AI-Ready Data

In brief

Machine learning is now widely used to analyze astronomical images. But a model learns whatever is in its training data, including the telescope's imperfections: a blurry or uneven point spread function, correlated noise, artifacts from undersampling. A model trained on such data may learn the instrument rather than the sky.

My approach is data-centric. Rather than building ever more complex models to make up for imperfect data, I use physics to improve the data first. Image reconstruction is where this starts: combining a telescope's undersampled exposures into well-sampled images whose point spread function is as close as possible to a chosen target, so that every output image comes with a known PSF, as IMCOM and Effortless do. In statistical terms, it is a linear inverse problem (Image Reconstruction).

In more depth

Much of my work is on inverse problems: reconstructing well-sampled images from a telescope's undersampled exposures is one, and Effortless, my algorithm for it, is built to be both scalable and interpretable. I validate such methods against simulated truth and measure how well each output meets its target, which is the discipline that uncertainty quantification asks of machine learning too.

What I have done

I led the refactoring of IMCOM into PyIMCOM, which reduced its computing cost by about an order of magnitude; teammates ran PyIMCOM to make the IMCOM coadds of the OpenUniverse2024 simulations, used in several team papers. I then developed Effortless, my algorithm for linear image reconstruction, which is tens of times faster than IMCOM, with a smaller memory footprint, and interpretable: where IMCOM's large matrix inversion is largely a black box, Effortless designs its weights in steps that can each be understood. My papers also use simple, checkable models from machine learning, for example a support vector regression that emulates the asteroseismic correction of APOKASC-3 (Cao and Pinsonneault 2025), and a robust regression that calibrates Effortless's measurements against injected stars (Cao 2026b). My formal training in data science includes the Erdős Institute Data Science Boot Camp.

Papers

The papers behind this work, from IMCOM to Effortless, are on Image Reconstruction; all my papers, with what I contributed to each, are on Publications.

What comes next

One goal is image products ready for machine learning: reconstructed images whose PSFs and noise are documented. I hope AI will help me reach the goals of my research on images, cosmology and stars, while everything I deliver stays checkable. Where AI enters each direction is on my Research Directions.