Graph neural network-accelerated Lagrangian fluid simulation

Article Mechanics

Deep learning for reduced order modelling and efficient temporal evolution of fluid simulations

Pranshu Pant et al.

Summary: A novel deep learning framework, DL-ROM, is developed to create a neural network capable of non-linear projections to reduced order states, which efficiently predicts future time steps of a simulation using 3D Autoencoder and 3D U-Net-based architectures. The model can achieve significant computational savings without the need for ground truth supervision or iteratively solving expensive equations, while reducing computational run times of iterative solvers by nearly two orders of magnitude.

PHYSICS OF FLUIDS (2021)

添加到收藏夹

Article Multidisciplinary Sciences

Machine learning-accelerated computational fluid dynamics

Dmitrii Kochkov et al.

Summary: Numerical simulation of fluids is crucial in modeling physical phenomena like weather and aerodynamics, but solving the Navier-Stokes equations at scale remains challenging due to computational costs. This study uses end-to-end deep learning to improve approximations in computational fluid dynamics for modeling two-dimensional turbulent flows, achieving computational speedups and stability while generalizing well to different scenarios. Leveraging machine learning and hardware accelerators, this approach showcases the potential to enhance simulations without sacrificing accuracy or generalization.

PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES OF THE UNITED STATES OF AMERICA (2021)

添加到收藏夹

Article Computer Science, Software Engineering

Neural UpFlow: A Scene Flow Learning Approach to Increase the Apparent Resolution of Particle-Based Liquids

Bruno Roy et al.

Summary: Our novel technique uses deep neural networks for high-resolution liquid generation based on scene flow estimation, inferring and synthesizing details from low-resolution simulations while encoding liquid properties through neighborhood contributions. The approach also includes particle-based interpolation and a key-event topological alignment constraint for varying discretizations, allowing the inference model to reward important differences in simulation setups.

PROCEEDINGS OF THE ACM ON COMPUTER GRAPHICS AND INTERACTIVE TECHNIQUES (2021)

添加到收藏夹

Article Computer Science, Software Engineering