2026 workshop program
Neural-network-based forward modeling of accretion disks
The work
Abstract
Accretion disks are ubiquitous in astrophysics, from planet-forming disks around young stars to disks around compact objects and supermassive black holes. Modeling their dynamics is essential for interpreting observations, but traditional simulations are computationally expensive, limiting large parameter surveys and inverse modeling.
We develop machine-learning tools with the goal to replace numerical simulations of accretion disks, using protoplanetary disks as an example. In these systems, embedded planets perturb the surrounding gas and dust, allowing planet properties to be inferred from disk structures. We use a data-driven Deep Operator Network to map physical parameters directly to hydrodynamic solutions, enabling rapid forward modeling and efficient inverse inference (Mao et al. 2023, ApJL, 950, 12). We also explore physics-informed neural networks (PINN) for solving time-dependent, two-dimensional compressible Navier-Stokes equations in rotating accretion flows (Mao et al. 2025, ApJL, 992, 20). Standard PINNs face challenges in convergence, long-time accuracy, and capturing high-spatial-frequency structures. By introducing adaptive activation functions, time-marching strategies, and network inheritance, we improve their efficiency and accuracy without requiring labeled simulation data. This work demonstrates the potential of operator learning and PINN for accretion disks and broader astrophysical fluid systems.
Our group at Peking University in Beijing is always looking for graduate students and postdocs interested in developing AI-based methods to accelerate, emulate, and ultimately rethink numerical simulations in astrophysics. Please get in touch at rbdong@pku.edu.cn. https://www.ruobingdong.com