[prak_1989]

I am trying to validate some of the statistics in isotropic turbulence by DNS. I am using Finite volume method, where the inviscid and viscous fluxes at the faces are interpolated by symmetric average.

I am getting a negative pressure which leads to NaN after certain iterations. What could be the problem ?

I am stuck for more than a week, someone please suggest me something.


Thanks

[FMDenaro]

Originally Posted by prak_1989

I am trying to validate some of the statistics in isotropic turbulence by DNS. I am using Finite volume method, where the inviscid and viscous fluxes at the faces are interpolated by symmetric average.

I am getting a negative pressure which leads to NaN after certain iterations. What could be the problem ?

I am stuck for more than a week, someone please suggest me something.


Thanks

for incompressible flows, negative pressure values mean nothing (only the gradient is present in the momentum equation) and cannot be the reason for your problem...
I think you get into numerical instability so you have to check if you satisfy the stability requirement that you discretization implies.

[mprinkey]

Is your mesh structured or unstructured? Collated or staggered? How are you evaluating your pressure terms? If you are using collated pressure/velocity grids, you MUST use Rhie-Chow to implement the pressure (correction) equation. Is your formulation fractional step or SIMPLE/PISO?

Also simple face averaging to construct convective terms is a very unstable formulation. Though it does minimize artificial diffusion on evenly space meshes, it does lead to instabilities arising from characteristic/information flow directions. Upwind convection terms are chosen to match the natural flow of information downstream. If you are going to violate that, you really need to know what you are doing.

[FMDenaro]

Originally Posted by mprinkey

Is your mesh structured or unstructured? Collated or staggered? How are you evaluating your pressure terms? If you are using collated pressure/velocity grids, you MUST use Rhie-Chow to implement the pressure (correction) equation. Is your formulation fractional step or SIMPLE/PISO?

Also simple face averaging to construct convective terms is a very unstable formulation. Though it does minimize artificial diffusion on evenly space meshes, it does lead to instabilities arising from characteristic/information flow directions. Upwind convection terms are chosen to match the natural flow of information downstream. If you are going to violate that, you really need to know what you are doing.

well, in case of DNS that's not a problem....as the scale between convection and diffusion must be comparable on the size of the DNS grid, you work at cell Reynolds number O(1). In this case, also the FTCS can work (it is unconditionally unstable only for the inviscid case), provided that your computational parameters fulfill the numerical stability constraints.

However, I agree that we need further details to understand better the issue.

[prak_1989]

I am sorry for not mentioning it earlier. I am solving 3-D compressible Navier-Stokes equation, with Reynolds No. 515... My domain L * B * H with L = B = H = 1.54E-4... I am using structured, collated grids with no. of grids = 96**3.. I am using 4th order R-K time stepping to march in time.

Pressure is calculated from P = (gamma - 1)*(rho* E - 0.5*rho*V**2)

P = Pressure
rho = density
E = int.energy + K.E
V = Velocity vector

[mprinkey]

Is this density-based or pressure-based solver? If you are using RK4, I'd assume density based.

In either case, you need to use Rhie-Chow to turn the pressure differences between cell pairs into mass fluxes on the faces. Collocated meshes require special treatment of the pressure terms to avoid pressure/density/velocity decoupling. Are you seeing checkerboard pressure patterns prior to the NaNs? If so, this is likely the problem.

[FMDenaro]

Your case is a very low-Mach problem, do you use some preconditioner or special low-Mach formulation?

[FMDenaro]

Originally Posted by prak_1989

I am sorry for not mentioning it earlier. I am solving 3-D compressible Navier-Stokes equation, with Reynolds No. 515... My domain L * B * H with L = B = H = 1.54E-4... I am using structured, collated grids with no. of grids = 96**3.. I am using 4th order R-K time stepping to march in time.

Pressure is calculated from P = (gamma - 1)*(rho* E - 0.5*rho*V**2)

P = Pressure
rho = density
E = int.energy + K.E
V = Velocity vector

I forgot a relevant question ...is your Re number based on lambda (the Taylor microscale)? Be carefull, your grid resolution could be not sufficient to work in a real DNS.

[prak_1989]

I tested the solver with Mt = 0.1 and it works well but not with 0.3. I think its a problem with scheme (Symmetric avg)

[FMDenaro]

Originally Posted by prak_1989

I tested the solver with Mt = 0.1 and it works well but not with 0.3. I think its a problem with scheme (Symmetric avg)

I would assume that an increasing in the Mach number corresponds to an increasing in the Reynolds number. Do you have enough grid points?

[prak_1989]

I am validating the results by my colleagues who did with 96**3 grid points, turbulent Mach No. (Mt = 0.3) and scheme used were spectral by one person and 5th order WENO for inviscid and central difference for viscous flux by other person.