NaN generation while trying to implement a moving wall BC

[spaceman-spence]

Hi everyone,

I have recently implemented a custom BC to account for a moving wall while simulating a standing shock/detonation to account for a no slip BC within a wave reference frame. Below I have included the method I use that gets called within the bcnormal method on the moving boundaries. cj_speed is the variable within my prob_parm keeping track of the wall velocity and is set in the input file. From my understanding, the ghost cell needs to be set such that the average velocity between the ghost cell and the interior cell is equal to the velocity I want the wall to be traveling at.

AMREX_FORCE_INLINE
void moving_wall(amrex::Real s_ext[NVAR],
  const int idir,
  const int sgn,
  amrex::GeometryData const& geomdata,
  ProbParmDevice const& prob_parm)
{
  amrex::Real rho = s_ext[URHO];
  amrex::Real u_wall = -prob_parm.cj_speed;
  amrex::Real u_ghost_old = s_ext[UMX] / rho; // u_ghost_old = -u_interior
  
  // modify ghost value
  s_ext[UMX] = rho * (2.0 * u_wall + u_ghost_old);
}

My main problem is when I run my detonation cases, where 20% of the domain is supersonic unburned fuel before the shock and the remaining domain is post shock, I develop NaN’s instantly along the post shock/detonation region. There are no issues within the incoming flow, only after the detonation, as seen below. Since it crashes on the first time step, I cant completely confirm no boundary layers develop on the incoming flow, but when probing individual points at the wall, their velocities are unchanged from the initialization.

If I slow down the wall velocity to <=10% of the speed I want the wall at, I can run the simulation for a couple hundred time steps, before NaN/s develop at the wall. 
Attached below is the simulation shortly before crash, then at the same location with the crash. They become concentrated at one location, close to where the detonation meets the oblique shock wave (which forms since the inflow is moving faster than the no slip wall is moving).

However, If I run a detonation simulation in a lab reference frame, I encounter no problems; the denotation travels down the extended domain with a boundary layer developing within the burned gas region as expected. Since the lab frame version works with no problems I am inclined to believe it is related to my BC implementation, not a numerical problem, since that is the only other change between the lab frame and wave frame code. Yet, as far as I can tell, my implementation is correct, and so I am not sure what else to consider to solve the issue.

Has anyone else tried to implement a moving wall or has anyone encountered a similar generation of NaNs? I am at a loss for what is causing them and would appreciate any suggestions on how I can further investigate the source of these NaNs.

Many thanks!
Spencer

[baperry2]

How are you specifying the full state in the ghost cells? Are you modifying the total energy of the ghost cell (s_ext[UEDEN]) appropriately to account for the changed kinetic energy due to the way you are changing velocity? If not, the temperature and pressure in ghost cells will take on bad values which would explain what you are seeing. This complexity is one reason we just compute the domain boundary fluxes for isothermal walls completely independently of the bcnormal function.

[spaceman-spence]

I see. I completely forgot that the kinetic energy would also need to be modified so I only modified the x-momentum from the default no-slip state that the bc normal function receives. I'll give that another look and see if the energy fixes the problem. Thank you!!

[spaceman-spence]

That appears to have fixed my problem! Thanks again for the help!

[RSuryaNarayan]

https://arc.aiaa.org/doi/abs/10.2514/6.2025-1386 I invite you to check out this :-)