Neural Network DPD

Leverages the universal approximation theorem to model arbitrary continuous non-linear functions without requiring a priori knowledge of the amplifier's physical characteristics. Unlike Volterra series or memory polynomials, neural networks can capture high-order non-linearities and complex interactions between AM-AM and AM-PM distortion without an exponential explosion in model coefficients.

• Models Doherty and envelope tracking PAs with severe non-linearity
• Captures complex cross-terms implicitly through hidden layer activations
• Avoids the curse of dimensionality inherent in polynomial basis expansions

Employs recurrent architectures such as LSTMs, GRUs, and tapped delay line feed-forward networks to model long-term memory effects caused by thermal dynamics, bias network impedance, and charge trapping in gallium nitride (GaN) transistors. These architectures maintain an internal state that captures temporal dependencies spanning hundreds of samples.

• Real-Valued Time-Delay Neural Networks (RVTDNN) process I/Q components with configurable memory depth
• Recurrent layers capture thermal time constants without explicit physics modeling
• Handles carrier aggregation scenarios with wideband memory effects

Supports both Indirect Learning Architecture (ILA) and Direct Learning Architecture (DLA) training paradigms. ILA identifies the predistorter by swapping the PA input and output, enabling non-iterative coefficient estimation. DLA minimizes the error between the desired linear output and actual PA output through gradient-based optimization, achieving superior linearization at the cost of iterative training.

• ILA enables fast initial coefficient acquisition without iterative loops
• DLA provides optimal minimum mean-squared error convergence
• Both architectures support online adaptation during live transmission

Integrates Crest Factor Reduction (CFR) and linearization into a single end-to-end neural network, jointly optimizing the trade-off between peak-to-average power ratio reduction and distortion compensation. This unified approach avoids the suboptimal cascading of separate CFR and DPD blocks, enabling operation closer to the amplifier's 1 dB compression point.

• Single model handles signal clipping and non-linearity correction
• Maximizes Power-Added Efficiency (PAE) by reducing required back-off
• Learns optimal peak regrowth control for specific modulation schemes

Addresses the unique challenge of beam-dependent non-linearity in hybrid beamforming arrays, where each beam experiences a different composite distortion due to per-antenna PA variations and mutual coupling. Neural networks can condition the predistorter on beamforming weights or spatial direction, enabling beam-aware linearization without requiring a separate DPD model per beam.

• Conditions predistortion on beam index or precoding matrix
• Compensates for antenna mutual coupling effects
• Scales to 64T64R and larger arrays without linear coefficient growth

Deploys on FPGA and ASIC platforms through quantization-aware training, weight pruning, and knowledge distillation. Post-training 8-bit integer quantization reduces memory footprint and inference latency while maintaining linearization performance within 0.5 dB of floating-point accuracy. Sparse neural network architectures exploit structured sparsity to accelerate matrix multiplications on dedicated neural processing units.

• INT8 quantization with minimal ACLR degradation
• Structured pruning removes redundant neurons and connections
• Real-time inference at sub-microsecond latency for 5G NR bandwidths

The fundamental problem that DPD solves. When a PA operates near saturation for efficiency, its output deviates from a linear function of the input, causing amplitude distortion (AM-AM) and phase distortion (AM-PM). This non-linearity generates intermodulation products that cause spectral regrowth into adjacent channels. Neural network DPD models must capture these complex, often device-specific, non-linear transfer functions.

A PA's output depends not only on the current input but also on past inputs due to:

• Thermal dynamics: Transistor heating changes gain over time
• Bias network impedance: Low-frequency envelope currents modulate the bias point
• Trapping effects: Charge capture in semiconductor materials

Neural networks with tapped delay lines or recurrent architectures are particularly well-suited to modeling these long-term temporal dependencies that simpler memoryless models miss.

The Volterra series provides the mathematical foundation for modeling non-linear dynamic systems with memory using multi-dimensional convolution kernels. The Generalized Memory Polynomial (GMP) is a practical, truncated form that includes cross-terms between the signal and its lagging/leading envelope values. These models serve as the theoretical baseline that neural network DPD architectures aim to outperform in terms of complexity and accuracy.

Two fundamental architectures for DPD coefficient identification:

• Indirect Learning Architecture (ILA): Swaps PA input/output to estimate the postdistorter, then copies it as the predistorter. Simpler but assumes commutability.
• Direct Learning Architecture (DLA): Iteratively minimizes the error between desired linear output and actual PA output. More accurate but requires a PA model for gradient computation.

Neural network DPD typically uses a direct learning approach with backpropagation.

The two critical metrics for evaluating DPD performance:

• Adjacent Channel Leakage Ratio (ACLR): Measures spectral regrowth into neighboring channels. Regulatory compliance typically requires -45 dBc or better.
• Error Vector Magnitude (EVM): Quantifies in-band signal quality by measuring constellation point deviation. Lower EVM means higher data rates.

Neural network DPD must simultaneously optimize both metrics, often achieving 2-3 dB improvement over conventional methods.

Modern high-efficiency PA architectures that present extreme linearization challenges:

• Doherty PAs: Combine main and peaking amplifiers, creating sharp non-linearity at the efficiency peak transition point.
• Envelope Tracking: Dynamically modulates supply voltage, introducing complex memory effects from the power supply modulator.

These architectures are primary candidates for neural network DPD due to their severe, non-standard distortion characteristics that exceed the modeling capacity of polynomial-based methods.