Oversample FD - #224
Oversample FD#224SeanMollet wants to merge 1 commit into
Conversation
The frequency domain dumps and probes (DFT during the run) were sampled at the Nyquist rate of the excitation, while the time domain dumps and probes use OverSampling (default 4) samples per Nyquist interval. At the Nyquist rate the spectrum just above the highest excited frequency aliases onto the upper band edge. Now all recording follows OverSampling. Horn antenna example (10-20 GHz), NF2FF box with frequencies: directivity at 20 GHz 15.87 -> 16.91 dBi, the value of the time domain dumps; the patterns at 10, 15 and 20 GHz now agree with them within 0.001 dB (were up to 18 dB off at 20 GHz). On the GPU engine the samples are one kernel each: timestepping 2.65 s -> 2.82 s on an RTX 5090. On the CPU engines the host work of the frequency domain dumps grows with the factor. Dipole_SAR.py: the reference peak SAR values, calibrated with Nyquist sampling, were ~2 % high (1 GHz in a 0-1.5 GHz excitation). They are recalibrated at the default OverSampling, where they have converged (OverSampling 8 and 16 within 0.01 %). The OverSampling documentation (Python, Matlab) now names the frequency domain recording and the probes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
Well in the Dipole_SAR tutorial with its rather large FD SAR dumpbox I see a speed penalty of ~20% while the SAR values only differs less than 0.5% and it seems the total number of timesteps which happens to end at has an even bigger influence. |
|
Yeah, I think a configurable option with a 1 default is great. |
|
I have rewritten this commit, the default is now 1 again but now can be set via CSXCAD per dump-box. Make sure to update and rebuild CSXCAD. |
2 participants
Oversample FD to reduce aliasing at the edge.
Tested on the Horn example and Dipole_SAR.
Somewhat expensive on a CPU, essentially free on a GPU. This changes results for anyone using FD dumps or probes. Happy to put it behind a separate variable defaulting to 1 if you'd rather preserve existing behavior.
Thanks,
Sean