2.2.14 measurement validation¶
The numerical record is stored in the repository, including input hashes, generated waveform settings, checkpoint hashes, sample counts and independent-check errors.
All experiments use TRes-GRU PA H27 and DPD H15, training/evaluation batch 16, seed 0, AdamW, initial learning rate 0.005, ReduceLROnPlateau, frame length 200. The eight-condition matrix uses 20 PA / 30 DPD epochs; synthetic stride 4, measured stride 16. These short runs check the full pipeline and scoring; they are not optimized model comparisons. The synthetic stimuli use 65,536 samples each and the generator's disclosed engineering waveform conventions, not complete standard-conformant RF packets.
| Input | QAM | PA data/model | DPD test carrier ACLR average (dBc) |
|---|---|---|---|
| NR 5 MHz | 16 | Rapp soft limiter | −63.44 |
| NR 20 MHz | 64 | Memory polynomial | −63.03 |
| NR 100 MHz | 256 | Saleh TWTA | −54.46 |
| Wi-Fi 6 20 MHz | 64 | Rapp AM/AM + AM/PM | −59.96 |
| Wi-Fi 6 80 MHz | 1024 | Generalized memory | −64.35 |
| Wi-Fi 7 320 MHz | 4096 | Doherty two-path | −46.78 |
| DPA 160 MHz | 1024 | Measured capture | −41.45 |
| DPA 200 MHz | 64 | Measured capture | −32.00 |
Each stored result and every raw train/validation/test PA-output split was checked against a separately written NumPy FFT periodogram average. The oracle builds its Hann window, FFTs and band masks directly, without SciPy Welch or production PSD/integration helpers. Every difference was below 1e-8 dB. The test suite also checks known −55/−60 dBc adjacent tones, carrier counts 1/4/5/10, odd FFT lengths, unavailable Nyquist bands and padding containing arbitrary nonzero model output.
APA reproduction¶
APA 200 MHz uses the bundled measured 256-QAM capture at 983.04 MSa/s,
5 × 40 MHz carrier bands and nperseg=19662. Both models trained for
300 epochs, frame stride 1. The validation-selected DPD checkpoint is epoch
222 (zero-based 221), scored on all 19,662 held-out test samples.
| Metric | Test value |
|---|---|
| Carrier ACLR left | −55.537388 dBc |
| Carrier ACLR right | −55.689390 dBc |
| Carrier ACLR average | −55.613389 dBc |
| Legacy carrier ACLR average, same checkpoint | −55.612043 dBc |
Full-band ACPR left (general-spectral-v1) |
−57.431678 dBc |
Full-band ACPR right (general-spectral-v1) |
−58.822323 dBc |
| NMSE | −45.380323 dB |
The oracle difference is below 2e-14 dB. Legacy/new agreement on APA shows that reaching the requested ~−55 dBc level does not come from changing measurement bands to improve a number. Full-band ACPR is reported separately because its integration bands and reference differ from carrier ACLR.
All DPD rows above are PA-model predictions, including models trained on measured captures. They do not establish hardware DPD performance. One seed is a reproduction check, not a statistical performance guarantee. The compute used an immutable candidate source snapshot, whose Python-file manifest hash is included with the numerical record.