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.
- Digital Predistortion
- Neural Networks
- Massive MIMO
- GPU
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
- Nonlinearity Correction in Massive MIMO Systems via Virtual DPD — Ph.D. Dissertation, Rice University (2022). The full story: PA measurements from a 16T MIMO array, mathematical models, the vDPD architecture using a NN to learn per-beam linearization.
- Virtual DPD Neural Network Predistortion for OFDM-Based MU-Massive MIMO — 55th Asilomar Conference (2021). The peer-reviewed framing: backpropagation through a NN model of the PA to train the predistorter without per-antenna training data.
- OFDM-Based Beam-Oriented Digital Predistortion for Massive MIMO — IEEE ISCAS 2021. Beam-oriented DPD applied in the OFDM guard-band subcarriers, before the precoder.
- Predistortion of OFDM Waveforms Using Guard-Band Subcarriers — 54th Asilomar Conference (2020). The guard-band DPD primitive that the beam-oriented and vDPD work builds on.
- Neural Network DPD via Backpropagation Through a Neural Network Model of the PA — 53rd Asilomar Conference (2019). The original NN-DPD-against-NN-PA-model paper that started the line.
- Design and Implementation of a Neural Network Based Predistorter for Enhanced Mobile Broadband — IEEE SiPS 2019. Implementation-side companion to the NN DPD work.
- GPU-Based Linearization of MIMO Arrays — IEEE SiPS 2020. GPU implementation of MIMO-array DPD, the platform-side counterpart to the Virtual DPD algorithm work. (Also appears under GPU-Based AI-RAN.)