OpenDPD Studio 2.2.14¶
Studio no longer mixes short validation/test preview scores with full-split training metrics. New PA and DPD training starts with TRes-GRU (PA H27, DPD H15), training/evaluation batch size 16.
A real measurement defect caused part of the discrepancy: the historical ACLR calculation included the final zero-padded evaluation segment. Its artificial signal edge could dominate adjacent-channel power. For example, the default filtered NR 20 MHz input read about −38/−35 dBc in the old calculation despite having much lower actual leakage.
The new default opendpd-spectral-v2 excludes padding before estimating
power. Training metrics, validation checkpoint selection and final test
results use the same calculation. Spectrum band overlays use the selected
carrier width. Historical legacy-opendpd-v1 remains available unchanged for
reproduction; already stored results are not relabeled. Modern Studio/CLI
runs use the new default; historical python main.py and paper scripts keep
the legacy metric protocol unless explicitly overridden.
Carrier ACLR is adjacent-carrier power relative to the strongest in-band carrier, in negative dBc. Full-band ACPR has different bands and reference power. IBE means in-band error power ratio; it is not demodulated EVM. See the metric definitions.
Validation includes six synthetic NR/Wi-Fi conditions (5–320 MHz, 16–4096 QAM, six virtual PAs), measured DPA 160/200 MHz captures and APA 200 MHz. Each condition trained both PA and DPD. Independent FFT integration agreed with stored ACLR within 1e-8 dB. APA TRes-GRU PA H27 / DPD H15, batch 16, 300 epochs each achieved −55.61 dBc test carrier ACLR, using a checkpoint selected only by validation ACLR. This is prediction through a PA model trained on measured APA data, not a new RF hardware measurement.