thesis emphasise the practical capabilitiesof AI in an astronomical context: Classification of astrophysical objectscan be conducted through supervised learning using known labels, aswell as unsupervised learning in a physics-agnostic process.

AB - In recent years, artificial intelligence (AI) has been applied in manyfields of research. It is particularly well suited to astronomy, in whichvery large datasets from sky surveys cover a wide range of observations.The upcoming Legacy Survey of Space and Time (LSST) presentsunprecedented big data challenges, requiring state-of-the-art methodsto produce, process and analyse information. Observations of Type Iasupernovae help constrain cosmological parameters such as the darkenergy equation of state, and AI will be instrumental in the nextgeneration of cosmological measurements due to limited spectroscopicresources. AI also has the ability to improve our astrophysical understandingby perceiving patterns in data which may not be obvious tohumans.In this thesis we investigate how advanced AI methods can be usedin classification tasks: to identify Type Ia supernovae for cosmologyfrom photometry using supervised learning; by determining alow-dimensional representation of stellar spectra, and inferring astrophysicalconcepts through unsupervised learning.In preparation for photometric classification of transients from LSST werun tests with different training samples. Using estimates of the depthto which the 4-metre Multi-Object Spectroscopic Telescope (4MOST)Time-Domain Extragalactic Survey (TiDES) can classify transients, wesimulate a magnitude-limited training sample reaching rAB = 22.5 mag.We run our simulations with the software snmachine, a photometricclassification pipeline using machine learning. The machine-learningalgorithms struggle to classify supernovae when the training sampleis magnitude-limited as its features are not representative of the testset. In contrast, representative training samples perform very well,particularly when redshift information is included. Classificationperformance noticeably improves when we combine the magnitude-limitedtraining sample with a simulated realistic sample of faint,high-redshift supernovae observed from larger spectroscopic facilities;the algorithms' range of average area under ROC curve (AUC) scoresover 10 runs increases from 0.547-0.628 to 0.946-0.969 and purity ofthe classified sample reaches 95% in all runs for 2 of the 4 algorithms.By creating new, artificial light curves using the augmentation softwareavocado, we achieve a purity in our classified sample of 95% in all10 runs performed for all machine-learning algorithms considered. Wealso reach a highest average AUC score of 0.986 with the artificialneural network algorithm. Having real faint supernovae to complementour magnitude-limited sample is a crucial requirement in optimisationof a 4MOST spectroscopic sample. However, our results are a proofof concept that augmentation is also necessary to achieve the bestclassification results.During our investigation into an optimised training sample, we assumedthat every training object has the correct class label. Spectroscopy is areliable method to confirm object classification and is used to define ourtraining sample. However, it is not necessarily perfect and we thereforeconsider the impact of potential misclassifications of training objects.Taking the predicted error rates in spectroscopic classification from theliterature, we apply contamination to a TiDES training sample usingsimulated LSST data. With the recurrent neural network from thesoftware SuperNNova, we determine appropriate hyperparametersusing a perfect, uncontaminated TiDES training sample and thentrain a model on its contaminated counterpart to study its effectson photometric classification. We find that a contaminated trainingsample produces very little difference in classification performance, evenwhen increasing contamination to 5%. Contamination causes moreobjects of both Type Ia and non-Ia to be classified as Ia, increasingefficiency, but decreasing purity, with changes of less than 1% onaverage. Similarly, we see a decrease of 0.1% in average accuracy, andno clear difference in AUC score, only varying at the fourth significantfigure. These results are promising for photometric classification.Contaminated training appears to have little impact and propagationto cosmological measurements is expected to be minimal.In a separate study, we apply deep learning to data in the EuropeanSouthern Observatory (ESO) archive using an autoencoder neuralnetwork with the aim of improving similarity-based searches using thenetwork's own interpretation of the data. We train the network toreconstruct stellar spectra by passing them through an informationbottleneck, creating a low-dimensional representation of the data. Wefind that this representation includes several informative dimensionsand, comparing to known astrophysical labels, see clear correlationsfor two key nodes; the network learns concepts of radial velocity andeffective temperature, completely unsupervised. The interpretationof the other informative nodes appears ambiguous, leaving room forfuture investigation.The results presented in this thesis emphasise the practical capabilitiesof AI in an astronomical context: Classification of astrophysical objectscan be conducted through supervised learning using known labels, aswell as unsupervised learning in a physics-agnostic process.

U2 - 10.17635/lancaster/thesis/1647

DO - 10.17635/lancaster/thesis/1647

M3 - Doctoral Thesis

PB - Lancaster University

ER -