Learning orbital dynamics of binary black hole systems from gravitational wave measurements

Abstract: We introduce a gravitational waveform inversion strategy that discovers mechanical models of binary black hole (BBH) systems. We show that only a single time series of (possibly noisy) waveform data is necessary to construct the equations of motion for a BBH system. Starting with a class of universal differential equations parameterized by feed-forward neural networks, our strategy involves the construction of a space of plausible mechanical models and a physics-informed constrained optimization within that space to minimize the waveform error. We apply our method to various BBH systems including extreme and comparable mass ratio systems in eccentric and non-eccentric orbits. We show the resulting differential equations apply to time durations longer than the training interval, and relativistic effects, such as perihelion precession, radiation reaction, and orbital plunge, are automatically accounted for. The methods outlined here provide a new, data-driven approach to studying the dynamics of binary black hole systems.

A Deep Learning approach to Large Scale Structure Cosmology

Abstract: I will review the current status of machine learning as it is applied to the large scale structures in the Universe.

Galaxy surveys, weak gravitational lensing, 21cm intensity radio projects and the CMB are providing us with maps of the large scale structures in the Universe.

These maps are typically analysed with tried and tested statistical techniques such as correlation functions and power spectra, which can relate observation to a theoretical model.

However, recently there has been many works exploring the use of machine learning methods to extract cosmological physics from astronomical data.

In this talk I will show a few examples of these works and give a more detailed exposition using my own recent work on probing the nature of dark matter using 21cm tomography with convolutional neural networks (arxiv:2108.07972).

Deep learning classification of transient noises using LIGOs auxiliary channel data

Abstract: The LIGO gravitational wave detector is a very sophisticated and sensitive device that can detect minute gravitational changes from space. In order to monitor the operation of these devices and the surrounding environment, various sensors are installed and operated around the detector. This is called an auxiliary channel, and it is divided into those directly related to gravitational waves (unsafe) and those not related (safe), and the safe channel is used for gravitational wave verification and detector characteristics research. The gravitational wave signal detected by the detector is observed in the main channel. In the main channel, not only gravitational-wave signals but also various types of signals (non-Gaussian transient noise) are detected. We call it Glitch and call it Blip, Helix, Low Frequency Burst, Low Frequency lines, Scratchy, Whistle, etc. by morphological classification. We will introduce a study of learning and classifying the various glitches that appear in the main channel using only the safe channel, and deriving a list of auxiliary channels that are correlated with the glitch by analyzing the trained deep learning model.

Artificial neural network as an efficient selection algorithm for high-redshift quasars

Abstract: Despite red colors of distant quasars due to the absorption by the neutral hydrogen in the intergalactic medium along the line of sight, discovering the quasars is still difficult since their red colors are similar to late-type stars' colors. Color selection criteria using a few broadbands are widely applied for separating quasars from stars on two-dimensional color-color spaces, however, it is only feasible for selecting bright quasars at a specific range of redshift. To alleviate the drawback of color selection, we tried to make a novel selection consisting of an artificial neural network (ANN) and the Bayesian information criterion (BIC). ANN determines whether an object is a quasar or non-quasar using multi-dimensional decision boundaries based on all the photometric results, and the BIC picks promising candidates by obtaining quasar and star best-fit SED models. Comparing confusion matrices of color criteria and our novel selections, we found that the former has few contaminant sources, however, it misses a lot of quasars except for quasars at z = 4.7-5.1. On the other hand, the latter can maximize the chance of finding high-redshift quasars as well as minimize the probability of including the stellar contaminant in candidates. As a result, we can efficiently select quasar candidates that are one magnitude fainter than quasars found by previous works.

Topological features of sliding window embeddings for signal detection

Abstract: Persistent homology of sliding window embeddings has been shown to be a useful feature extraction tool in time series classification. Convolutional neural networks, which have become famous for their high efficacy when dealing with image data inputs, have also seen some appreciation for their performance on time series analysis problems including signal detection and classification. Due to the nature of the convolution layers, it can be effective to naively combine normalized raw input with features extracted by explicit processes such as TDA. This has turned out to be effective as a synergistic combination for difficult signal detection and classification problems, such as the detection of gravitational wave signatures in interferometer readings.

Artificial Intelligence in Transient Astronomy: Dataset for CNN-TDA Net Classifier of Real and Bogus

Abstract: Transients, fast decaying astrophysical events, such as supernovae (SNe), kilonovae (KNe), and gamma-ray bursts (GRBs), have become important astrophysical objects since the advent of time-domain astronomy. Transients are commonly found in an image which shows the difference between the observed image and a reference image through image subtraction (difference image). However, image subtraction is not always perfect, and the difference image almost always contains not only true transients, but also a number of image subtraction artifacts that could be confused as astrophysical sources. Exclusion of artifacts is critical, but it involves a time-consuming process of visual inspection of many (thousands and more) transient candidates, which delays the transient search process. For fast and reliable real (transient) and bogus (artifact) classification of signals in difference images, we have been constructing CNN-TDA Net, a deep learning architecture which combines the convolutional neural network (CNN) and the topological data analysis (TDA) methods. Here, we introduce previous studies utilizing CNN for real and bogus classification, and present the image datasets that have been used for training CNN-TDA Net. To make the dataset we used images taken almost every night for two months as a part of the Intensive Monitoring Survey of Nearby Galaxies with the LOAO 1-m telescope. Postage stamp images of real/bogus objects were generated by gpPy (automatic image process pipeline), and labeling was conducted by matching detected sources in difference images with reported transient objects and visual inspection. As a result, the postage stamp image dataset for an initial test of CNN-TDANet is constructed, containing 351 transient images and 3,365 bogus images (real:bogus=1:9.6). The real transient sample consists of three SNe at dim and bright phases, and two variable objects. Finally, we will also show future works for improving the AI test dataset.

A brief tutorial on CNN-TDA Net for transient detection

Abstract: We provide an introductory tutorial session of CNN-TDA Net for transient and bogus classifications. We will explain how to compute and analyze the main tool of TDA of persistent diagram from the data and put it in a CNN. We will compare the performance between CNN and CNN-TDA Net. All will be done on Google Colab.

How and why TDA improves CNN: Transient vs Bogus classifications and gravitational wave detections

Abstract: Topological Data Analysis (TDA) characterizes the global structure of data based on topological invariants, while Convolutional Neural Network (CNN) is capable of characterizing local features. A combined model of CNN-TDA Net, a family of multimodal networks, that takes the input image and the corresponding topological features simultaneously together for classification problems, significantly improves the CNN performance. Although its success has recently been reported in various applications, however, there is a lack of explanation as to why how topological features improve the discriminative power of the original CNN. In this talk, we demonstrate the effects of topological features on a CNN using Transient vs Bogus classification and gravitational wave detection problemswith Grad-CAM analysis. With Grad-CAM analysis on multimodal networks, we demonstrate that adding topological feature enforces a CNN to concentrate on more important regions across images than task-irrelevant artifacts such as background and texture.