Learning Fluid Simulation with PyFR

The sequence used throughout the document is:

PyFR is an open-source CFD solver that uses a high-order Flux Reconstruction method on unstructured meshes. Its present documented governing systems are compressible Euler, compressible Navier–Stokes, artificial-compressibility Euler, and artificial-compressibility Navier–Stokes. These cover inviscid and viscous gas or liquid flow, including unsteady and turbulent flow problems.

1.2 Why PyFR is distinctive

• High-order accuracy: each mesh element can represent a curved and smoothly varying solution with a polynomial rather than a single cell-average value.

• Scale-resolving focus: it is particularly suited to unsteady turbulent structures, wakes, vortices, and flow features for which excessive numerical dissipation is undesirable.

• Unstructured mixed-element meshes: useful for complex engineering geometries.

• Cross-platform execution: current documentation describes CPU and multiple accelerator backends, including CUDA, HIP, OpenCL, Metal, and OpenMP.

• Parallel operation: meshes can be partitioned and cases can run across multiple processes/devices.

1.3 What is directly in scope

1.4 Areas that require caution or another solver

PyFR is primarily a fluid-flow solver. The following are not standard out-of-the-box workflows in the current user documentation and should not be promised as ordinary PyFR capabilities without a specific extension, coupling strategy, or development effort:

• Structural stress, deformation, fatigue, or crash mechanics.

• Electromagnetics, RF, antennas, quantum-device fields, or circuit simulation.

• Full conjugate heat transfer through both fluids and solid structures.

• Multiphase free surfaces, droplets, cavitation, or particle-laden flow as standard models.

• Detailed combustion chemistry and reacting species transport as a standard model.

• Complete meteorology with atmospheric stratification, humidity, radiation, and weather physics.

• True moving/sliding meshes for complete rotor–stator or multi-rotor motion unless a verified PyFR-specific method or extension is available.

2.1 Simple concepts: learn to describe the flow

2.2 Core concepts: understand why the flow behaves that way

• Conservation of mass: fluid mass cannot appear or disappear inside the domain unless a source is deliberately introduced.

• Conservation of momentum: pressure, viscosity, and body forces change the fluid’s motion.

• Conservation of energy: important for compressible flows where temperature, pressure, density, and speed are coupled.

• Inviscid versus viscous: Euler equations omit viscous stresses; Navier–Stokes equations include them.

• Compressible versus incompressible: density variations are retained in compressible modelling and suppressed through an incompressible formulation when appropriate.

• Initial conditions: the flow state at simulation start.

• Boundary conditions: rules applied at inlets, outlets, walls, far-field boundaries, symmetry planes, and periodic pairs.

• Mesh: the geometric partition on which the solution is represented.

• Time step: the increment through which an unsteady solution advances.

2.3 Medium concepts: learn the organizing parameters

A small set of nondimensional numbers organizes a large fraction of engineering CFD:

• Reynolds number, Re = ρUL/μ: compares inertial effects with viscous effects. Higher values usually produce thinner boundary layers, stronger separation sensitivity, and more turbulent scales.

• Mach number, Ma = U/a: compares flow speed with the speed of sound. It indicates when compressibility, wave propagation, and shocks may matter.

• Strouhal number, St = fL/U: describes the nondimensional frequency of periodic unsteadiness such as vortex shedding.

• Pressure coefficient, Cp: presents surface pressure in a normalized form useful for comparing geometries and speeds.

• Lift and drag coefficients, CL and CD: normalize aerodynamic forces so designs can be compared.

2.4 Advanced concepts: understand high-fidelity simulation

• Spatial order (p-order): the polynomial order used inside each element. Raising order can increase accuracy without changing the element count, but increases cost and may expose mesh-quality problems.

• Flux Reconstruction: a high-order discretization framework that reconstructs fluxes and corrects them to communicate consistently across element interfaces.

• Direct Numerical Simulation (DNS): resolves all dynamically relevant turbulence scales; extremely expensive at engineering Reynolds numbers.

• Large-Eddy Simulation (LES) and implicit LES: resolve large turbulent structures while modelling or numerically dissipating smaller scales.

• Aliasing: errors caused by nonlinear terms being under-integrated. Anti-aliasing and adequate quadrature can improve robustness and accuracy.

• Shock capturing: stabilization for discontinuities; current PyFR options include entropy filtering and artificial viscosity for relevant compressible systems.

• Statistical convergence: mean values and spectra are meaningful only after transients have passed and sufficient independent flow events have been sampled.

• Scalability: how efficiently a case uses more GPUs/CPUs; partition balance, communication, and element mix matter.

3.1 The four documented system choices

3.2 A minimum setup vocabulary

External aerodynamics asks how air moves around a body and how the pressure and viscous stresses produce forces, moments, separation, wakes, and noise.

Internal-flow simulation predicts pressure loss, flow uniformity, recirculation, mixing, pulsation, and wall loading inside ducts, channels, manifolds, nozzles, valves, and equipment passages.

This field examines flow through blades, vanes, ducts, intakes, compressors, turbines, fans, propulsors, jets, and energy-conversion passages.

These low-speed flows focus on wind loads, wakes, pedestrian-level wind, ventilation, pollutant transport approximations, terrain/obstacle effects, and wind-energy interactions.

High-speed CFD deals with density and temperature changes, pressure waves, expansion fans, shock waves, and their interaction with boundary layers and wakes.

Unsteadiness is often the reason to use PyFR: vortices, turbulence, periodic shedding, force fluctuations, and acoustic-source behaviour are poorly represented by a single steady picture.

10.1 Start with the engineering decision

A useful simulation begins with a question that can be answered by measurable outputs. "Simulate a drone" is too vague. "Estimate the change in mean lift, drag, and pitching moment when angle of attack increases from 4° to 8° at a specified Reynolds and Mach number" is actionable.

10.2 Build a physics checklist

• Fluid and property model: air, water, ideal gas, viscosity law.

• Compressibility: does density change significantly or are waves/shocks important?

• Viscosity: are wall shear, boundary layers, drag, separation, or mixing important?

• Time dependence: is a mean solution enough, or are fluctuating loads/frequencies required?

• Turbulence: laminar, transitional, turbulent, synthetic inflow, LES/implicit LES, or DNS intent?

• Thermal assumptions: adiabatic or isothermal walls; is solid conduction absent?

• Motion: fixed geometry, rotating reference approximation, or a capability that must be separately verified?

10.3 Geometry and domain simplification

• Remove tiny details that do not affect the decision but force extremely small elements.

• Retain edges, gaps, radii, and appendages that control separation, leakage, or wake formation.

• Place far-field boundaries sufficiently far from external bodies.

• Provide adequate upstream and downstream lengths for internal flows and wakes.

• Use symmetry or periodicity only when the physics truly supports it.

10.4 Mesh thinking for a high-order solver

10.5 The PyFR file and command mental model

Current PyFR documentation presents import, partition, run, restart, region, and export commands, with parallel runs using a mesh partitioning matching the process/device count.

10.6 Monitor before trusting

• Residual or update behaviour: is the solution stable and behaving consistently?

• NaN checks and crash diagnostics: do not treat a completed write as proof of healthy physics.

• Mass flow and conservation: does inlet/outlet balance make physical sense?

• Force histories: have initial transients decayed and has a periodic/statistical regime formed?

• Time-step statistics: is adaptive stepping behaving as intended?

• Field inspection: are there nonphysical oscillations, boundary reflections, or mesh-imprinted patterns?

• Hardware utilization and partition balance: is the problem large enough and balanced for the selected devices?

10.7 Post-processing outputs by decision type

The official test cases provide a useful backbone. The order below is selected for learning, not for industrial importance.

11.1 Learning task for every case

12.1 Verification versus validation

12.2 Minimum credibility ladder

12.3 Common novice mistakes

12.4 Useful PyFR output capabilities

• Fluid-force integration for pressure and viscous force/moment histories.

• Time averaging and standard-deviation statistics for primitive-variable expressions.

• Synthetic-eddy turbulence injection for selected inflow-region use cases.

• Ffowcs Williams–Hawkings far-field noise approximation at specified observer points.

• Sampling, residual, integration, writer, NaN-check, and in-situ visualization plugins described in the user guide.

13.1 Capstone simulation specification template

PyFR becomes valuable when these questions are answered systematically. The objective is not merely to produce high-resolution contours; it is to produce defensible engineering evidence.