Research Projects

The physical processes driving galaxy formation span an extraordinary scale range—from ~100 Mpc (where the universe is homogeneous) to ~10 AU (the Schwarzschild radius of supermassive black holes). Even cutting-edge cosmological simulations, with adaptive refinement, cover only ~8 orders of magnitude, leaving crucial processes like star formation, black hole accretion, and feedback heating under-resolved. These are typically incorporated via phenomenological subgrid models calibrated mainly on stellar observables, which limits predictive power, especially for multiphase circumgalactic medium (CGM) properties revealed by UV, X-ray, SZ, and fast radio burst observations. At the heart of this problem lie radiative boundary layers (RBLs) between cool, dense clouds and hot, diffuse halos. RBLs regulate mixing, cooling, and the fate of baryons—whether they condense to form stars or are driven into outflows. Resolving them in cosmological simulations would demand ~0.1 pc resolution, far beyond the ~100 pc achievable today. This proposal pioneers a new approach: replacing traditional subgrid recipes with physics-anchored machine learning (ML) surrogates trained on high-resolution simulations of turbulent radiative mixing layers (TRMLs), idealized setups that capture RBL physics. Scientific Objectives: Generate high-resolution (up to 4096² in 2D, 1024³ in 3D) TRML simulations with AthenaK, spanning shear velocities, density contrasts, metallicities, and radiation fields. Train convolutional neural networks (CNNs) and symmetry-preserving variants on these datasets to learn subgrid fluxes and source terms (mass, momentum, energy, passive scalars) driving the coarse-grained evolution of multiphase gas. Validate these ML surrogates by embedding them into low-resolution hydrodynamic simulations, checking their ability to reproduce TRML physics. Integrate trained ML models into the SIMBA cosmological code, replacing conventional treatments of RBL cooling and mixing, and assess impacts on galaxy and CGM properties across cosmic time. Hypothesis:
We hypothesize that physics-informed ML surrogates trained on well-resolved TRMLs can accurately emulate unresolved subgrid RBL physics in galaxy simulations, offering a robust alternative to empirical models. This should yield more predictive simulations of multiphase CGM and star formation, directly tied to first-principles hydrodynamics rather than calibration. Methodology & Experiments: TRML Simulations: Use AthenaK to run 2D/3D simulations across a broad physical parameter space, generating time-resolved datasets that capture turbulent steady states of radiative mixing. ML Development: Train CNN-based models to map coarse flow variables (density, velocity, temperature, gradients) to subgrid fluxes and source terms, embedding conservation laws and symmetries via custom loss functions. Validation in Idealized Flows: Integrate ML modules back into AthenaK to test them in coarse-resolution TRMLs and radiative cloud-crushing simulations, benchmarking against high-resolution results. Cosmological Embedding: Implement ML closures within the SIMBA (GIZMO-based) galaxy formation framework, validating with fluid-mixing tests, cosmological cube runs, and zoom-ins on Milky Way-mass and group halos. Significance:
If successful, this work will transform galaxy simulation methodology by replacing empirically tuned subgrid prescriptions with ML surrogates grounded in high-resolution physics. It promises significantly improved predictions of CGM structure, star formation rates, and black hole growth—directly relevant for interpreting early galaxies unveiled by JWST. The public TRML datasets and open ML modules will also become foundational community resources, democratizing rigorous subgrid modeling. Drawing from advances in climate and CFD, this project pioneers the application of physics-driven ML to cosmic hydrodynamics, opening a new era of predictive, cross-scale galaxy simulations.