论
Member
A memoryless DPD corrects static AM-AM and AM-PM, and on narrowband it looks like magic. Put the same GaN PA behind a 100 MHz-plus 5G NR waveform and the magic fades: ACPR improves a few dB, then stalls, and the upper and lower adjacent channels stop matching. What you are seeing is memory, and no lookup table can fix it.
Three mechanisms dominate. Thermal memory: junction temperature follows the envelope on microsecond-to-millisecond scales, so gain depends on what the PA was doing a moment ago. Bias modulation: the drain bias network has finite impedance at envelope frequencies, so the rail moves with the envelope; GaN trapping (gate lag, drain lag) adds time constants from microseconds to seconds.
The fix starts with the model: a generalized memory polynomial (GMP) with lagging and leading envelope terms captures most of it. Checklist: characterize with the real modulated signal, never CW two-tone alone; check ACPR asymmetry before touching the coefficients; measure the bias network impedance at envelope frequencies; keep the observation bandwidth at least five times the signal bandwidth; and watch the crest factor, because a DPD trained on one PAPR underperforms on another.
Three mechanisms dominate. Thermal memory: junction temperature follows the envelope on microsecond-to-millisecond scales, so gain depends on what the PA was doing a moment ago. Bias modulation: the drain bias network has finite impedance at envelope frequencies, so the rail moves with the envelope; GaN trapping (gate lag, drain lag) adds time constants from microseconds to seconds.
The fix starts with the model: a generalized memory polynomial (GMP) with lagging and leading envelope terms captures most of it. Checklist: characterize with the real modulated signal, never CW two-tone alone; check ACPR asymmetry before touching the coefficients; measure the bias network impedance at envelope frequencies; keep the observation bandwidth at least five times the signal bandwidth; and watch the crest factor, because a DPD trained on one PAPR underperforms on another.
