me

ChangHoon Hahn

My research interests include astrostatistics and ML, galaxy evolution, and cosmology.

Simulation-Based Inference


Probabilistic inference is a pivotal step in any scientific experiment. It provides the statistical framework for comparing theoretical models that encapsulate our knowledge of physics processes to observations and derive constraints on the parameters of interest — i.e. physics. In cosmology, for instance, we use inference to measure the contents of the Universe from 3D maps of galaxies. In galaxy evolution, inference allows us to measure the physical properties of galaxies from their spectra.

A new class of methods in simulation-based inference (SBI; also known as "likelihood-free inference") is transforming probabilistic inference. SBI enables rigorous inference using only simulations that describe the observations and the physical phenomena of interest. Cutting-edge SBI methods exploit neural density estimation and can efficiently infer high-dimensional posteriors, the probability distribution of parameters given observations. SBI relaxes the assumptions in standard approaches for more accurate and rigorous inference. It also enables us to leverage the predictive power of high-fidelity simulations to more tightly constrain physics.

I am a leading expert on SBI. I pioneered SBI in large-scale structure [1607.01782] and galaxy evolution studies. I have also developed new SBI methods [1803.06348]. Last semester (Fall 2021), I co-developed and instructed a graduate course on SBI at Princeton. I also organized workshops on SBI at the Flatiron Institute and at Berkeley Institute for Data Science that brought together SBI experts from across disciplines to share the latest techniques, use cases and applications, and to discuss open challenges. Much of my current research focuses on developing new SBI methods and exploiting them to maximize the scientific return of upcoming galaxy surveys, starting with DESI.

ML for Big Data Astronomy


With upcoming observations from DESI, PFS, JWST, Rubin, and Roman, astronomy is entering the era of Big Data. ML techniques offer new ways to tackle the upcoming massive data sets. One key application of ML is for accelerating our models and simulations: e.g. in [1910.04255] we used ML to accurately construct N-body simulations with massive neutrinos that originally takes 700 CPU hours in just 4 mins. Similarly, in [1911.11778] we used neural emulators to make SED models >1000x times faster.

ML also provides essential tools to construct scalable analyses through SBI. Certain SBI methods enable amortized inference by using neural density estimate of the posterior over the full conditional space of observations. Once trained, they can be applied to data to infer the posterior from each observation in seconds, effectively eliminating the computational cost of inference. Using SBI, we can deploy rigorous and sophisticated analyses to the billions of galaxies that will be observed by the next-generation experiments. For an example, check out my interactive slides from a recent conference.

I am also developing unsupervised deep generative models for generating realistic simulations. In particular, I am building generative models that incorporate our knowledge of physics. For example, most of the light in galaxy spectra comes from stellar populations that can be modeled using stellar population synthesis (SPS). I am currently developing a generative model for galaxy spectra that incorporates SPS within its architecture. It also includes physical transformations such as redshifting within its causal structure. This physics-driven generative model produces galaxy spectra that are more realistic and orders of magnitude faster than current models. More broadly, these ML generative models will be crucial to meet the simulation demands of upcoming surveys. For more details, check out my interactive slides from a recent conference.