1 August 2022 in Physics of Fluids
The simulation of fluid dynamics, typically by numerically solving partial differential equations, is an essential tool in many areas of science and engineering. However, the high computational cost can limit application in practice and may prohibit exploring large parameter spaces. Recent deep-learning approaches have demonstrated the potential to yield surrogate models for the simulation of fluid dynamics. While such models exhibit lower accuracy in comparison, their low runtime makes them appealing for design-space exploration. We introduce two novel graph neural network (GNN) models, multi-scale (MuS)-GNN and rotation-equivariant (RE) MuS-GNN, for extrapolating the time evolution of the fluid flow. In both models, previous states are processed through multiple coarsening of the graph, which enables faster information propagation through the network and improves the capture and forecast of the system state, particularly in problems encompassing phenomena spanning a range of length scales. Additionally, REMuS-GNN is architecturally equivariant to rotations, which allows the network to learn the underlying physics more efficiently, leading to improved accuracy and generalization. We analyze these models using two canonical fluid models: advection and incompressible fluid dynamics. Our results show that the proposed GNN models can generalize from uniform advection fields to high-gradient fields on complex domains. The multi-scale graph architecture allows for inference of incompressible Navier–Stokes solutions, within a range of Reynolds numbers and design parameters, more effectively than a baseline single-scale GNN. Simulations obtained with MuS-GNN and REMuS-GNN are between two and four orders of magnitude faster than the numerical solutions on which they were trained.
19 July 2023 in Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences
Over the last decade, deep learning (DL), a branch of machine learning, has experienced rapid progress. Powerful tools for tasks that have been traditionally complex to automate have been developed, such as image synthesis and natural language processing. In the context of simulating fluid dynamics, this has led to a series of novel DL methods for replacing or augmenting conventional numerical solvers. We broadly classify these methods into physics- and data-driven methods. Physics-driven methods, generally, tune a DL model to provide an analytical and differentiable solution to a given fluid dynamics problem by minimizing the residuals of the governing partial differential equations. Data-driven methods provide a fast and approximate solution to any fluid dynamics problem that shares some physical properties with the observations used when tuning the DL model’s parameters. Meanwhile, the symbiosis of numerical solvers and DL has led to promising results in turbulence modelling and accelerating iterative solvers. However, these methods present some challenges. Exclusively data-driven flow simulators often suffer from poor extrapolation, error accumulation in time-dependent simulations, as well as difficulties in training against turbulent flows. Substantial effort is, therefore, being invested into approaches that may improve the current state of the art.
1 August 2025 in Physics of Fluids
Rotational equivariance arises in physical problems as a common symmetry of partial differential equations, including the Navier–Stokes equations governing fluid phenomena. Architectural changes are necessary when guaranteeing rotational equivariance in neural networks, which incur additional computational costs, leading to increased inference times by up to an order of magnitude, as measured in our numerical studies. We introduce a new method for rotational equivariance in graph neural networks that achieves high predictive accuracy while maintaining a smaller computational footprint than comparable approaches. We establish rotational equivariance by transforming vector features and spatial neighborhood information into local, node-specific vector bases. The resulting architecture follows an encode-process-decode paradigm. Vector features are transformed before being encoded. When message-passing, i.e., in the process stage, latent features are interpreted as a set of vectors and undergo a similar transformation and are thus always processed in the receivers' basis. The results are transformed back into the reference system after decoding, resulting in rotational equivariance. The networks receive full physical and geometric information of the neighborhood without costly computations. Three different scenarios are experimentally investigated, ranging from an advection problem to incompressible Navier–Stokes flow around an ellipse and to a highly complex transonic cylinder flow scenario. We compare with state-of-the-art approaches and demonstrate how our method achieves comparable accuracy while reducing computational costs. Our method is the only one to consistently improve upon a data-augmentation baseline, and does so with an error reduction of 25.3%. It is 1.6 times faster than the next best model that guarantees equivariance.
1 November 2013 in Clinical Neurophysiology