Author: Denis Avetisyan
A new deep learning approach streamlines the design of Doherty power amplifiers, achieving extended efficiency ranges and improved performance.
This work presents a deep learning-driven methodology for the inverse design of Doherty power amplifiers with pixelated output combiners and demonstrates significant gains in RF circuit efficiency through load modulation.
Achieving both high efficiency and extended bandwidth remains a persistent challenge in power amplifier design. This is addressed in ‘Deep Learning-Driven Black-Box Doherty Power Amplifier with Pixelated Output Combiner and Extended Efficiency Range’, which introduces a deep learning methodology for the inverse design of Doherty power amplifiers incorporating pixelated output combiners. By leveraging a convolutional neural network as a surrogate model within a genetic algorithm framework, the authors demonstrate the synthesis of highly efficient designs exceeding 74% drain efficiency and delivering over 44.1 dBm at 2.75 GHz, while maintaining over 52% efficiency at a 9-dB back-off power level. Could this approach unlock a new paradigm for rapidly prototyping and optimizing RF power amplifiers for increasingly complex communication systems?
Decoding Efficiency: The Challenges of Modern Power Amplifiers
The relentless push for faster data speeds and increased network capacity in 5G communications presents a significant challenge to power amplifier (PA) design. These amplifiers, critical components in transmitting wireless signals, must now deliver both high efficiency – minimizing wasted energy and heat – and exceptional linearity. Linearity ensures signal fidelity, preventing distortion that degrades data quality and can lead to connection errors. Achieving both simultaneously is difficult, as improvements in one area often come at the expense of the other; however, without both characteristics, 5G networks cannot reliably support the bandwidth-intensive applications and massive device connectivity that define their potential. Consequently, research focuses intently on novel PA architectures and design techniques to overcome these limitations and realize the full promise of 5G technology.
Doherty power amplifiers have long been favored for their potential to boost efficiency in radio frequency transmission, a crucial factor in extending battery life for mobile devices and reducing energy consumption in base stations. However, realizing this efficiency gain isn’t straightforward. The amplifier’s architecture, relying on carefully tuned impedance matching networks and the precise biasing of multiple transistors, introduces significant design complexity. Optimizing these parameters for the wide range of new signals employed in contemporary 5G communication-with their varying modulation schemes and carrier frequencies-presents a formidable challenge. Traditional Doherty designs, while effective for established waveforms, often require substantial re-engineering and painstaking adjustments to maintain both high efficiency and the required signal fidelity when confronted with these novel and increasingly complex signals.
The pursuit of highly efficient power amplifier (PA) designs is significantly hampered by the computational burden of accurate modeling. Simulating PA behavior-accounting for nonlinearities, impedance matching, and the complex interplay of multiple frequencies-demands substantial processing power and time. Traditional techniques often rely on harmonic balance or time-domain simulations, which, while precise, scale poorly with increasing signal bandwidth and component complexity. Each design iteration, requiring numerous simulations to optimize performance metrics like power-added efficiency and linearity, can take hours or even days. This slow turnaround hinders exploration of the vast design space, limiting the potential for innovative PA architectures and preventing timely responses to the evolving demands of modern communication systems. Consequently, engineers are often forced to compromise between accuracy and speed, potentially leading to suboptimal designs and reduced overall system performance.
Surrogate Modeling: Accelerating Design with Deep Learning
Deep learning techniques provide a methodology for constructing surrogate models that approximate the behavior of Power Amplifiers (PAs) with significantly reduced computational cost. Traditional PA design relies heavily on electromagnetic simulations, which are resource-intensive and time-consuming, especially when exploring a large design space. Deep learning models, trained on data generated from these simulations, can learn the complex relationships between design parameters and PA performance metrics. This allows for the creation of a computationally efficient model – the surrogate – capable of predicting PA behavior without requiring repeated full electromagnetic simulations. The resulting reduction in computational burden enables faster design iterations, automated optimization, and real-time performance analysis.
A Convolutional Neural Network (CNN) functions as a surrogate model by learning the complex relationship between input design parameters and the resulting S-parameters of a Power Amplifier (PA). Traditional electromagnetic simulations, used to determine S-parameters, are computationally intensive and can require significant processing time. The CNN, once trained on a dataset of PA designs and their corresponding simulated S-parameters, can predict S-parameters for new designs with substantially reduced computational cost. This prediction capability stems from the CNN’s ability to extract hierarchical features from the input design, effectively mimicking the behavior modeled by full-wave electromagnetic solvers. The accuracy of the CNN surrogate model is dependent on the size and quality of the training dataset, and validation against independent simulation data is crucial to ensure reliable predictions.
The implementation of a deep learning-based surrogate model enables a significant reduction in computational time associated with power amplifier (PA) design space exploration. By accurately predicting PA performance characteristics – such as S-parameters, gain, and linearity – based on variations in design parameters, the surrogate model circumvents the need for repeated, resource-intensive electromagnetic simulations. This accelerated evaluation process facilitates rapid prototyping and optimization, allowing designers to efficiently investigate a wider range of design options and identify configurations that meet specific performance targets with greater speed and reduced computational cost. Consequently, design cycles are shortened, and the potential for improved PA performance is maximized.
Genetic Algorithms: Optimizing Architecture Through Iteration
A Genetic Algorithm (GA) was implemented to optimize the design of the pixelated combiner network within the Doherty Power Amplifier (PA) architecture. This optimization process involves iteratively refining network parameters – specifically, the combiner’s pixel arrangement and impedance matching – to maximize PA performance. The GA operates by defining a population of potential network designs, evaluating each design’s performance via electromagnetic simulation, selecting the fittest designs for reproduction, and applying genetic operators – crossover and mutation – to generate a new population. This cycle repeats until a design meeting the specified performance goals is achieved, automating a process traditionally reliant on manual tuning and expert intuition.
The Genetic Algorithm (GA) operates by cyclically modifying the parameters of the pixelated combiner network and evaluating the resulting performance predictions generated by the Convolutional Neural Network (CNN) surrogate model. Each iteration involves generating a population of network configurations, assessing their predicted performance via the CNN – eliminating poorly performing designs – and then applying genetic operators such as crossover and mutation to create a new generation of parameters. This process leverages the CNN’s ability to rapidly estimate performance, avoiding computationally expensive full-wave electromagnetic simulations for each parameter set, and guides the GA towards optimal network configurations based on the CNN’s predictions as the fitness function.
Implementation of the automated optimization process, leveraging a genetic algorithm and CNN surrogate model, yielded a Doherty power amplifier design demonstrating a peak drain efficiency of 76%. This performance was achieved concurrently with an output power exceeding 44.8 dBm at a carrier frequency of 2.75 GHz. These results indicate a substantial improvement in PA efficiency and output capability through the application of the described automated design methodology.
High Performance and Spectral Purity: A Synergistic Approach
The pursuit of both high performance and signal fidelity in power amplifiers (PAs) is addressed through an optimized Doherty architecture, skillfully combining load modulation techniques with digital predistortion (DPD). This approach allows the PA to operate with enhanced linearity, minimizing signal distortion across a broad bandwidth, while simultaneously maximizing power efficiency. Load modulation dynamically adjusts the impedance seen by the amplifier, optimizing its performance for varying signal levels. Complementing this, DPD actively compensates for remaining nonlinearities introduced by the PA, effectively “pre-correcting” the signal to achieve a cleaner output. The synergistic effect of these technologies results in a PA capable of delivering substantial power with minimal distortion, crucial for demanding applications like 5G New Radio (NR) communication systems where spectral purity is paramount.
Digital predistortion (DPD) proves highly effective in refining signal transmission by actively counteracting inherent distortions within the power amplifier. This technique demonstrably enhances signal quality, as evidenced by achieved metrics of -61.1 dBc for Adjacent Channel Leakage Ratio (ACLR) and 1.2% for Error Vector Magnitude (EVM) when processing a 20 MHz 5G New Radio-like signal. These figures represent significant improvements in spectral purity and signal fidelity, minimizing interference with neighboring channels and ensuring reliable data conveyance – critical factors for the performance of modern wireless communication systems. The precision of DPD allows for a cleaner signal output, contributing to more efficient use of bandwidth and improved overall network capacity.
The power amplifier (PA) design prioritizes both high efficiency and signal fidelity, demonstrably maintaining greater than 51% efficiency even at a substantial 99 dB back-off level – a critical metric for minimizing energy waste in modern communication systems. This sustained performance is further bolstered by fabrication utilizing the Rogers 4350B substrate, a material chosen for its low loss tangent and consistent dielectric properties. These characteristics are essential for preserving signal integrity, especially when transmitting the complex waveforms inherent in 5G New Radio (NR) signals, and contribute to the PA’s overall reliability and consistent operation within the demanding parameters of contemporary wireless infrastructure.
The presented methodology leverages deep learning not merely as an optimization tool, but as a means to fundamentally reshape the design process of Doherty power amplifiers. This approach mirrors a broader philosophical tenet – that understanding a system necessitates exploring its underlying patterns. The pixelated output combiner, conceived through inverse design, exemplifies this; it’s not simply a component added to an existing architecture, but a reconfiguration born from analyzing the data. As Mary Wollstonecraft stated, “It is time to revive the dormant energy of woman, and to wake her to a sense of her own powers.” Similarly, this work awakens dormant potential within RF circuit design by allowing data to guide the creation of novel structures and extended efficiency ranges, moving beyond conventional limitations through rigorous analysis and creative hypothesis – a true exploration of system patterns.
Beyond the Black Box
The demonstrated success of deep learning in navigating the Doherty power amplifier design space, while promising, merely shifts the locus of complexity. The current methodology excels at finding solutions – at optimization – but provides limited insight into why those solutions function as they do. Future work must prioritize explainability; a truly robust design process requires understanding the underlying relationships between network architecture, circuit parameters, and performance metrics, not simply achieving a high efficiency number. The pixelated combiner, while effectively modulating load impedance, presents a design challenge ripe for further exploration-specifically, how to move beyond empirical optimization towards analytically predictable behavior.
A persistent limitation lies in the dependence on training data generated through conventional electromagnetic simulation. This introduces inherent biases and restricts the exploration of truly novel topologies. A compelling avenue for future research involves integrating physics-informed neural networks, allowing the deep learning model to directly incorporate fundamental electromagnetic principles. This could unlock designs beyond the reach of current simulation techniques and, crucially, offer a degree of generalization currently absent from purely data-driven approaches.
Ultimately, the field must acknowledge that performance metrics – efficiency, linearity, bandwidth – are merely proxies for desired system behavior. A more holistic approach will focus on optimizing for specific application requirements, rather than maximizing isolated parameters. This necessitates a shift from ‘black box’ optimization to a framework where deep learning acts as an intelligent assistant, guiding the designer towards solutions that are both performant and fundamentally understandable.
Original article: https://arxiv.org/pdf/2603.16565.pdf
Contact the author: https://www.linkedin.com/in/avetisyan/
See also:
- Spotting the Loops in Autonomous Systems
- Seeing Through the Lies: A New Approach to Detecting Image Forgeries
- Staying Ahead of the Fakes: A New Approach to Detecting AI-Generated Images
- Julia Roberts, 58, Turns Heads With Sexy Plunging Dress at the Golden Globes
- Palantir and Tesla: A Tale of Two Stocks
- TV Shows That Race-Bent Villains and Confused Everyone
- The Glitch in the Machine: Spotting AI-Generated Images Beyond the Obvious
- How to rank up with Tuvalkane – Soulframe
- Gold Rate Forecast
- The 25 Marvel Projects That Race-Bent Characters and Lost Black Fans
2026-03-19 00:08