Neural Network Digital Pre-Distortion

The core principle of NN DPD is learning the inverse of the Power Amplifier's (PA) non-linear transfer function. A neural network is trained to model the complex baseband distortion such that when the pre-distorted signal passes through the PA, the cascade results in a linear output.

• Training Data: Uses time-aligned input and output I/Q samples from the PA.
• Architecture: Typically employs feedforward or recurrent networks with memory taps to capture memory effects (long-term thermal and electrical hysteresis).
• Objective: Minimizes the Error Vector Magnitude (EVM) and Adjacent Channel Leakage Ratio (ACLR) at the PA output.

Unlike static look-up tables, NN DPD systems can adapt coefficients in real-time to track changes in the PA's behavior due to aging, temperature drift, or antenna load mismatch (VSWR changes).

• Online Learning: Uses stochastic gradient descent on streaming I/Q samples to continuously refine the pre-distorter.
• Tracking Loops: Integrates with the transmit observation path receiver (feedback loop) to capture the PA output.
• Benefit: Maintains spectral mask compliance without factory re-calibration.

NN DPD can be implemented using two primary learning architectures, each with distinct trade-offs in stability and complexity.

• Indirect Learning Architecture (ILA): Trains a post-distorter on the PA output and copies its weights to the pre-distorter. It is stable but mathematically sub-optimal in noisy conditions.
• Direct Learning Architecture (DLA): Backpropagates the error directly through a modeled PA to update the pre-distorter. It achieves better linearization but requires an accurate PA model to avoid instability.
• Augmented ILA: A hybrid approach using a small auxiliary neural network to correct the bias inherent in the classical ILA structure.

Wideband signals (e.g., 100 MHz for 5G NR) suffer from significant memory effects where the PA output depends on past inputs, not just the current one. NN DPD excels at modeling these long-term dependencies.

• Time-Delay Neural Networks (TDNN): Explicitly feed delayed versions of the I/Q signal into the input layer.
• Recurrent Neural Networks (RNN/LSTM): Use internal hidden states to implicitly capture temporal dependencies, often outperforming TDNNs for severe memory effects.
• Volterra Series Replacement: Neural networks can model high-order non-linearities with memory using far fewer parameters than a full Volterra series expansion.

Deploying NN DPD on a Field-Programmable Gate Array (FPGA) or Application-Specific Integrated Circuit (ASIC) requires extreme optimization to meet the latency budget of a few nanoseconds.

• Quantization-Aware Training (QAT): Trains the network with simulated low-bit precision (e.g., INT8 or INT4) to prevent accuracy collapse during fixed-point conversion.
• Weight Pruning: Removes near-zero connections to reduce the number of multiply-accumulate operations (MACs).
• Knowledge Distillation: Trains a compact 'student' network to mimic a high-capacity 'teacher' network, preserving linearization performance with a fraction of the resources.

Modern NN DPD architectures extend beyond single-band, single-antenna systems to handle the cross-talk and intermodulation products found in advanced radio front-ends.

• Dual-Band DPD: A single neural network processes two widely separated carriers simultaneously, suppressing intermodulation distortion that falls into the receive band.
• MIMO DPD: Compensates for crosstalk between multiple transmit chains and beam-dependent impedance loading in phased arrays.
• Cross-Coupling Modeling: The network input layer includes I/Q streams from adjacent transmitters to learn the interaction matrix.

The fundamental physical phenomenon that Neural Network Digital Pre-Distortion (NN-DPD) aims to correct. As a power amplifier approaches its saturation region to maximize power-added efficiency (PAE) , it exhibits amplitude-to-amplitude (AM-AM) and amplitude-to-phase (AM-PM) distortion. This non-linear behavior causes spectral regrowth into adjacent channels and degrades the error vector magnitude (EVM) of the transmitted signal. NN-DPD learns a precise inverse model of this complex, often memory-influenced, transfer function.

A classical mathematical framework for modeling non-linear dynamic systems with memory, which serves as the theoretical precursor to many NN-DPD architectures. The Volterra series represents the system output as a sum of multi-dimensional convolution integrals. However, its practical use is limited by the exponential growth in the number of coefficients required to model higher-order non-linearities. Neural networks, particularly recurrent or convolutional architectures, can be seen as efficient, learned approximators of truncated Volterra kernels.

The dominant training paradigm for NN-DPD. Instead of directly learning the pre-distorter, a post-distorter model is first identified by placing an identical copy of the neural network after the power amplifier. The error between the attenuated PA output and the desired linear signal is minimized. Once trained, the learned parameters are copied to the pre-distorter block placed before the PA. This elegant closed-loop structure avoids the need for a pre-existing inverse PA model.

A critical capability of advanced NN-DPD that distinguishes it from static, memoryless look-up tables (LUTs). Thermal memory effects (due to transistor heating) and electrical memory effects (due to bias circuit impedance) cause the PA's distortion to depend on the history of the input signal. Architectures like Temporal Convolutional Networks (TCNs) or Gated Recurrent Units (GRUs) are employed to capture these long-range temporal dependencies and apply a dynamic, history-aware correction.

A specialized neural network architecture inherently suited for NN-DPD because it directly processes the in-phase (I) and quadrature (Q) components of the baseband signal as a single complex entity. By using complex-valued weights and activation functions, and performing backpropagation via Wirtinger calculus, these networks preserve the phase relationships critical for canceling AM-PM distortion. This often leads to more compact and physically meaningful models compared to dual-channel real-valued networks.

A methodology that bridges classical signal processing and deep learning for DPD. An iterative optimization algorithm, such as the Iterative Shrinkage-Thresholding Algorithm (ISTA) , is unrolled into a fixed number of neural network layers. Each layer corresponds to one algorithm iteration, and key parameters (like step sizes) are made learnable. This results in a highly parameter-efficient NN-DPD model with a strong theoretical prior, requiring less training data than a generic black-box network.