A deep learning-enabled universal inverse design framework for rapid synthesis of complex multi-port electromagnetic structures and circuits

Background

Millimeter-wave and terahertz integrated circuits are foundational to the advancement of next-generation wireless communication, sensing, and radar systems. These high-frequency technologies demand sophisticated electromagnetic (EM) structures—such as antennas, filters, and power dividers—that can operate efficiently at extremely small scales and high bandwidths. The design of such components is inherently complex, as it requires precise control over scattering parameters, impedance matching, and radiation patterns across broad frequency ranges. As wireless standards evolve and device miniaturization accelerates, there is a growing need for design methodologies that can rapidly produce high-performance, application-specific EM structures with unprecedented flexibility in geometry and function.

Current approaches to EM structure design rely heavily on human expertise, parameterized templates, and iterative electromagnetic simulations. These traditional methods are hampered by several critical limitations: the design space is restricted to a small set of pre-defined topologies, which constrains achievable performance and innovation. The optimization process is slow and computationally expensive, often requiring weeks or months of repeated full-wave EM simulations for each design iteration. Furthermore, existing machine learning surrogates, when used, are typically tailored to specific device types or fixed layouts, lacking the generalization needed for arbitrary multi-port or broadband designs. As a result, many designs fall short of theoretical performance limits, and the co-design of passive EM structures with active circuits remains a labor-intensive and suboptimal process.

Technology

A deep learning-enabled universal inverse design methodology for radio-frequency (RF) and millimeter-wave (mmWave) systems leverages convolutional neural networks (CNNs) to synthesize complex, arbitrary-shaped multi-port electromagnetic (EM) structures and circuits. This approach represents planar EM structures as pixelated binary images, enabling the CNN surrogate to rapidly and accurately predict multi-port scattering parameters and radiation patterns across broad frequency ranges. The CNN is trained using a two-step transfer learning process—first on large, low-fidelity simulation datasets, then fine-tuned with high-fidelity data—allowing it to generalize to a wide variety of structures. An optimization loop, powered by heuristic algorithms and customizable cost functions, iteratively proposes new candidate structures, which are quickly evaluated by the CNN to converge on designs with desired EM properties such as impedance matching, power division, and phase control. This framework supports the synthesis of a wide array of devices, from antennas and filters to power dividers and combiners, and enables seamless co-design with active circuits, drastically reducing design cycles from months to minutes.

What differentiates this technology is its ability to transcend the limitations of traditional EM design methodologies, which are constrained by pre-defined templates, slow simulation times, and limited generalization. The pixel-based representation and deep learning surrogate unlock an exponentially larger design space, allowing for the discovery of novel, non-intuitive structures that outperform conventional designs in bandwidth, efficiency, and spectral purity. The CNN surrogate’s speed enables rapid optimization and supports arbitrary multi-port configurations, while transfer learning ensures adaptability to different fabrication processes, frequency bands, and physical constraints. This universal framework is validated through both simulation and experimental fabrication, demonstrating superior performance in both passive and active RF/mmWave components, and paving the way for automated, high-performance integrated circuit design that was previously unattainable with manual or template-driven approaches.

Benefits

• Drastically reduces design time from months to minutes by replacing costly EM simulations with fast deep learning surrogate models.
• Enables universal inverse design of arbitrary-shaped multi-port electromagnetic structures, expanding the design space beyond traditional template-based methods.
• Supports co-design of passive EM components with active circuits, improving overall system performance such as broadband gain and impedance matching.
• Achieves high prediction accuracy of complex scattering parameters and radiation patterns across broad frequency ranges using convolutional neural networks.
• Facilitates discovery of novel, non-intuitive EM structures with superior performance metrics like bandwidth, power division, and phase control.
• Generalizes across different fabrication processes, frequency bands, and device sizes through transfer learning and data augmentation.
• Automates optimization with customizable cost functions and heuristic algorithms, enabling flexible targeting of amplitude, phase, and spectral characteristics.
• Validated experimentally with fabricated devices, demonstrating practical applicability in RF/mmWave integrated circuits and antenna systems.

Applications

• Automated RF/mmWave IC design
• Custom multi-port antenna synthesis
• Rapid filter and diplexer development
• Broadband amplifier co-design
• Automated power divider/combiner creation

Additional info

This deep learning-enabled universal inverse design methodology synthesizes complex multi-port electromagnetic structures for RF/mmWave systems. It leverages trained convolutional neural networks as fast EM surrogates, predicting scattering parameters from pixelated binary images. An optimization loop iteratively refines candidate structures based on customizable cost functions, enabling rapid creation of arbitrary-shaped devices and co-design with active circuits.