Physics-Informed Machine Learning for Low-Complexity Digital Pre-Distortion under Approximate Physics

Sammanfattning

Massive multiple input multiple output (MIMO) base stations run a separate digital pre-distorter (DPD)per antenna at hundreds of MHz, so the per-sample cost of the underlyingpower amplifier (PA) model is a first-order deployment constraint.Classical generalized memory polynomial (GMP) models are cheap andperform well while neural networks (NN) can outperform them but at substantiallygreater cost. This thesis compares three ways of embedding the GMP into an NN forPA behavioural modelling and DPD. A physics-informed neural network(PINN) loss, physics-guided envelope-power input features, and aresidual network summing a GMP prediction with a neuralcorrection. All models are evaluated on a single PA dataset from theRF~WebLab platform on a common floating-point-operation (FLOP) axis. The soft loss adds no benefit and at large weight caps the network atthe GMP's own ceiling. Envelope-power features help only the pure NNon the harder DPD direction, with the physically motivated even-onlypowers outperforming the all-powers variant. The residual architectureoutperforms the pure NN in both accuracy and FLOP cost, and alsoimproves on the already strong GMP prediction. The best residual DPDreaches within 0.55 dB of the empirical normalized mean square error (NMSE) floor and improvesadjacent-channel error power ratio by roughly 4 dB over the bestpolynomial DPD, at 6 to 22 times lower FLOP counts than a pure NNacross multiple NMSE and adjacent channel error power ratio (ACEPR) targets. A data-fraction sweep furthershows the residual reaches the same NMSE targets at roughly100 times less training data than the pure NN. The injection point of the physics prior matters more than itspresence, an approximate model that adds nothing as a loss term yieldsa consistent gain as a forward-path component.