Gravitational-Wave Anomalous Knowledge with Recurrent Autoencoders.
The goal of this project is to leverage machine learning for the detection of gravitational waves originating from sources other than compact binary coalescences. Motivated by the current dominance of compact binary coalescence searches within the algorithmic landscape, we seek to develop robust algorithms capable of identifying signals from more challenging or unknown sources, aiming to unveil novel phenomena and contribute to our understanding of new physics. GWAK uses a deep recurrent autoencoder-based model specifically designed for detecting non-compact binary coalescence models. Adopting a semi-supervised approach enables the exploration of sources that defy modeling, while also providing valuable information on the resemblance of signals to known sources. The added advantage of fast inference not only facilitates efficient detection but also opens avenues for multi-messenger astronomy, enhancing the project’s significance in the broader context of gravitational wave research.

For more information, please see here.

Gravitational-wave inference as a service prototype.
The goal of this project is to establish a versatile platform capable of deploying real-time machine learning algorithms within the realm of gravitational-wave science. Motivated by the paradigm of inference as a service, we aim to harness existing computing resources efficiently, ensuring real-time capabilities. We have demonstrated sub-second latencies through the implementation of distributed inference in both noise cleaning and compact binary coalescence detection pipelines, specifically tailored for gravitational-wave computing resources. This addresses not only the need for rapid analysis in gravitational-wave science but also signifies a pivotal advancement in the field by showcasing the feasibility of deploying machine learning algorithms with remarkable speed and efficiency.

For more information, please see here.

The Zwicky Transient Facility Source Classification Project
Understanding the ever-changing nature of celestial objects, from stars to galactic centers, is crucial for astrophysics. The Zwicky Transient Facility (ZTF) uses time series observations to track the variability of over a billion sources. Given the immense scale of this data, we employ automated methods to thoroughly analyze it. This project uses neural network and XGBoost machine learning algorithms to categorize ZTF sources into different classes, based on a carefully crafted training set. The classification catalog resulting from this work is accessible to the public, and the software is open-source, ready to adapt for future time-domain surveys.

For more information, please see here.

aframe: A machine-learning pipeline for real-time detection of gravitational waves from compact binary coalescences

Swiftly detecting gravitational waves is essential to enable timely follow-up observations by telescopes. Neural networks have showcased their ability to understand complex patterns and make rapid decisions with minimal computational resources, making them an appealing choice for detecting gravitational waves. However, integrating these advanced algorithms into the field of gravitational-wave astrophysics has proven to be a challenging task. In this project, we introduce the first-ever fully machine-learning based system designed to swiftly detect gravitational waves from compact binary coalescences with low latency. Our innovative pipeline not only demonstrates near state-of-the-art sensitivity but also achieves this at a fraction of the time and computational cost required by traditional methods like matched filtering pipelines.

Acceleration of ML Inputs for IceCube: IceCube operates in an extreme environment with limited resources, i.e. power. Calibrating and feature extracting the individual channels consumes a significant fraction of the compute resources at the South Pole and is required to generate inputs for the current generations of IceCube machine learning-based algorithms. Accelerating these inputs will allow IceCube to reduce the power requirements for the computing infrastructure and overall time to result.