SkAI-Funded Research Projects: Year 1 (2024-25)

Lead PIs: Noelle Samia (NU), Gautham Narayan (UIUC)
Co-PIs: Adam Miller (NU), Daniel Apley (NU), Matt Krafczyk (UIUC)
Collaborators: John O’Brien (UIUC), Amanda Wasserman (UIUC), Jiezhong Wu (NU)
Project Summary: This project aims to build a foundational AI model that would form a core part of a community pipeline for Rubin to identify (including anomalies) and characterize transients and set priorities for follow-up observations. The foundation model will be based on a new custom variational autoencoder (VAE) architecture, optimized for the sparse, multimodal and irregular data, and both observational data and emulated data from (physical) transient models. In year 2, the VAE will be replaced by a transformer.

Lead PIs: Adam Miller (NU), Han Liu (NU)
Co-PIs: Noelle Samia (NU), Gautham Narayan (UIUC)
Collaborators: Nabeel Rehemtulla (NU), Ved Shah (NU), Hong-Yu Chen (NU), Qinjie Lin (NU), Weijian Li (NU)
Project Summary: This project seeks to predict nonuniformly sampled time-series data using artificial intelligence with uncertainty quantification. The main goal is astronomical time-series data, mainly from stars within the Zwicky Transient Facility (ZTF) survey, but ultimately applicable to Rubin/Legacy Survey of Space and Time (LSST) and even other domains.

Lead PIs: Emma Alexander (NU), Tjitske Starkenburg (NU)
Co-PIs: Aggelos Katsaggelos (NU), Alex Drlica-Wagner (UC), Joshua Frieman (UC), Chihway Chang (UC), Rebecca Willett (UC), Michael Zevin (Adler)
Collaborator: Travis Rector (U Alaska Anchorage)
Project Summary: This project has the following objectives: (1) remove distortions from astrophysical images by modeling the image formation process, (2) initially focus on simpler distortion models (convolution, noising), and (3) subsequently, develop a system to handle gravitational lensing. These objectives will be met using (1) a differentiable image synthesis pipeline comprising three components: (a) implicit image representations (e.g., SIREN) to act as per-image prior, (b) conditional generative model (normalizing flows or diffusion) to capture data distribution, and (c) differentiable nonlinear forward model to capture physics-based distortions (e.g., lensing); and (2) uncertainty quantification.

Lead PIs: Joaquin Vieira (UIUC), William Moses (UIUC)
Co-PIs: Michael Zevin (Adler), Laura Trouille (Adler), Tom Crawford (UC)
Collaborators: Kedar Phadke (UIUC), Vimarsh Sathia (UIUC), Melanie Archipley (UC), Sydnee O’Donnell (UIUC), Siyuan Brant Qian (UIUC), Jennifer Li (UIUC), Felipe Menanteau (UIUC)
Project Summary: This project aims to develop an automated Bayesian inference tool for millimeter source association [cross-matching South Pole Telescope (SPT) sources with other surveys]. Correct associations to observations in different wavelengths are necessary to reconstruct spectral energy distributions (SEDs) and infer properties and redshifts of astrophysical objects. Bayesian inference is one of the methods to do this, but due to large numbers of objects that will be available from the current and future surveys, faster methods are necessary. This project aims to use machine learning (ML) to automate Bayesian inference, focusing on advancements in differentiable and probabilistic programming. Unlike existing differential programming problems, this project will incorporate the use of discrete variables, which are not handled well by existing automatic differentiation (AD) frameworks. The team plans to build a Zooniverse project to confirm source associations via visual inspection by volunteers.

Lead PIs: Daniel Holz (UC), Rina Foygel Barber (UC)
Project Summary: This project is on coincidence detection for time-domain multiwavelength and multimessenger data. Specifically, the project includes testing assumptions in current coincidence-detection techniques and describing uncertainty quantification for this method. Additionally, options for more robust detection techniques will be explored.

Lead PIs: Rebecca Willett (UC), Chihway Chang (UC) 
Co-PIs: Joshua Frieman (UC), Moritz Münchmeyer (UW–Madison), Daniel Sanz-Alonso (UC)
Collaborators: Yuuki Omori (UC), Shi-Hui Zang (UW–Madison)
Project Summary: This project aims to use AI tools to reconstruct the initial matter density field, given measured observations (e.g., galaxy counts).

Lead PIs: Ermin Wei (NU), Jeff McMahon (UC)
Co-PI: Yuxin Chen (UC)
Project Summary: This project targets AI-informed astronomical instrument design for millimeter-wave spectrometers. The project is broken into two tasks: Task 1 seeks to leverage deep neural networks to more expansively explore the space of demultiplexer spectrometer designs by building a mapping between topology and performance; Task 2 seeks to leverage multifidelity simulations (deep kernel Bayesian optimization) to efficiently explore the design space. Potential impacts would be in the design of new astronomical instrumentation and the AI tools for optimizing instrument design.

Lead PIs: Aravindan Vijayaraghavan (NU), Alex Drlica-Wagner (UC)
Collaborators: Alexander Ji (UC), Adam Miller (NU), Aleksandra Ćiprijanović (Fermilab), Brian Nord (Fermilab), Emma Alexander (NU)
Project Summary: This project has the following objectives: (1) develop reinforcement-learning (RL) algorithms for dynamic, real-time astronomical observation scheduling with deferred, multi-objective rewards; and (2) use active learning and weak-to-strong supervision to optimize target selection for population-level parameter inference. These objectives will be met by (1) leveraging state-of-the-art AI techniques, including RL, active learning, and conformal prediction; and (2) using ∼500k observations from the Dark Energy Camera (DECam) for model training and validation.

Lead PIs: Aleksandra Ćiprijanović (Fermilab), Sandeep Madireddy (Argonne National Laboratory)
Co-PIs: Alexander Ji (UC), Allison Strom (NU), Tjitske Starkenburg (NU), Michael Zevin (Adler), Michael Maire (UC), Aggelos Katsaggelos (NU)
Collaborator: Gourd Khullar (U Washington)
Project Summary: This project aims to develop AI methods for studying the evolution of galaxies, including generative models to create synthetic data and inference tools for learning galaxy properties.