AI-Driven Design Automation Platform for Optimizing Active and Passive Components in Radio-Frequency Integrated Circuits

Background

The design and synthesis of radio-frequency integrated circuits (RFICs), especially those operating at millimeter-wave (mmWave) and higher frequencies, are foundational to next-generation wireless systems, including applications in smart cities, autonomous vehicles, and advanced communications infrastructure. As the demand for higher data rates and more sophisticated wireless capabilities grows, there is an increasing need for RFICs that can deliver high performance across multiple metrics such as gain, bandwidth, efficiency, and power output. However, the complexity of these circuits, which involve intricate interactions between active devices (like transistors and amplifiers) and passive electromagnetic (EM) structures (such as matching networks and transmission lines), makes their design highly challenging. The vast, high-dimensional design space and the need to account for strong coupling and parasitic effects further complicate the process, necessitating innovative solutions to streamline and optimize RFIC development.

Traditional approaches to RFIC design are predominantly manual, relying on the expertise and intuition of experienced engineers who iteratively refine circuit topologies and parameters using fixed templates and labor-intensive optimization cycles. This process is not only time-consuming and costly but also inherently limited by human experience and the practical constraints of exploring only a small fraction of the possible design space. Automated methods developed to date, such as those based on neural networks trained on parameter sweeps or metaheuristic algorithms, have significant shortcomings: they require large, expensive labeled datasets, struggle with topology optimization, exhibit poor convergence in high-dimensional spaces, and often fail to accurately incorporate complex EM effects and parasitics. As a result, current methodologies are unable to fully address the demands of modern RFIC design, highlighting the urgent need for more robust, scalable, and intelligent automation tools.

Technology

This technology is an AI-driven framework for automated synthesis of radio-frequency integrated circuits (RFICs), with a particular focus on millimeter-wave power amplifiers. It integrates a reinforcement learning (RL) engine with an inverse design synthesizer to concurrently optimize both active and passive circuit components. The RL engine autonomously explores and tunes active circuit parameters—such as transistor sizes, bias voltages, and stage configurations—using a multi-objective reward function and real-time circuit simulation feedback. In parallel, the inverse design synthesizer employs genetic algorithms and a convolutional neural network-based electromagnetic (EM) emulator to rapidly generate and evaluate pixelated EM structures for passive components like matching networks. This iterative and integrated approach enables automated architecture selection, topology optimization, and parameter tuning, translating high-level code-based specifications directly into optimized physical layouts. The system has demonstrated its effectiveness through the successful design and fabrication of a 120 GHz power amplifier, validating its end-to-end capabilities from specification to working silicon.

What differentiates this technology is its holistic and experimentally validated application of AI to RFIC design, overcoming the limitations of traditional manual methods and previous automation attempts. Unlike conventional approaches that rely on fixed architectures, human intuition, or require large labeled datasets, this framework leverages RL for sequential exploration and topology optimization without pre-defined templates. Its integration with inverse design allows for the creation of passive EM structures with exponentially greater design freedom than traditional methods, while the use of a fast EM emulator accelerates the optimization process. The system supports multi-objective optimization, rapid synthesis post-training, and transfer learning for new objectives, drastically reducing design time and lowering the expertise barrier. By uniting active and passive component optimization and effectively handling complex EM coupling and parasitics, this technology enables the exploration of vast design spaces and delivers high-performance RFICs with unprecedented speed and flexibility, positioning it as a significant advancement in electronic design automation.

Benefits

• Automates complex RFIC design, significantly reducing design time from weeks or months to minutes after training.
• Enables simultaneous optimization of active circuit topologies and passive electromagnetic structures, improving overall circuit performance.
• Eliminates the need for large labeled datasets by leveraging reinforcement learning for sequential exploration and optimization.
• Supports multi-objective optimization across key performance metrics such as gain, bandwidth, efficiency, and saturation power.
• Incorporates parasitic effects and electromagnetic coupling for more accurate and realistic circuit designs.
• Facilitates transfer learning to quickly adapt to new design objectives, reducing additional training time and cost.
• Validated experimentally with a 120 GHz power amplifier, demonstrating practical viability and accuracy of the approach.
• Offers commercialization potential through direct use in RF/mmWave PA synthesis and integration into existing CAD software platforms.

Applications

• Automated mmWave power amplifier design
• RFIC design acceleration for wireless
• EDA tool integration for RF synthesis
• Custom RFIC solutions for manufacturers

Additional info

These technological advancements describe AI-driven methodologies for automated integrated circuit synthesis. They concurrently optimize active components using reinforcement learning and passive electromagnetic structures via inverse design with genetic algorithms and neural network emulation. This iterative integration enables rapid, multi-objective optimization from specifications to physical layouts.