chance.tarver
Projects

2019–2022 · Lead author

AI-based Virtual DPD for MIMO

Neural-network digital predistorters trained against a model of the power amplifier instead of the real thing, so DPD becomes a software problem that scales to massive MIMO.

Per-antenna DPD on a massive MIMO array is intractable if you need real training data from every PA. Virtual DPD sidesteps the problem: model the PA once, then train the predistorter end-to-end against the model, so each new antenna becomes a software adapt instead of a feedback-loop hardware deployment. The technique also moves the DPD block before the precoder, where dimensionality is lower and the linearity that matters is per-beam, not per-antenna.

This line of work was the focus of my Ph.D. dissertation and produced a multi-year run of conference and journal papers leading up to it. It is one of the lines I’m proudest of: it took a problem the field thought needed more hardware and turned it into a model-and-train problem. The current AI DPD for High-Efficiency PAs work picks up the same thread for envelope-tracked PA architectures.

Selected publications