Godunov Loss Functions for Modelling of Hyperbolic Conservation Laws

Abstract

Machine learning techniques are being used as an alternative to traditional numerical discretization methods for solving hyperbolic partial differential equations (PDEs) relevant to fluid flow. Whilst numerical methods are higher fidelity, they are computationally expensive. Machine learning methods on the other hand are lower fidelity but can provide significant speed-ups. The emergence of physics-informed neural networks (PINNs) in fluid dynamics has allowed scientists to directly use PDEs for evaluating loss functions. The downfall of this approach is that the differential form of systems is invalid at regions of shock inherent in hyperbolic PDEs such as the compressible Euler equations. To circumvent this problem we propose the Godunov loss function: a loss based on the finite volume method (FVM) that crucially incorporates the flux of Godunov-type methods. These Godunov-type methods are also known as approximate Riemann solvers and evaluate intercell fluxes in an entropy-satisfying and non-oscillatory manner, yielding more physically accurate shocks. Our approach leads to superior performance compared to standard PINNs that use regularized PDE-based losses as well as FVM-based losses, as tested on the 2D Riemann problem in the context of time-stepping and super-resolution reconstruction.

arXiv commentPublished in Computer Methods for Applied Mechanics and Engineering

References in corpus (17)

• Super-resolution reconstruction of turbulent flows with machine learning
• Learning data driven discretizations for partial differential equations
• A Comprehensive Physics-Informed Machine Learning Framework for Predictive Turbulence Modeling
• Physics-Informed Machine Learning Approach for Augmenting Turbulence Models: A Comprehensive Framework
• Nonlinear mode decomposition with convolutional neural networks for fluid dynamics
• Deep Neural Networks for Data-Driven Turbulence Models
• Cluster-based reduced-order modelling of a mixing layer
• PhyCRNet: Physics-informed Convolutional-Recurrent Network for Solving Spatiotemporal PDEs
• JAX-FLUIDS: A fully-differentiable high-order computational fluid dynamics solver for compressible two-phase flows
• Deep Model Predictive Control with Online Learning for Complex Physical Systems
• RiemannONets: Interpretable Neural Operators for Riemann Problems
• Data-Driven Selection of Actuators for Optimal Control of Airfoil Separation
• RoeNets: Predicting Discontinuity of Hyperbolic Systems from Continuous Data
• Physics-Informed CNNs for Super-Resolution of Sparse Observations on Dynamical Systems
• A six-point neuron-based ENO (NENO6) scheme for compressible fluid dynamics
• Deep smoothness WENO scheme for two-dimensional hyperbolic conservation laws: A deep learning approach for learning smoothness indicators
• On the TVD property of second order methods for 2D scalar conservation laws