Dall-EM: AI-Enabled Generative Design System for Rapid Synthesis of Electromagnetic Passive Components with Controllable Complexity

Background

Millimeter-wave (mmWave) and terahertz (THz) integrated circuits (ICs) and printed circuit boards (PCBs) are critical for next-generation wireless communication, sensing, and radar systems. These frequency ranges enable high data rates, miniaturization, and new applications, but also introduce significant design challenges due to the complex electromagnetic (EM) behavior of passive components at such small wavelengths. Traditionally, the design of these EM structures relies on expert-driven workflows, where human designers iteratively select and optimize from a limited set of classical circuit templates—such as transmission lines and resonators—using time-consuming electromagnetic simulations and manual tuning. The growing demand for rapid innovation, increased performance, and reduced costs in mmWave and THz technologies has created a pressing need for more efficient, automated, and versatile design methodologies that can explore a broader range of EM topologies.

Current approaches to EM component synthesis are hindered by several limitations. The reliance on human intuition and experience restricts the exploration of unconventional or highly complex structures, confining the design process to a narrow subset of possible solutions. Even recent AI-based EM emulators, which predict EM properties from structure images, fall short because they cannot generate new designs with controllable complexity or emulate the interpretable forms favored by human designers. This lack of flexibility and control results in suboptimal prediction accuracy and limits practical utility. Furthermore, the traditional bottom-up design process is both slow and expensive, often requiring months or years and substantial financial investment to achieve acceptable performance. As a result, there is a clear need for new methodologies that can efficiently and automatically synthesize EM structures tailored to specific performance requirements, while offering control over structural complexity and interpretability.

Technology

This technology is an AI-driven methodology for the rapid synthesis and design of electromagnetic (EM) passive components and circuits, specifically targeting millimeter-wave (mmWave) and terahertz applications. It utilizes a conditional diffusion model—a type of generative AI—trained on simulated classical EM structures to create EM layouts from random noise, guided by user-specified scattering parameters (S-parameters). The system encodes EM structures as binary matrices and allows precise control over the structural complexity of the generated designs, ranging from simple, interpretable classical forms to highly complex, pixelated topologies. This complexity is quantified using a spatial entropy metric. The training process employs a two-stage approach with transfer learning, leveraging both low- and high-fidelity simulation data to efficiently develop accurate surrogate models and the generative diffusion model. The resulting EM structures can be further optimized and verified using commercial EM solvers, and the entire workflow—from inputting target performance to generating and validating a physical layout—can be completed in minutes.

What differentiates this technology is its ability to perform true inverse design: users specify the desired EM performance, and the system automatically generates a physical structure that achieves it, overcoming the limitations of traditional, expert-driven workflows. Unlike previous AI-based EM emulators, this approach offers controllable structural complexity, enabling the generation of both human-interpretable and novel, unconventional topologies. The use of generative diffusion models conditioned on S-parameters allows for the creation of multi-port EM structures with flexible port placement, supporting a wide range of applications from filters to antennas. The integration of transfer learning and data augmentation dramatically reduces the need for costly EM simulations, while the surrogate model accelerates optimization and design verification. This methodology not only expands the design space beyond classical templates but also reduces design time and cost, making it a transformative solution for the development of next-generation wireless communication and sensing hardware.

Benefits

• Significantly reduces design time and cost for millimeter-wave and terahertz EM components from months to minutes.
• Enables inverse design of EM structures directly from target scattering parameters, automating topology selection and layout synthesis.
• Provides controllable structural complexity, allowing generation of both classical interpretable designs and novel, complex topologies.
• Improves design accuracy and efficiency using AI surrogate models and generative diffusion models trained with transfer learning.
• Supports multi-port EM structures and integration with active circuit elements for comprehensive RF front-end design.
• Facilitates easy verification and optimization through compatibility with commercial EM solvers and discrete optimization algorithms.
• Expands the accessible design space beyond traditional human-driven methods, enabling discovery of innovative EM topologies.
• Integrates with CAD/EDA tools, offering interpretable layouts that can be modified by designers and fabricated using standard IC or PCB processes.

Applications

• Rapid mmWave PCB component design
• Automated RF filter synthesis
• EDA tool integration for EM layout
• Custom multi-port passive device generation
• Inverse design for terahertz circuits

Additional info

This AI-driven system synthesizes electromagnetic passive components and circuits for millimeter-wave and terahertz applications. It leverages a conditional diffusion model, trained on simulated EM structures, to generate physical layouts (binary matrices) from random noise. The generation is guided by target scattering parameters and allows control over structural complexity, ranging from interpretable classical forms to novel, pixelated topologies.