Machine Learning-Enabled Inverse Design Platform for Synthesizing Broadband, High-Efficiency Millimeter-Wave Power Amplifiers

Background

Millimeter-wave (mmWave) and radio frequency (RF) power amplifiers are critical components in modern wireless communication systems, particularly as the industry advances toward 5G and future 6G networks. These systems demand front-end components capable of supporting extremely wide bandwidths—often spanning multiple frequency bands simultaneously—to accommodate high data rates and flexible spectrum allocation. Achieving high efficiency and linearity across such broad frequency ranges is essential for reliable, high-speed data transmission, carrier aggregation, and spectrum sharing. However, the design of broadband matching networks, which are responsible for efficiently transferring power from the amplifier to the antenna over wide frequency spans, remains a significant challenge. As wireless standards push into higher frequencies and wider bandwidths, the need for innovative matching network solutions that can deliver both broadband performance and high efficiency becomes increasingly urgent.

Traditional approaches to designing matching networks for mmWave and RF power amplifiers typically rely on optimizing a limited set of predefined circuit topologies, such as cascaded transmission lines or lumped element networks. These methods restrict the designer to a narrow subset of possible geometries, often resulting in sub-optimal performance—especially in terms of bandwidth, loss, and the ability to support multi-band operation. Moreover, the optimization process is computationally intensive, as it requires repeated full-wave electromagnetic (EM) simulations for each design iteration, which can take hours or even days. This makes it impractical to explore the vast space of potential network configurations, particularly for irregular or non-intuitive geometries that might offer superior performance. As a result, current solutions struggle to meet the demands of next-generation wireless systems, leaving a gap for more efficient, flexible, and scalable design methodologies.

Technology

This technology leverages a machine learning-enabled inverse design methodology to synthesize broadband, high-efficiency matching networks for millimeter-wave and radio frequency power amplifiers. The approach discretizes the matching network into a pixel grid, where each pixel represents either metal or void, enabling the exploration of an enormous design space with (2^{n \times n}) possible configurations. A deep convolutional neural network (CNN) acts as a rapid electromagnetic emulator, predicting the RF performance of these pixelated structures across wide frequency ranges, which dramatically accelerates the design process compared to traditional full-wave electromagnetic simulations. The methodology employs evolutionary algorithms for fast optimization, achieving near-optimal broadband, low-loss solutions suitable for a variety of topologies, including multi-port combiners and Doherty networks. A prototype fabricated in 90 nm SiGe technology demonstrated exceptional performance, achieving a 30-94 GHz saturated output power 3 dB bandwidth and supporting concurrent multi-band high-speed data modulation.

What differentiates this technology is its unprecedented ability to explore and optimize a vast, previously inaccessible design space, thanks to the pixel-based discretization and the use of deep learning for rapid electromagnetic prediction. Unlike conventional methods that rely on fixed, template-based topologies and time-intensive simulations, this solution enables the synthesis of irregular, non-intuitive geometries that maximize broadband efficiency and minimize loss. The CNN-based EM emulator provides accurate, wideband predictions in minutes, facilitating rapid iterative optimization and enabling functionalities such as concurrent multi-band operation and carrier aggregation. The approach is scalable, adaptable to various circuit architectures, and supports both integrated and PCB implementations. Its demonstrated record bandwidth, low insertion loss, and flexible topology make it a powerful tool for next-generation wireless communication systems, setting a new benchmark for performance and design agility in RF and mmWave power amplifier networks.

Benefits

• Enables synthesis of broadband, high-efficiency matching networks with over 100% fractional bandwidth (30-94 GHz) for mmWave power amplifiers.
• Utilizes deep learning (CNN) as a fast electromagnetic emulator, drastically reducing design optimization time from hours/days to minutes.
• Explores an enormous design space (2^(n×n) configurations) allowing discovery of irregular, non-intuitive geometries for near-optimal RF performance.
• Supports concurrent multi-band operation, enabling simultaneous multi-frequency transmission and carrier aggregation for enhanced throughput.
• Achieves low insertion loss (0.55-1 dB) and high power-added efficiency (up to 24.7%) across wide frequency ranges.
• Flexible and scalable design methodology applicable to various matching network topologies including multi-port combiners and Doherty networks.
• Demonstrated practical implementation in compact silicon technology with validated high-speed data modulation and multi-band capabilities.
• Provides a novel paradigm surpassing traditional fixed-template designs by integrating machine learning with electromagnetic theory for next-generation wireless systems.

Applications

• 5G/6G mmWave RF front-ends
• Broadband phased array transceivers
• Concurrent multi-band wireless systems
• High-speed mobile data modems
• Spectrum sharing communication platforms

Additional info

This methodology employs machine learning-enabled inverse design for synthesizing broadband, high-efficiency matching networks in millimeter-wave power amplifiers. It discretizes the physical network into a pixel grid, utilizing a deep convolutional neural network as a rapid electromagnetic emulator to predict RF performance. Evolutionary algorithms then optimize these pixelated structures, achieving record bandwidth and concurrent multi-band capabilities.