Physics-Informed DPD

Integrates Volterra series kernels and memory polynomial structures directly into the neural network architecture. Rather than learning amplifier behavior from scratch, the model is constrained by known physical laws—including AM/AM and AM/PM characteristics—ensuring predictions remain physically plausible even when extrapolating beyond training data. This hard-coding of domain knowledge prevents the network from learning spurious correlations that violate amplifier physics.

Reduces training data requirements by 60-80% compared to purely data-driven neural DPD approaches. By encoding the underlying differential equations governing electron transport in GaN HEMTs and thermal memory dynamics, the model generalizes to unseen signal conditions—varying bandwidths, PAPR levels, and carrier configurations—without requiring exhaustive retraining. Critical for Doherty amplifier architectures where load modulation creates complex nonlinear operating regimes.

Explicitly models S-parameter coupling matrices and active impedance mismatch as physical priors within the learning framework. The network incorporates knowledge of how beamforming weight changes alter the impedance seen by each PA element, enabling joint compensation of:

• Antenna mutual coupling effects
• Cross-coupling distortion between adjacent elements
• Load modulation from dynamic beam steering This eliminates the need for per-beam DPD coefficient tables.

Incorporates electro-thermal models as physics-based regularizers that capture both short-term and long-term memory effects. The neural network learns residual corrections around a known thermal impedance model, accurately predicting how self-heating in GaN power amplifiers alters gain and phase response over time. This hybrid approach outperforms pure black-box models for signals with high peak-to-average power ratios that induce rapid thermal transients.

Enables online training algorithms that update only the residual neural network weights while keeping the physics-based backbone fixed. This dramatically reduces computational complexity for closed-loop adaptation—the physical model handles the bulk nonlinearity, while the neural component learns small corrections for:

• Aging and drift effects
• Environmental temperature variations
• Manufacturing variances across array elements Adaptation converges in fewer iterations than fully learned approaches.

Unlike purely data-driven models that fail catastrophically outside their training distribution, physics-informed DPD maintains linearization performance when encountering novel signal conditions. The embedded amplifier physics—including gain compression curves, saturation behavior, and harmonic generation mechanisms—provides a robust inductive bias. This is essential for multi-band DPD architectures where concurrent transmission creates nonlinear interaction products not present in single-band training data.

The mathematical foundation for physics-informed DPD. Behavioral models capture the nonlinear relationship between input and output signals without requiring detailed semiconductor physics.

• Volterra Series: The most general polynomial model with memory
• Memory Polynomial: A simplified, widely-used subset of Volterra
• Generalized Memory Polynomial (GMP): Adds cross-terms for improved accuracy

These models provide the inductive bias that physics-informed neural networks use to constrain their learning space.

The primary physical phenomenon that makes MIMO DPD challenging. As beamforming weights change, the load impedance seen by each power amplifier varies dynamically.

• Causes PA nonlinear behavior to shift with beam angle
• Creates a non-stationary distortion problem
• Physics-informed models encode this impedance relationship directly

Understanding this mechanism is critical for designing DPD that generalizes across beam states.

An adaptive linearization strategy that explicitly compensates for time-varying load impedance. This is the direct application of physics-informed principles.

• Models the PA as a function of both input signal and instantaneous load
• Uses S-parameters or coupling matrices as physical priors
• Reduces the dimensionality of the learning problem

Load modulation DPD is often the practical implementation of physics-informed approaches in beamforming arrays.

Signal processing that mitigates antenna mutual coupling—the electromagnetic interaction where energy from one element induces currents in neighbors.

• Alters individual element impedance and radiation patterns
• Creates spatial distortion not present in isolated PAs
• Physics-informed DPD incorporates coupling matrices as known priors

Jointly addressing coupling and nonlinearity is a key advantage of physics-informed approaches over purely data-driven methods.

Deep learning approaches that learn the inverse PA transfer function. Physics-informed DPD enhances these by embedding domain knowledge into the architecture or loss function.

• Standard NN DPD: Pure data-driven, requires extensive training
• Physics-Informed NN DPD: Augments loss with physical constraints
• Hybrid Models: Combines analytical basis functions with learned corrections

The physics-informed variant achieves superior generalization with fewer training samples.

The mathematical techniques used to extract DPD parameters. Physics-informed approaches modify these algorithms to operate within physically plausible parameter spaces.

• Least Squares (LS): Batch estimation minimizing error norm
• Recursive Least Squares (RLS): Adaptive, sample-by-sample updates
• Regularized LS: Adds physics-based constraints to prevent overfitting

Constraining coefficient search to physically realizable values improves robustness against measurement noise.