2026 workshop program
Modelling Hydrodynamics of Self Gravitating Molecular Clouds using Physics Informed Neural Networks
The work
Abstract
We present a physics-informed neural network (PINN) based model to simulate fragmentation of molecular clouds under self-gravity, a fundamental process in star formation. Simulating such a system requires solving nonlinear, time-dependent hydrodynamic equations that become increasingly expensive at high spatial resolution with traditional methods. In this work, we explore the use of PINNs to study the fragmentation of molecular clouds due to gravitational instability. The instability is triggered by external perturbations, such as shocks from distant supernova explosions, which can lead to the formation of dense collapsing cores that may serve as seeds for star formation. We apply PINNs to learn the dynamics of such systems without relying on simulation data and benchmark against Athena++ simulations. We find that standard multilayer perceptron (MLP)–based PINNs capture the evolution of subsonic perturbations in two dimensions, maintaining errors below 10% over several free-fall times. However, gravitational collapse leads to highly localized, high-density structures that are difficult for standard PINN training to resolve over longer timescales. To address this, we introduce a residual-based adaptive collocation strategy that concentrates training points in high-density regions. This adaptive sampling improves the network’s ability to capture the nonlinear collapse dynamics. As a result, the maximum error is reduced by approximately 27% compared to uniform sampling. Our results demonstrate that simple MLP-based PINNs with adaptive collocation points, trained solely through physical constraints, can effectively model self-gravitating hydrodynamical systems, highlighting their potential as a mesh-free alternative for astrophysical simulations.