| Author | Message | |||
     ufgator New member Username: ufgator Post Number: 1 Registered: 03-2010 |
I am using Flex version 5.0.7 and am attempting to run a time dependent thermal simulation. I have successfully simulated the steady state case, but in including the time dependence I receive an error: memory allocation failed: MBalloc(20175776)=0 -called from sparsemx::resize -called from loading::map_matrix -called from timesolve::evolve -called from control::do_runjob -called from ccontrol::runphase I've attached the run file and would appreciate any insight as to what may be causing this error. I am currently running this on Windows 7. Thanks.
| |||
| rgnelson Moderator Username: rgnelson Post Number: 1336 Registered: 06-2003 |
1) With a LIMITED REGION, all layers and surfaces that are named within the definition block contain the region. In your case, by naming the Package, SiC and GaN layers "Void", you have created very thin slots that extend all the way through the substrate. I suspect this is not what you want. It requires an enormous mesh, which has apparently exceeded the storage capacity of your computer. If you leave the voided regions out of the region declaration entirely, the thin features will not be included in the thick layers, and the mesh requirements will be much, much smaller. See attached. 2) There is a meshing error that is probably due to the extremely thin layer 3. You might be able to fix this with a larger NGRID, or with a MESH_SPACING control in layer 3, or by using coordinate scaling in layer 3 (see "coordinate scaling" in the help index). The problem will run even with the error, if you click "continue" through the error reports. 3) you have an initial discontinuity in the temperature distribution (60 internally and 100 on the bottom surface). This requires a dense mesh at the bottom and tiny timesteps at the beginning. I assume you don't really care about the first millisecond or so, so you could initialize with a smooth distribution in z (e.g. exp(-z/k), with k a reasonable distribution width). You can supress the initial smoothing and regridding with SELECT INITGRIDLIMIT=0 SMOOTHINIT=OFF 4) you have requested to complete the simulation in a single timestep (TIME 0 to 1 by 1). This is not going to happen, and it gives FlexPDE a headache because FlexPDE analyzes TWO initial timesteps to determine an effective starting stepsize. Either leave the "BY" clause out, or give it a much smaller number. If you intended only to see the equilibrium distribution, then you should delete the dt() term in the equation. 5) You can save some time on re-runs with identical domains if you TRANSFER data at the end of the problem and use TRANSFERMESH to import the mesh. Comment out the TRANSFERMESH on the first run. Notice that if the TRANSFER output command is executed more than once there will be a sequence number in the file name. 6) You should put in some MONITORS so you can see how the solution is progressing. 7) You say your are using version 5.0.7, but this is a very old version. You should update to 5.1.4. This is a free download from our website www.pdesolutions.com/v5support.html 8) grid(x,y,0) makes a very strange plot. The proper format for a cutplane plot is "grid(x,y) on z=<something>" or "grid(x,y) on 'named surface'". 9) Even though the thin features on the active layer are driving the mesh size, you may want to manually request more mesh cells in this area for better resolution of the hot spot. In the "Active Area" region, use MESH_DENSITY=<width>/<number> where <width> is the width of the feature and <number> is 2 or 3 or 4.
| |||
| ufgator New member Username: ufgator Post Number: 2 Registered: 03-2010 |
Thanks so much for all of your help! Your suggestions have made a huge improvement in my simulations, and even decreased the run time from a few hours to a few minutes. Thanks again! | |||
| ufgator Junior Member Username: ufgator Post Number: 3 Registered: 03-2010 |
I am trying to implement a thin region within a layer in which heat is generated. I get a floating point equals zero error, which doesn't occur if I make the layer thicker. I tried your suggestion of increasing ngrid, controlling the mesh_density, and using the scaling. Either I get different errors or extremely long run times. With scaling the z axis I get: bad nodecell patch Too many neighbors Broken Node Ring! An internal processing error has occured. Please contact PDE Solutions Inc. --Called from compmesh3::reconnect --Called from compmesh3::scan_flat_cells --Called from mesh3::relax_hooks3 --Called from domainspec::grid_layer --Called from mesher3::new I attempted to use the same setup in the sample run file for scaled z. I would appreciate any help.
| |||
| rgnelson Moderator Username: rgnelson Post Number: 1342 Registered: 06-2003 |
You have the right idea in the zscaling you have done, but I don't think you have been nearly aggressive enough in your scaling numbers. In the attached, I have used 1/2 and 1000 for your scaling numbers, and added a scale of 10000 for layer 4. This grids successfully and runs quickly, but gets a mesh generation error in trying to build the cut-plane plot meshes. I don't yet know what causes this, but you can click through the error dialogs and get plots.
|
GaN HEMT