Terminology » 6G PA Linearization : Memory Polynomials vs AI

6G PA Linearization : Memory Polynomials vs AI

memory polynomial power amplifier linearization ai 6g

Introduction : wireless communications, the Radio Frequency (RF) Power Amplifier (PA) is a critical component, but it is inherently non linear. To maximize energy efficiency, PAs are often driven near their saturation points. However, this introduces severe non linear distortions, causing out of band spectral regrowth (which interferes with adjacent channels) and in-band distortion (which degrades signal quality).

To counter this, engineers use Digital Predistortion (DPD), a technique that applies an inverse distortion to the signal in the digital domain before it reaches the PA, effectively canceling out the PA’s physical distortion.

As we move toward 6G, networks will rely on ultra-wide bandwidths, extremely high Peak to Average Power Ratios (PAPR) and highly dynamic operational modes like Integrated Sensing and Communication (ISAC). These 6G conditions induce severe memory effects; where the PA’s current output is affected not just by the current input, but by the history of past inputs due to thermal and electrical trapping. Linearizing PAs under these extreme conditions requires looking beyond traditional mathematical models.

Traditional Memory Polynomials

For the past several generations of wireless technology (4G and 5G), Memory Polynomials (MPs) and Generalized Memory Polynomials (GMPs) have been the industry standard for DPD.

Memory Polynomials are derived from the Volterra series. They use rigid mathematical formulas to model the PA’s behavior by considering both the current signal sample and a finite number of past signal samples (memory taps). By calculating the appropriate coefficients for these polynomials, the system can map the inverse non-linearity of the amplifier.

Drawbacks for 6G:

  • While Memory Polynomials are highly effective for narrower bandwidths, they suffer from the “curse of dimensionality” when applied to 6G. As the signal bandwidth widens into the gigahertz range, the memory effects become “deeper.” To accurately model this using polynomials, the number of required coefficients and cross-terms explodes. This exponential increase in mathematical complexity makes real-time calculation too slow and power-hungry for practical 6G deployments.
  • Moreover this technique struggles to adapt quickly to rapid environmental or configuration changes, such as the rapid switching between communication and radar sensing modes expected in 6G ISAC systems.

AI-Based Neural Networks

To overcome the mathematical bottlenecks of polynomials, the industry is shifting toward Artificial Intelligence (AI) and Machine Learning (ML) for 6G PA linearization.

Instead of relying on predefined mathematical equations, Neural Network-based DPD uses architectures like Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), or Long Short Term Memory (LSTM) networks. These models are fed massive datasets of input output signal pairs from the PA. Through training, the neural network independently “learns” the complex, inverse non-linear behavior of the amplifier.

Benefits:

  • Neural Networks are universal function approximators. They excel at mapping highly complex, non-linear relationships and deep memory effects across ultra wide bandwidths without the exponential scaling of coefficients that plagues polynomial models.
  • NNs are also highly adaptable; a trained model can dynamically adjust to changing load impedances, fluctuating power supplies (such as in Envelope Tracking), or mode switching in ISAC systems.
  • Once trained, the inference (the actual application of the predistortion) can be executed efficiently on modern FPGAs or dedicated AI hardware accelerators.

Key differences

Feature/AttributeTraditional Memory polynomialsAI based Neural Networks
Core MechanismPredetermined mathematical formulas (Volterra based) with extracted coefficients.Data driven architecture that “learns” the inverse distortion from input/output training data
Handling of Memory EffectsUses specific delay taps; struggles mathematically with “deep” or long term memory effects.Naturally captures complex, deep, and long-term memory effects (especially models like LSTMs)
6G Bandwidth ScalabilityPoor; Suffers from the “curse of dimensionality.” Wider bandwidths require an exponential, unmanageable increase in coefficients.Excellent; Handles ultra-wide bandwidths and extreme non-linearities efficiently without exponential coefficient explosion.
AdaptabilityRigid; typically requires the calculation of an entirely new set of coefficients if the PA’s operating mode or environment changes.Highly dynamic; can adjust in real-time to changing variables like load mismatch, temperature shifts, or ISAC mode switching.

Summary: As 6G pushes power amplifiers into ultra wide bandwidths and extreme non linearities, traditional memory polynomials are hitting their mathematical limits due to an unmanageable explosion of required coefficients. AI based neural networks provide a much needed breakthrough, offering a dynamic, data driven approach that efficiently models deep memory effects and adapts in real time to changing RF environments.