AI in Science
My approach to AI in science is data-centric and checkable. I use physics to improve the data that AI learns from, so that models learn the sky rather than the instrument, and I hold everything AI helps me produce to one standard: however much AI helped, my deliverables are white boxes. My work on AI as a focus began with my Schmidt AI in Science Fellowship at MIDAS, after earlier papers that already used machine learning (Cao and Pinsonneault 2025; Cao 2026b), and most of it lies ahead.
- published scoring rubric
- open code
- public daily tables
- every score labeled AI-generated
- scores are pointers, not judgments
Where I stand
My part in AI is at its two ends: rather than developing large models myself, I make both the data that go into them and the results that come out checkable.
In precision science, systematic errors are an ethical matter: a biased pipeline can lead a whole community to the wrong conclusions about the Universe. So I prefer methods I can interpret. I use AI to extend physical understanding, not to replace physical models with opaque networks. The best check against error is scrutiny by the community, and automation should never outpace the scientific method.
As data and computation stop being what limits progress in astronomy, and AI takes over more of the routine work, what stays scarce is innovation and judgment: seeing a question or a method nobody has tried, knowing what a method assumes and how it fails, and turning a result into physical understanding. That is where I want to spend my own effort.
- AI-Ready Data: Using physics to improve the data that AI learns from, starting with image reconstruction.
- Working with AI: How I use AI in my daily work: reading the literature, writing papers, teaching, acknowledging AI, and applying for jobs.
- MapHub: Daily relevance scores of new arXiv preprints for a research community, open and checkable; the first deliverable of my fellowship.
Where AI enters each of my research directions is on Research Directions. New to the topic? The Glossary explains the terms in plain language.
Terms on this page (2)
- MIDAS
- The Michigan Institute for Data and AI in Society, the University of Michigan's institute for data science and AI, which hosts the Schmidt AI in Science postdoctoral fellows there.
- Systematics, systematic errors
- Errors that push results in a consistent direction, instead of scattering them randomly. More data do not average them away, so they must be understood and corrected. Uncorrected noise patterns, such as stripes, can become systematics when they mimic the signal.