Deep Learning-Driven Inverse Design of Doherty Power Amplifiers Using Pixelated Combiners and Dual-State Impedance Synthesis

arXiv:2606.18395 · eess.SP, cs.AI, cs.AR, cs.SY, eess.SY · Submitted 2026-08-24 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Deep Learning-Driven Inverse Design of Doherty Power Amplifiers Using Pixelated Combiners and Dual-State Impedance Synthesis".

Jane: The paper was written by Han Zhou, Haojie Chang, David Widén and Christian Fager from Tampere University and Chalmers University of Technology.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and Authors: Tom: So, the core problem they are addressing in "Deep Learning-Driven Inverse Design of Doherty Power Amplifiers Using Pixelated Combiners and Dual-State Impedance Synthesis" is that designing the output combiner for a Doherty PA is just incredibly difficult because it has to handle so many conflicting functions.

Jane: It needs to manage load modulation, impedance matching, and phase compensation all in one network, which is a massive task for traditional iterative design methods.

Lu: The fact that they are using deep learning means we' are moving away from just brute force simulation sweeps toward a predictive model that should dramatically speed up the design process.

Meng: I agree with Lu; the idea of replacing exhaustive EM simulations with an AI surrogate is a massive practical win for development cycles.

Lalam: This is about shifting the paradigm, enabling us to envision designs that optimize not just for peak performance but also for long-term adaptability and cultural shifts toward computational efficiency.

Summary of Findings: Tom: Moving into the summary of results, the team has produced two GaN HEMT Doherty PA prototypes that are genuinely impressive in terms of performance metrics.

Jane: The fact that both achieved saturated output power exceeding forty-four point two dBm shows they have managed to push the limits of these circuits successfully within that two point six to two point eight GHz range.

Lu: What’s really striking is the drain efficiency, with peak efficiency hitting over seventy-one percent and a back-off efficiency reaching up to sixty-four percent, which is a huge validation of the design approach.

Meng: The measurable output power and high drain efficiency are exactly what engineers want to see; it suggests these results are not just theoretical but very tangible in the lab.

Lalam: Achieving this level of performance speaks to a cultural shift toward highly efficient, cutting-edge technology that empowers communication systems globally.

Improvements and Methodology: Tom: The key improvement lies in the methodology, specifically their dual-state impedance synthesis combined with deep CNN models.

Jane: They are not just guessing; they’ve created a detailed workflow where a deep convolutional neural network acts as a surrogate model to predict S-parameters incredibly quickly.

Lu: That CNN is then fed into a genetic algorithm, which is such an elegant way to explore the design space and evolve pixelated layouts that meet complex performance criteria.

Meng: From an engineering standpoint, the dual-state synthesis—separately targeting peak power versus back-off power—is what guarantees that this approach works for real-world Doherty operation.

Lalam: This is how we improve the culture of engineering; by using these advanced AI tools to ensure robust performance across a scalable and efficient design space, as seen in this method.

Conclusion and Wrap-up: Tom: So, we’ve covered a lot of ground—the challenges, the impressive results in "Deep Learning-Driven Inverse Design of Doherty Power Amplifiers Using Pixelated Combiners and Dual-State Impedance Synthesis," and the methodology is truly cutting edge.

Jane: The ability to achieve such high efficiency while maintaining excellent linearity under modulated signals with DPD is a huge step forward for communication technology.

Lu: I think the implications for advanced RF design, where AI guides every single pixel, are just staggering. It’s a new frontier in circuit architecture.

Meng: It feels like we're seeing the practical implementation of AI as an actual tool that solves problems rather than just as a theoretical concept; this is a working solution to complex RF challenges.

Lalam: We hope this research, the "Deep Learning-Driven Inverse Design of Doherty Power Amplifiers Using Pixelated Combiners and Dual-State Impedance Synthesis," inspires future builds toward better, more efficient global communication infrastructure.

Han Zhou, Haojie Chang, David Widén, Christian Fager

Tampere University · Chalmers University of Technology

eess.SP, cs.AI, cs.AR, cs.SY, eess.SY

Submitted: 2026-08-24

Updated: 2026-08-25

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: This paper presents a novel methodology for designing three-port Doherty power amplifier (PA) combiners by integrating deep learning with genetic algorithms.

Key concepts

Doherty Power Amplifier (DPA): A DPA is a complex circuit where the output combiner must manage multiple conflicting functions, including load modulation, impedance matching, and phase compensation. This is a massive challenge for traditional iterative design methods.
Deep Convolutional Neural Network (CNN) Surrogate Model: The CNN serves as an AI surrogate model that predicts S-parameters extremely quickly. This predictive model replaces exhaustive electromagnetic simulations, offering a practical win by dramatically speeding up the circuit development process.
Dual-State Impedance Synthesis: This is a specific methodology used in the design workflow. It involves separately targeting two critical operational states—peak power and back-off power. This ensures the resulting pixelated layouts are robust for real-world Doherty PA operation.

Terminology

Inverse Design — 8× in this paper · explained in 2 episodes
Genetic Algorithm — 7× in this paper
Convolutional Neural Networks (CNNs) — 2× in this paper · explained in 2 episodes
Accuracy — 1× in this paper · explained in 1 episode
Adam — 1× in this paper · explained in 2 episodes
Adaptability — 1× in this paper · explained in 1 episode
Core — 1× in this paper · explained in 2 episodes
Energy Efficiency — 1× in this paper · explained in 3 episodes
Initial Data — 1× in this paper · explained in 2 episodes
Loss Function — 1× in this paper · explained in 2 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

This paper presents a novel methodology for designing three-port Doherty power amplifier (PA) combiners by integrating deep learning with genetic algorithms. This approach addresses the highly challenging task of synthesizing networks that must simultaneously provide load modulation, impedance matching, and phase compensation, offering a faster and more efficient alternative to traditional iterative electromagnetic (EM) parameter sweeps.

The Design Challenge

The output combiner of a Doherty PA is a critical component that fundamentally governs Doherty PA behavior and ultimately sets the limits on achievable back-off efficiency. Traditional design methods rely on parameterized models with pre-selected topologies, which are often iterative, time-consuming, and prone to converging on local optima. While deep learning has been applied to simpler architectures like wideband class B PAs or harmonic-tuned class F PAs, applying pixelated layout techniques to the complex synthesis of three-port load-modulated networks has previously remained unexplored.

How it works

The researchers propose a dual-state impedance synthesis framework that utilizes a deep convolutional neural network (CNN) as a fast EM surrogate model. The methodology involves several key technical components:

((

A circuit layout is represented as a binary matrix where "1 indicates the presence of a metal pixel and 0" non-metal, discretized into a 15 × 15 grid.
A deep CNN architecture consisting of 12 convolutional blocks with residual network structures and six fully connected layers is trained on approximately 600K augmented samples to predict frequency-dependent S-parameters in sub-millisecond time.
A Genetic Algorithm (GA) is integrated with the CNN to perform dual-state impedance optimization, evolving pixelated layouts that meet specific targets for both peak and back-off power conditions.
The synthesis targets ensure that within the 2.6–2.8 GHz band, the real part of the impedance remains within ±10% of target values while minimizing the imaginary components.

)

Experimental Validation

To validate the framework, two GaN HEMT Doherty PA prototypes were designed and fabricated using a 20-mil Rogers 4350B substrate. The design process incorporated transistor parasitics and packaging elements to ensure accuracy. The authors evaluated the prototypes through three distinct testing methods:

((

Small-signal measurements to confirm gain and input return loss.
Large-signal continuous-wave (CW) measurements to evaluate drain efficiency and output power.
Modulated signal testing using a 20 MHz OFDM signal with a 7-dB PAPR to assess linearity via digital predistortion (DPD).

)

Key Results

The experimental results demonstrate that the proposed methodology produces high-performance amplifiers that compare favorably to state-of-the-art designs. Both prototypes achieved a measured saturated output power exceeding 44.2 dBm and peak drain efficiency levels above 71.2%. Specifically, at a 6-dB back-off level, the prototypes maintained high efficiency, with measured values ranging from 56.3% to 64.6%. Furthermore, after applying digital predistortion, each prototype achieved an adjacent channel leakage ratio (ACLR) better than −51.3 dBc, proving the strong linearizability of the synthesized designs under modulated signal testing.of the proposed deep learning-driven design methodology offers a scalable and efficient solution for advancing multi-port load-modulation RF PA architectures.

Improvements for AI systems

To improve existing AI systems in the field of Electronic Design Automation (EDA) and RF engineering based on this research, I propose the following specific architectural and functional enhancements:

Multi-State Objective Optimization via Dual-Constraint Loss Functions

The paper introduces a dual-state synthesis approach (Peak Power vs. Back-off). Current AI design tools often optimize for a single operating point (e.g., saturated power).

By implementing a multi-task learning framework where the loss function is weighted by the divergence from target impedances at multiple operating points simultaneously, an AI system can move beyond single-point optimization to trajectory-aware design. This allows the AI to synthesize components that maintain high efficiency across a wide dynamic range, rather than just at peak performance.

Hybrid Neural-Evolutionary Surrogate Modeling (CNN + GA Integration)

Most current surrogate models act as standalone predictors. I would integrate a Deep CNN surrogate directly into the inner loop of a Genetic Algorithm (GA) to create an Active Learning Design Environment.

The improved system would use the CNN to provide sub-millisecond gradient-free feedback to the GA, allowing it to navigate high-dimensional, non-convex design spaces (like pixelated layouts) without the prohibitive cost of full-wave EM simulations. This enables the AI to discover non-intuitive, non-human topological layouts that traditional iterative parameter sweeps would miss.

Pixelated Layout Representation for Geometric Topology Synthesis

Current AI for circuit design often relies on template-based or parameterized models (adjusting length/width of known components). I would implement the paper's pixelated binary matrix representation as a standard input format for generative models.

This allows an AI to perform true topology synthesis—designing the actual physical arrangement of metal pixels from scratch—rather than merely optimizing the dimensions of pre-defined components. This enables the discovery of complex, multi-port networks (like three-port combiners) that are difficult to model with standard analytical equations.

Physics-Informed Impedance Mapping

The paper demonstrates how to bridge S-parameters and Z-parameters through ABCD matrix conversion within the training pipeline. An improved AI system would incorporate Physics-Informed Neural Networks (PINNs) that enforce Kirchhoff’s laws and impedance matching constraints directly into the neural network's architecture.

This would ensure that the AI's predicted S-parameters are not just statistically likely, but physically realizable and consistent with fundamental electromagnetic principles, significantly reducing the reality gap between simulated AI designs and fabricated hardware.