KAAI Program I, Summer 2026
KAAI has gotten off to a really exciting start!
, KAAI's inaugural Fellow and a postdoctoral researcher at Columbia University, studies the growing connections between AI, astrophysics, and cosmology.
He develops machine learning models that leverage both observational and simulated datasets for scientific inference and emulation. His research spans a wide range of topics, including galaxy clustering, galaxy formation, galaxy clusters, and dust attenuation in cosmology. He is particularly interested in building robust and trustworthy AI for scientific discovery, combining Bayesian methods and explainable AI techniques to improve the reliability and interpretability of complex models.
KAAI 2026 Cycle, Project 1
✦ Foundation Models for Astrophysics, Agentic Systems and Virtual Universes
KAAI Fellow: Matt Ho
Students: Anshul Kumar, Xiaowen Zhang, Yixi Zhao, Chaipat Tirapomprasert, Fatemeh Hafezianzadeh
Faculty: , Rupert Croft and Tiziana Di Matteo
This project asks whether the same kind of large-scale AI that has transformed language and images can also learn the physics of the universe. Modern cosmology rests on two pillars: "virtual universes", enormous computer simulations that grow galaxies, black holes and the cosmic web forward from the Big Bang, and surveys that photograph billions of real galaxies. Our goal is to build foundation models that speak both languages at once, so that a simulated galaxy and an observed one can be compared, matched and understood within a single shared picture, and to develop agentic systems: AI that does not merely analyse data but runs, steers and calibrates the simulations themselves. Over the summer we have been probing what a galaxy foundation model has actually learned, finding that its internal map spontaneously rediscovers the classical Hubble sequence of galaxy shapes and separates out features such as spiral bars. We have built models that translate between different kinds of cosmological data and recover the parameters of the universe with high accuracy across several independent simulation codes, taught networks to sharpen coarse simulations into fine structural detail across a wide range of scales, developed AI agents that tune simulations automatically by proposing parameters, running jobs and keeping their own laboratory notebook, and generated realistic mock survey images so that theory and observation can be set side by side on equal terms.
Cosmological parameters recovered from simulated galaxy catalogues across four independent simulation codes
A coarse simulation and the fine structure recovered from it
The agentic simulator loop, in which an AI agent reads a task description, proposes parameters, submits jobs and records what it has learned.
A foundation model's internal map of galaxies, in which the axis from ellipticals to disks and the presence of a bar emerge on their own
Nested simulation volumes spanning a range of cosmological scales
The group works closely and informally, with daily KAAI coffee meetings and hack sessions keeping ideas moving between the astronomy and machine learning sides. KAAI also hosts workshops and hackathons that bring together dozens of participants, with registered teams drawn from many departments across 91视频 and the University of Pittsburgh.