Convolutional Neural Network DPD (CNN-DPD)

Unlike polynomial models that require manual selection of basis functions and truncation orders, CNN-DPD uses 1D convolutional layers to automatically learn the most relevant temporal features from raw I/Q data. The network discovers hierarchical representations: early layers capture short-term memory effects, while deeper layers model long-range dependencies and complex envelope interactions. This eliminates the need for domain expertise in Volterra kernel selection and reduces model development time.

CNN-DPD architectures achieve superior linearization performance with fewer parameters than equivalent memory polynomial or RVTDNN models. Key mechanisms include:

• Weight sharing across temporal positions via convolutional kernels
• Local receptive fields that focus on relevant signal history
• Pooling layers that reduce dimensionality while preserving critical features

This compact representation reduces FPGA resource utilization and enables lower-latency real-time inference compared to fully-connected alternatives.

Dilated convolutions and multi-branch architectures enable CNN-DPD to simultaneously capture short-term memory effects (thermal trapping, bias modulation) and long-term dependencies (self-heating, charge trapping) across different time scales. Parallel convolutional paths with varying kernel sizes and dilation rates process the input at multiple temporal resolutions, then fuse the extracted features. This multi-scale approach is particularly effective for GaN power amplifiers exhibiting complex memory spanning nanoseconds to milliseconds.

CNN-DPD natively processes complex baseband signals as two-channel inputs (I and Q), preserving the phase relationships critical for linearization. Unlike real-valued networks that treat I and Q independently, complex-aware architectures can employ:

• Complex convolutional layers with learned real and imaginary kernels
• Complex activation functions like modReLU that respect phase information
• Complex batch normalization for stable training

This preserves the envelope-phase coupling essential for compensating AM-PM conversion and cross-modulation distortion.

CNN-DPD models trained on diverse signal conditions demonstrate robust cross-domain generalization. A single trained network can linearize across:

• Varying signal bandwidths (20 MHz to 400 MHz)
• Different modulation formats (QPSK to 256-QAM)
• Multiple average power levels and PAPR profiles
• Temperature and supply voltage variations

This reduces the need for extensive per-condition calibration and lookup table storage, simplifying deployment in dynamic 5G NR environments.

CNN-DPD benefits from mature deep learning optimization techniques including Adam optimizers, learning rate scheduling, and gradient clipping. The convolutional inductive bias provides a strong prior that accelerates convergence compared to fully-connected architectures. Typical training requires fewer iterations to reach target ACLR and EVM specifications. Batch normalization between layers mitigates internal covariate shift, enabling stable training even with high-PAPR waveforms that cause gradient variance in polynomial models.

A Volterra-based behavioral model that serves as the classical baseline against which CNN-DPD is benchmarked. GMP captures nonlinear memory effects through cross-terms between delayed samples and envelope powers.

• Incorporates lagging and leading envelope terms
• Effective for moderate bandwidths below 100 MHz
• Struggles with the high-order nonlinearities in GaN mmWave PAs
• CNN-DPD typically achieves 2-3 dB better ACLR than GMP at wide bandwidths

A training framework where the predistorter is identified by placing it after the PA model in the estimation loop. The CNN is trained to minimize the error between its output and the PA input.

• Avoids the need to compute an explicit inverse model
• Assumes the postdistorter equals the predistorter
• Widely used for offline CNN-DPD coefficient extraction
• Can suffer from noise bias in the presence of feedback path distortion

A linearization method that captures the combined nonlinear distortion of an entire mmWave phased array in the far-field. CNN-DPD excels here by learning beam-dependent distortion patterns.

• Addresses antenna crosstalk and active impedance mismatch
• Uses a single observation receiver in the far-field
• Eliminates per-element feedback chains
• CNN architectures can learn beam-index-dependent predistortion functions

A precursor architecture to CNN-DPD that processes I and Q components separately through tapped delay lines. RVTDNN introduced neural networks to DPD but lacks the hierarchical feature extraction of CNNs.

• Uses fully connected layers with time-delayed inputs
• Models memory effects through temporal context windows
• Requires explicit feature engineering of envelope terms
• CNN-DPD replaces manual feature design with learned convolutional filters

A recurrent neural network approach that models long-range temporal dependencies in PA behavior. LSTM-DPD competes with CNN-DPD for capturing extended memory effects.

• Excels at thermal and trapping memory spanning microseconds
• Sequential processing limits parallelization during inference
• CNN-DPD offers lower inference latency for real-time applications
• Hybrid CNN-LSTM architectures combine spatial and temporal feature extraction

An iterative training method that minimizes the error between the desired linear output and actual PA output. DLA directly optimizes the predistorter without the ILA's postdistorter assumption.

• More robust to measurement noise than ILA
• Requires a differentiable PA model for backpropagation
• CNN-DPD trained via DLA can adapt to changing operating conditions
• Enables online adaptation when combined with real-time coefficient updates