Wednesday 12 March 2025
The quest for a more efficient way to model rarefied gas dynamics has been an ongoing challenge in the field of fluid mechanics. Rarefied gases are those that exist at extremely low pressures, where the distance between molecules is significant compared to their size. Understanding how these gases behave is crucial for optimizing flows through microelectromechanical systems, air filtration devices, and shale gas extraction.
Traditionally, researchers have relied on numerical methods like discrete velocity and direct simulation Monte Carlo (DSMC) to simulate rarefied gas dynamics. However, these approaches demand intense computational resources and memory, making them impractical for complex geometries or large-scale simulations.
Enter physics-informed neural networks (PINNs), a meshless and adaptable alternative that has been gaining traction in recent years. By incorporating continuity and Cauchy momentum exchange equations into the loss function, PINNs can learn to solve non-linear partial differential equations (PDEs) with unprecedented accuracy.
In a new study, researchers have applied PINNs to rarefied gas dynamics, leveraging their ability to model complex flows in microflows. The team trained a PINN using a limited number of DSMC-generated rarefied gas microflows in the transition regime (0.1 < Kn < 3), achieving under 2% error on these residuals and effectively filtering out DSMC's inherent statistical noise.
The results are impressive, with the PINN showing strong predictive capabilities for tested flow fields at a range of Knudsen numbers (Kn). Notably, each DSMC simulation required approximately 20 hours on four graphics processing units (GPUs), while the PINN training took under two hours on a single GPU, with evaluations taking mere seconds.
The implications are significant. By leveraging PINNs to model rarefied gas dynamics, researchers can now tackle complex geometries and large-scale simulations more efficiently than ever before. This has far-reaching potential for optimizing flows in microelectromechanical systems, air filtration devices, and shale gas extraction, as well as other applications where rarefied gases play a crucial role.
The study’s findings also highlight the potential of PINNs to bridge the gap between numerical methods and physical phenomena. By incorporating physical constraints into the neural network architecture, researchers can create models that not only predict complex flows but also provide insights into the underlying physics driving those flows.
As the field continues to evolve, it will be exciting to see how PINNs are applied to an increasingly diverse range of problems.
Cite this article: “Physics-Informed Neural Networks Revolutionize Modeling of Rarefied Gas Dynamics”, The Science Archive, 2025.
Rarefied Gas Dynamics, Physics-Informed Neural Networks, Microflows, Transition Regime, Knudsen Numbers, Discrete Velocity Method, Direct Simulation Monte Carlo, Non-Linear Partial Differential Equations, Meshless Methods, Computational Fluid Dynamics.
