Learning Forum

[rcyork]

1. When I solve a magnetostatic model in Maxwell 3D with the percent error of 0.01%, I see the resultant excitation curve (i.e. field strength at a specific point vs excitation current, or integrated energy over all volume vs excitation current) is scattered (not smooth).
 
What causes this? Is the percent error not small enough? How small should the percent error should be to obtain a smooth excitation curve?
 
2. In Maxwell 3D, sometimes I get an error message as follows:
 

[error] ProjectS1Opt, Design:Maxwell3DDesign1 (Magnetostatic), IPC failed for child process 3dnms. Please contact ANSYS technical support. -- Simulating on machine: localhost (10:59:47 PM  Nov 22, 2019)

 
What would cause these error messages, and how can I avoid this?
 
3. Often I see the total energy changes by much compared to the previous passes and then gets back, towards the end of the calculation (i.e. near convergence). How can this happen, or is this kind of behavior considered normal?
 
4. Regarding the “smooth BH curve” and "iterative solver": With these enabled, is Maxwell supposed to give the same calculation result as that with them not enabled, or can it be different from what is obtained without smoothing the BH curve and/or the direct solver?
 
5. I wish to accelerate calculations, i.e. reduce the time required for getting the calculations converged with a sufficiently high precision. What are the most crucial: CPU clocks, CPU core counts, RAM, storage I/O speed?

[mchristi]

 

1. When I solve a magnetostatic model in Maxwell 3D with the percent error of 0.01%, I see the resultant excitation curve (i.e. field strength at a specific point vs excitation current, or integrated energy over all volume vs excitation current) is scattered (not smooth).  What causes this? Is the percent error not small enough? How small should the percent error should be to obtain a smooth excitation curve?

• This is not caused by percent error being too large. The default value of 1% is sufficient for the adaptive refinement process for simulations which are interested in the total overall solution (and not the solution at a particular location such as a point, line). Reducing to 0.01% is more than sufficient and will provide a solution with even better overall global energy accuracy.


• Jumping energy, % energy error or field strength at a particular location has other causes. Possible reasons include: mesh is too coarse around the point/line of interest, BH curve is not smooth, BH curve does not extended far enough into saturation, BH curve does not have enough points or nonlinear residual is too large.

 

Some possible solutions are:

a)       Apply a global length-based mesh operation on all objects in the entire model including the background region

b)      Check that slope of the BH curve (1^st derivative) is smooth

c)       Check that BH curve is extended well into saturation, usually > 2.5T

d)      Check that the BH curve has at least 20 points

e)      Reduce nonlinear convergence from 0.001 to 0.0001 or even 1e-5

f)        Draw dummy object such as a box around the point (or line) of interest and apply a mesh operation such as 1000 tetrahedra to the dummy object

 

2. In Maxwell 3D, sometimes I get an error message as follows: [error] ProjectS1Opt, Design: Maxwell3DDesign1 (Magnetostatic), IPC failed for child process 3dnms. Please contact ANSYS technical support. -- Simulating on machine: localhost (10:59:47 PM  Nov 22, 2019)   What would cause these error messages, and how can I avoid this?

        g)       This is most likely caused by insufficient RAM, by insufficient hard disk storage or could be a license issue.  Check the Maxwell PROFILE to ensure RAM is not being exceeded by the solver (you can also try to solve on another machine with more RAM).  Check your hard disk to ensure there is adequate storage remaining, after getting this error.  Finally, verify that the communication to the ANSYS license server is not being interrupted and that the IP address of the computer (if dynamic) is not changing.

 

3. Often I see the total energy changes by much compared to the previous passes and then gets back, towards the end of the calculation (i.e. near convergence). How can this happen, or is this kind of behavior considered normal?

      h)      This is usually due to either a significant mesh refinement between two passes in a critical area of the model (or) due to problems with the nonlinear convergence. To resolve mesh refinement issues, try a) above. For resolving nonlinear convergence issues, try b) c) d) e) above.  

 

4. Regarding the “smooth BH curve” and "iterative solver": With these enabled, is Maxwell supposed to give the same calculation result as that with them not enabled, or can it be different from what is obtained without smoothing the BH curve and/or the direct solver?

      i)        The results can be slightly different when using “smooth BH curve” and "iterative solver" but there should not be a significant difference if b) c) d) e) above are satisfied.  As more than 20 points are added to the BH curve, the difference between using “smooth BH curve” and "iterative solver" gets smaller/insignificant.

 

[noji]

Thank you for your comments. I had my colleague (@rcyork) posted my questions above on my behalf, however, as he was not able to include figures, my original intentions may have not been very clear. Let me elaborate on my questions with figures while addressing your comments.

Ad 1m. By "the resultant excitation curve" what I meant was the field strength at a specific point vs excitation current, or integrated energy over all volume vs excitation current, as can be seen in the figure. When I fitted these points in the plots with polynomials, I see that the points deviate from the fit by as much as about 4%. This behavior seems unphysical to me, and I wanted to find out what the root cause of this issue was.

According to the mesh statistics (see the next figure), the minimum edge lengths from the adaptive refinement process in the vicinity of the region of interest are in the order of a few millimeters, would a) applying a global length-based mesh operation or f) drawing a dummy object and applying a mesh operation solve this problem of non-smooth excitation curve? Or e) is it crucial to reduce the nonlinear convergence to a smaller value to get a smoother curve?

I would tend to think that b) the slope of the BH curve is smooth, c) the BH curve is extended into saturation, with the highest point being at 3 T, and d) the BH curve has 50 data points. Regarding c), the calculation results remain practically unchanged by adding a data point (by extrapolation) to the BH curve.

Ad 3m. Regarding this convergence behavior, please have a look at the figure. You wrote "a significant mesh refinement between two passes"; I set the refinement per pass at 10% or smaller, but can this induce a significant mesh refinement? If so, what would be an appropriate value?

Ad 2m. I have had RAM increased on my computer, and this error has gone away. Regarding my last question, would you think of anything I need to be mindful of when acquiring a workstation for magnetostatic analysis in Maxwell? I've found and read two threads on this forum (this and this). What would you prioritize: CPU clocks, CPU core counts, RAM, storage I/O speed, or something else?