Upcoming Event: NSF-Simons CosmicAI Hybrid Seminar Series

1) From astrochemical models to surrogate models: understanding the chemistry of star formation, 2) Adaptive Tree Reduced-Order Model for Parametric Chemical Kinetics, 3) Tree-based adaptive approximation

Mélisse Bonfand-Caldeira, Arjun Vijaywargiya and George Biros, Speaker 1: Postdoc at the University of Virginia Speaker 2: Postdoctoral Fellow at the Oden Institute, University of Texas at Austin Speaker 3: W. A. ``Tex'' Moncrief Chair in Simulation-Based Engineering Sciences at the Oden Institute, University of Texas at Austin

Abstract

Speaker 1: Astrochemical models are powerful tools for tracing the chemical evolution of star-forming regions, where atoms and simple molecules transform into more complex species under the influence of diverse physical processes. Coupled with realistic, time-dependent physical models, they provide a detailed view of the intertwined physical and chemical evolution from cold pre-stellar cores to protostars and their surrounding disks, helping us understand how chemistry shapes the environments in which planets form. However, the chemical networks used in these simulations can contain thousands of species and hundreds of thousands of reactions and processes, making astrochemical models computationally expensive and limiting their application across large parameter spaces and complex physical models. In the Accelerated Universe group, we are developing surrogate models that aim to capture the complex chemical evolution at minimum computational cost, enabling fast predictions across a broad range of physical conditions. Part 1: I will introduce a benchmark of time-dependent astrochemical simulations performed across a grid of physical models, sampling gas densities, temperatures, and radiation fields representative of turbulent molecular clouds and dense cores, using physical conditions extracted from 3D hydro- and -magnetohydrodynamic simulations of star formation. The chemistry is computed using three chemical networks of increasing complexity, providing datasets with different levels of fidelity for training and validating surrogate models.

Speaker 2: I will present an adaptive clustering-based reduced-order model (ROM) for astrochemical kinetics, which conserves the network's linear invariants, enforces positivity at interval boundaries, and optionally falls back to the high-fidelity solver when the active latent space inadequately represents the state. The ROM reduces simulation cost by evolving a low-dimensional latent representation from which dominant solution features can be extracted. Instead of a single global latent space, the method clusters snapshots and constructs a local latent space for each cluster. In two chemical networks tests our method accurately reproduces target abundances with roughly 3-, 5-fold dimension reduction.

Speaker 3: Surrogate models for expensive deterministic simulators are commonly optimized for average error, small average error can often conceal poor localized learning. I will discuss adaptive trees that can learn spatial partitioning and piecewise approximations that can provide provable approximation errors. I will also discuss high-performance computing algorithms for the efficient construction of such trees and using them as surrogates for dynamical systems.

Biography

Speaker 1: Mélisse Bonfand-Caldeira: I am a postdoctoral researcher in the Departments of Chemistry and Astronomy at the University of Virginia. My research combines observations from state-of-the-art radio telescopes with numerical simulations to investigate how complex organic molecules form in star-forming regions. I obtained my Ph.D. from the Max Planck Institute for Radio Astronomy in 2019. I then worked at the Laboratoire d’Astrophysique de Bordeaux in France before joining UVA four years ago.

Speaker 2: Arjun Vijaywargiya is a Peter O’Donnell Jr. Postdoctoral Fellow at the Oden Institute at the University of Texas at Austin, and concurrently a Cosmic AI Fellow at the NSF–Simons AI Institute for the Study of Cosmic Origins. He works with Prof. George Biros and Prof. Stella Offner on surrogate modeling of astrochemical networks.

Speaker 3: George Biros is the W. A. ``Tex'' Moncrief Chair in Simulation-Based Engineering Sciences in the Oden Institute for Computational Engineering and Sciences and has Full Professor appointments with the departments of Mechanical Engineering and Computer Science (by courtesy) at The University of Texas at Austin. With collaborators, he received the ACM Gordon Bell Prize in 2003 and in 2010. He is a 2023 SIAM Fellow.