Machine‑learning surrogates speed RF hardware design and cut power losses: two GaN‑HEMT Doherty amplifiers found by a CNN‑guided inverse‑design pipeline delivered >74% peak drain efficiency and >51% average PAE under 5G‑like signals while satisfying linearity requirements.

Main Finding

A deep-learning-driven inverse design pipeline—using a CNN electromagnetic (EM) surrogate of pixelated passive networks combined with a genetic-algorithm blackbox optimizer—enables practical synthesis and fabrication of multi-port pixelated Doherty output combiners. Two fabricated GaN‑HEMT Doherty PA prototypes realized the designs and validated the approach experimentally, achieving >74% peak drain efficiency, >44.1 dBm output power at 2.75 GHz, and >52% drain efficiency at 9 dB back-off; under a 20 MHz 5G‑NR-like waveform (9 dB PAPR) with digital predistortion, each prototype reached >51% average PAE while meeting ACLR < −60.8 dBc.

Key Points

• Inverse-design approach: a deep convolutional neural network trained as an EM surrogate predicts S‑parameters of pixelated passive networks quickly and accurately, enabling fast evaluation inside an optimizer loop.
• Optimization stack: the CNN surrogate is embedded in a blackbox Doherty framework and paired with a genetic algorithm (GA) to search the discrete pixel design space for multi‑port combiners.
• Design novelty: synthesis of complex, three‑port pixelated output combiners that extend efficiency over back‑off using fully symmetrical device implementations.
• Prototypes: two Doherty PA prototypes built with GaN HEMT transistors and three‑port pixelated combiners were fabricated and tested at 2.75 GHz.
• Measured performance:
  • Peak drain efficiency > 74%
  • Output power > 44.1 dBm
  • Drain efficiency > 52% at 9 dB back‑off
  • With a 20‑MHz, 9 dB‑PAPR 5G‑NR-like signal and DPD: average PAE > 51% and ACLR ≤ −60.8 dBc
• Validation: physical fabrication and over‑the‑air measurements confirm the surrogate-driven design transfers to real hardware.

Data & Methods

• Surrogate modeling:
  • A deep convolutional neural network was developed to serve as an EM surrogate mapping pixelated network layouts to S‑parameters. The surrogate enables orders‑of‑magnitude faster evaluations than full-wave EM simulation (enabling global search), though quantitative speedups are not specified in the summary.
  • The surrogate allows handling of multi‑port, discrete pixel layout spaces where standard continuous synthesis is hard.
• Optimization:
  • A genetic algorithm (GA) optimizer runs in a blackbox Doherty framework, using the CNN surrogate to evaluate candidate combiners and select layouts that maximize efficiency and power metrics across operating points (including back‑off).
  • The blackbox Doherty model integrates device nonlinear behavior and network S‑parameters to evaluate system-level metrics.
• Prototyping and measurements:
  • Two prototypes were implemented using GaN HEMT transistors and three‑port pixelated combiners derived from the optimizer output.
  • RF characterization at 2.75 GHz measured drain efficiency, output power, back‑off efficiency, and linearity metrics.
  • Realistic waveform testing: a 20‑MHz 5G‑NR‑like waveform with 9 dB PAPR; digital predistortion (DPD) was applied to evaluate average PAE and ACLR.
• Notes / limitations:
  • Exact training dataset size, CNN architecture/hyperparameters, and surrogate error metrics are not detailed in the provided summary.
  • Comparative runtime or accuracy statistics versus standard EM-driven design loops are not reported here.

Implications for AI Economics

• Productivity and R&D scaling:
  • Surrogate-assisted inverse design reduces the marginal cost and time of exploring high‑dimensional, discrete hardware design spaces by replacing costly EM simulations with fast ML inference. This raises R&D productivity and can shorten design cycles for RF components.
  • Faster iteration lowers the effective cost of experimentation, increasing returns to scale for firms that invest in ML design infrastructure.
• Capital and labor effects:
  • The approach raises demand for ML‑fluent RF engineers and design-tooling specialists, shifting skill premia toward hybrid expertise (RF + ML).
  • Routine parts of RF layout search may become automated, enabling engineers to focus on higher‑level system integration and validation; some low‑skill design tasks could be displaced.
• Barriers to entry and competition:
  • Surrogate-driven design tools can lower technical barriers to entry for smaller firms and startups by reducing simulation infrastructure needs, potentially increasing competition in hardware and RF markets.
  • At the same time, firms that develop proprietary, high-quality surrogate models may gain competitive advantage and capture larger rents, reinforcing winner‑take‑most dynamics in specialized design-tool markets.
• Cost structure of network deployment and operations:
  • Improved PA efficiency across back‑off directly reduces base station power consumption and operational costs (OPEX), which scales across operators and networks. This can alter network economics, lowering total cost of ownership and accelerating rollouts, especially where energy costs are significant.
  • Better energy efficiency also has climate/ESG implications, lowering emissions per unit data transmitted.
• Innovation diffusion and generality:
  • The methodological template—train an ML surrogate of a costly simulator and embed it in an optimizer—generalizes beyond Doherty PAs to other analog/microwave components, antennas, and possibly broader engineering domains. Widespread adoption could accelerate hardware innovation across multiple industries.
• Policy and IP considerations:
  • The value of trained surrogates raises questions about data/IP (who owns training data, surrogate models, and optimized layouts). Proprietary surrogates could create gatekeepers; open‑source surrogates would democratize access.
  • Regulators and standard bodies may need to account for rapid tool-driven design when assessing compliance, safety, and reproducibility.
• Market and environmental externalities:
  • Higher‑efficiency PAs can reduce energy demand from telecom networks, affecting energy markets and possibly delaying investment in generation capacity; conversely, lower costs and improved performance could increase demand for data, partially offsetting efficiency gains (rebound effects).
• Overall assessment:
  • This work exemplifies how ML surrogates can materially change the economics of hardware R&D—lowering costs and accelerating iteration—while shifting value toward firms and workers that combine domain and ML capabilities. The net effect is likely faster, cheaper hardware innovation with important distributional and policy implications for industry structure, labor markets, and energy use.