Deep Learning-Driven Black-Box Doherty Power Amplifier with Pixelated Output Combiner and Extended Efficiency Range

This paper presents a deep learning-driven inverse design methodology that utilizes a CNN-based electromagnetic surrogate model and genetic algorithm to synthesize Doherty power amplifiers with pixelated output combiners, achieving high efficiency and linearity across extended back-off ranges in fabricated GaN prototypes.

The Gist

The Big Picture: The "Smart Chef" and the "Pixelated Puzzle"

Imagine you are trying to build a high-performance engine for a race car (in this case, a Power Amplifier for your 5G phone or Wi-Fi router). This engine needs to be incredibly efficient, especially when the car is cruising at a steady speed (low power) rather than speeding at the limit (peak power).

For decades, engineers have used a specific engine design called a Doherty Amplifier. Think of this like a hybrid car: it has a main engine that runs most of the time and a backup engine that kicks in only when you need a burst of speed. The tricky part is the combiner—the mechanical gearbox that connects these two engines to the wheels. If the gearbox is designed poorly, energy is wasted as heat, and the car gets sluggish.

The Problem: Designing this "gearbox" usually requires engineers to guess, simulate, and tweak tiny wires and metal shapes over and over again. It's like trying to solve a maze by walking every single path until you find the exit. It takes forever, and you might miss the perfect path.

The Solution: This paper introduces a Deep Learning "Smart Chef" that can instantly taste a dish (simulate the physics) and tell you exactly how to cook it, without you having to try every recipe.

---

Key Concepts Explained

1. The "Pixelated" Puzzle (The Canvas)

Traditionally, engineers design the gearbox using standard shapes: straight lines, circles, and curves (like building with LEGOs).

• The Innovation: This team decided to treat the design area like a digital image or a Minecraft world. They broke the space down into a tiny grid of squares (pixels), like a 15x15 checkerboard.
• The Rule: Each square can either be Metal (1) or Empty Space (0).
• Why? This creates a massive number of possible shapes. Instead of just a few standard LEGO structures, they can create millions of unique, organic, and complex shapes that a human would never think to draw. It's like going from building with LEGOs to sculpting with clay.

2. The "Crystal Ball" (The Deep Learning Model)

To find the best shape, you usually have to run a super-computer simulation for every single possibility. If you have millions of shapes, this would take years.

• The Innovation: The team trained a Deep Convolutional Neural Network (CNN). Think of this as a Crystal Ball or a Super-Intelligent Apprentice.
• How it works: They showed the AI thousands of examples of "Pixel Puzzles" and what their electrical performance was. The AI learned the patterns. Now, if you show it a new puzzle, it can predict the result in a fraction of a second instead of hours. It acts as a "surrogate model," skipping the heavy math and giving a near-perfect guess instantly.

3. The "Evolutionary Search" (The Genetic Algorithm)

Now that they have the AI, how do they find the best shape?

• The Innovation: They used a Genetic Algorithm (GA). Imagine a game of "Hot and Cold" played by evolution.
  1. The AI generates 4,000 random pixel puzzles.
  2. It asks the "Crystal Ball" (the trained AI) to predict how they perform.
  3. It keeps the top 10 performers (the "fittest" ones).
  4. It "breeds" them: It takes the top half of one puzzle and the bottom half of another to create new "offspring" puzzles.
  5. It occasionally makes a random "mutation" (flipping a pixel from metal to empty) to try something new.
  6. It repeats this process, evolving better and better designs until it finds the perfect one.

4. The "Black Box" Approach

Usually, engineers design the amplifier and the gearbox separately, then try to glue them together.

• The Innovation: This paper uses a "Black Box" method. They don't care about the internal wires of the amplifier; they just look at the "Input" (what the transistor wants) and the "Output" (what the antenna needs). They let the AI design the entire gearbox from scratch to perfectly match those two points. It's like hiring an architect to design a bridge between two cliffs without worrying about the specific type of steel used, as long as the bridge holds.

---

The Results: What Did They Build?

The team built two physical prototypes (real circuit boards) to prove this works.

• The Test: They used these amplifiers with a signal that mimics a busy 5G network (lots of data, high peaks and valleys).
• The Performance:
  • Efficiency: Even when the signal was weak (9 dB "back-off," which is like driving at 30 mph instead of 100 mph), the amplifiers were still incredibly efficient (over 50%). Usually, amplifiers waste a lot of energy at low power. These didn't.
  • Power: They pumped out a lot of power (over 44 dBm), which is strong enough to cover a large area.
  • Cleanliness: After applying a digital "clean-up" filter (called DPD), the signal was so clean that it didn't interfere with neighboring channels. It was like a singer hitting the perfect note without any squeaks or cracks.

Why Does This Matter?

1. Saves Energy: Mobile networks consume huge amounts of electricity. If amplifiers are more efficient, we save money and reduce carbon emissions.
2. Faster Design: What used to take months of trial-and-error now takes days (or even hours) because the AI does the heavy lifting.
3. Better Performance: By exploring shapes humans wouldn't think of (the pixelated approach), they found solutions that are better than traditional designs.

Summary Analogy

Imagine you need to find the fastest route through a giant, foggy city to get to a destination.

• Old Way: A human driver tries one street, gets stuck, turns around, tries another. It takes all day.
• This Paper's Way:
  1. You give a Super-Computer (AI) a map of the whole city (the pixelated grid).
  2. The AI instantly simulates millions of routes in a second.
  3. An Evolutionary Engine picks the best routes, mixes them, and improves them over and over.
  4. In minutes, it hands you the perfect, fastest route that no human driver would have ever discovered.

The authors successfully used this method to build a "super-efficient" engine for 5G networks, proving that AI can design better hardware than humans can alone.

Technical Summary

1. Problem Statement

Modern communication systems (e.g., 5G) utilize advanced modulation schemes that result in high Peak-to-Average Power Ratios (PAPR). This necessitates Power Amplifiers (PAs) that maintain high efficiency not just at saturation, but also at deep power back-off levels to ensure energy sustainability.

• Limitations of Traditional Doherty PAs: While Doherty PAs are widely used for their back-off efficiency, their design relies on complex, cascaded networks (matching sections, offset lines, impedance transformers). Traditional design methods use "bottom-up" approaches with predefined topologies (lumped or distributed elements) and iterative electromagnetic (EM) simulations.
• Design Bottlenecks: These conventional methods restrict the design space to specific topologies, require extensive simulation time, and often fail to discover globally optimal solutions for complex three-port combiner networks, especially when using fully symmetrical transistors.
• The Gap: There is a lack of inverse design methodologies capable of synthesizing complex, non-standard Doherty combiner structures that can fully exploit the capabilities of modern transistors without being constrained by manual topology selection.

2. Methodology

The authors propose a Deep Learning-Driven Inverse Design Framework that combines a "black-box" Doherty synthesis approach with a pixelated output combiner network.

A. Black-Box Doherty Theory

• Instead of designing individual sub-networks, the authors derive the required two-port impedance matrix ($Z_{2P}$) directly from transistor load-pull data.
• Using the main and auxiliary amplifier load-pull data (optimal impedances at peak and back-off power), they calculate the necessary phase delay ($\theta$) and impedance parameters to achieve a target back-off level (specifically 9 dB).
• This $Z_{2P}$ is mathematically transformed into a lossless three-port combiner requirement, serving as the target specification for the inverse design.

B. Pixelated Layout Design Space

• The output combiner is not designed with standard transmission lines but is represented as a $15 \times 15$ binary pixel grid ($225$ pixels total).
• Each pixel is either "1" (metal) or "0" (dielectric). This creates a massive design space ($2^{225}$ possible structures), allowing for the discovery of non-intuitive, compact, and highly efficient topologies.
• The physical size is approximately $27.36 \text{ mm} \times 27.36 \text{ mm}$.

C. Deep Learning Surrogate Model (CNN)

• To navigate the vast design space without running time-consuming full-wave EM simulations for every candidate, a Deep Convolutional Neural Network (CNN) is trained as an EM surrogate model.
• Architecture: A ResNet-style architecture with 12 convolutional layers and 6 fully connected layers.
• Training: The model is trained on a dataset of ~77,000 simulated pixelated structures (generated via Keysight ADS Momentum). It maps the binary pixel layout (input) to the S-parameters (output).
• Performance: The trained model achieves a Mean Absolute Error (MAE) of 0.068, enabling rapid prediction of S-parameters in milliseconds.

D. Genetic Algorithm (GA) Optimization

• A Genetic Algorithm is used to perform the inverse design.
• Process:
  1. An initial population of random binary matrices is generated.
  2. The trained CNN predicts the S-parameters for each candidate.
  3. A fitness function calculates the deviation between the predicted S-parameters and the target S-parameters (derived from the black-box theory).
  4. Through selection, crossover, and mutation, the population evolves toward the optimal pixelated layout that satisfies the target impedance behavior.

3. Key Contributions

1. First Pixelated Doherty Combiner: This work presents the first Doherty PA design utilizing a pixelated output combiner synthesized via deep learning inverse design.
2. Black-Box + Deep Learning Integration: It successfully integrates a black-box synthesis theory (using load-pull data) with a top-down pixelated inverse design, eliminating the need for predefined circuit topologies.
3. Symmetrical Device Utilization: The method enables extended back-off efficiency using fully symmetrical Main and Auxiliary amplifiers (identical transistors), simplifying the design compared to asymmetric approaches.
4. Efficiency Acceleration: The CNN surrogate model replaces thousands of hours of EM simulation with seconds of inference, making the exploration of a $2^{225}$ design space computationally feasible.

4. Experimental Results

Two prototypes were fabricated using GaN HEMT transistors (CG2H40010F) on Rogers 4350B substrates.

• Continuous Wave (CW) Performance (at 2.75 GHz):

  • Prototype 1: Peak Drain Efficiency (DE) > 74%, Output Power > 44.1 dBm. At 9-dB back-off, DE remained 55%.
  • Prototype 2: Peak DE > 76%, Output Power > 44.8 dBm. At 9-dB back-off, DE remained 51%.
  • Both prototypes maintained >52% efficiency at 9-dB back-off across the 2.65–2.85 GHz band.

• Modulated Signal Performance (20 MHz 5G NR-like, 9-dB PAPR):

  • After applying Digital Predistortion (DPD):
    • Average PAE: > 51%.
    • ACLR: Better than −60.8 dBc.
    • EVM: < 1.2%.
  • This demonstrates that the complex pixelated structures do not compromise linearity when properly linearized.

• Comparison: The prototypes outperform or match state-of-the-art symmetrical Doherty PAs and Load-Modulated Balanced Amplifiers (LMBAs) in terms of back-off efficiency and compactness (normalized size).

5. Significance

• Paradigm Shift in RF Design: This paper validates that deep learning-driven inverse design can move beyond simple filters or splitters to complex, active RF front-end components like Doherty PAs.
• Design Freedom: By removing topological constraints, the method discovers solutions that human designers might overlook, leading to more compact and efficient circuits.
• Scalability: The framework is scalable to other frequencies and transistor technologies, offering a pathway to automate the design of high-efficiency PAs for future 6G and beyond communication systems.
• Practical Viability: The successful fabrication and measurement of two distinct prototypes (showing different pixelated layouts achieving the same target) prove that the method is robust and reproducible in a real-world manufacturing environment.