Optimising Physics-Informed Neural Network Solvers for Turbulence Modelling: A Study on Solver Constraints Against a Data-Driven Approach
Open Access
Abstract
Physics-informed neural networks (PINNs) have emerged as a promising approach for simulating nonlinear physical systems, particularly in the field of fluid dynamics and turbulence modelling. Traditional turbulence models often rely on simplifying assumptions or closed numerical models, which simplify the flow, leading to inaccurate flow predictions or long solve times. This study examines solver constraints in a PINNs solver, aiming to generate an understanding of an optimal PINNs solver with reduced constraints compared with the numerically closed models used in traditional computational fluid dynamics (CFD). PINNs were implemented in a periodic hill flow case and compared with a simple data-driven approach to neural network modelling to show the limitations of a data-driven model on a small dataset (as is common in engineering design). A standard full equation PINNs model with predicted first-order stress terms was compared against reduced-boundary models and reduced-order models, with different levels of assumptions made about the flow to monitor the effect on the flow field predictions. The results in all cases showed good agreement against direct numerical simulation (DNS) data, with only boundary conditions provided for training as in numerical modelling. The efficacy of reduced-order models was shown using a continuity only model to accurately predict the flow fields within 0.147 and 2.6 percentage errors for streamwise and transverse velocities, respectively, and a modified mixing length model was used to show the effect of poor assumptions on the model, including poor convergence at the flow boundaries, despite a reduced solve time compared with a numerically closed equation set. The results agree with contemporary literature, indicating that physics-informed neural networks are a significant improvement in solve time compared with a data-driven approach, with a novel proposition of numerically derived unclosed equation sets being a good representation of a turbulent system. In conclusion, it is shown that numerically unclosed systems can be efficiently solved using reduced-order equation sets, potentially leading to a reduced compute requirement compared with traditional solver methods.Keywords
Funding Information
- Engineering and Physical Sciences Research Council (EP/T517926/1)
This publication has 23 references indexed in Scilit:
- Physics-informed neural networks for solving Reynolds-averaged Navier–Stokes equationsPhysics of Fluids, 2022
- Data-driven modeling and prediction of non-linearizable dynamics via spectral submanifoldsNature Communications, 2022
- Machine learning-augmented turbulence modeling for RANS simulations of massively separated flowsPhysical Review Fluids, 2021
- Prandtl’s extended mixing length model applied to the two-dimensional turbulent classical far wakeProceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 2021
- NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equationsJournal of Computational Physics, 2021
- Comparison and verification of turbulence Reynolds-averaged Navier–Stokes closures to model spatially varied flowsScientific Reports, 2020
- Discovering governing equations from data by sparse identification of nonlinear dynamical systemsProceedings of the National Academy of Sciences, 2016
- High-order compact schemes for incompressible flows: A simple and efficient method with quasi-spectral accuracyJournal of Computational Physics, 2009
- Flow over periodic hills – Numerical and experimental study in a wide range of Reynolds numbersComputers & Fluids, 2009
- Artificial neural networks for solving ordinary and partial differential equationsIEEE Transactions on Neural Networks, 1998