Microwave Integrated Circuits Design with Relational Induction Neural Network

Jie Liu, Zhi-Xi Chen, Wenhui Dong, Xiao Wang, Jia Shi, Hong-Liang Teng, Xi‐Wang Dai, Stephen S.‐t. Yau, Chang-Hong Liang, Pingfa FengarXiv 2019 · DOI · arXiv

Abstract

The automation design of microwave integrated circuits (MWIC) has long been viewed as a fundamental challenge for artificial intelligence owing to its larger solution space and structural complexity than Go. Here, we developed a novel artificial agent, termed Relational Induction Neural Network, that can lead to an automotive design of MWIC and avoid brute-force computing to examine every possible solution, which is a significant breakthrough in the field of electronics. Through the experiments on microwave transmission line circuit, filter circuit and antenna circuit design tasks, strongly competitive results are obtained respectively. Compared with the traditional reinforcement learning method, the learning curve shows that the proposed architecture is able to quickly converge to the pre-designed MWIC model and the convergence rate is up to four orders of magnitude. This is the first study which has been shown that an agent through training or learning to automatically induct the relationship between MWIC's structures without incorporating any of the additional prior knowledge. Notably, the relationship can be explained in terms of the MWIC theory and electromagnetic field distribution. Our work bridges the divide between artificial intelligence and MWIC and can extend to mechanical wave, mechanics and other related fields.

Introduction

Microwave Integrated Circuit (MWIC) design is often considered a "black art" in electronics. Engineers must balance intricate structural details with precise electromagnetic performance, a process that typically requires years of experience and extensive trial-and-error using Computer-Aided Design (CAD) tools. From a computational standpoint, the design space is immense. Representing a circuit as a mesh of 500 vertices, where each can move in multiple directions, creates a solution space of approximately $4^{500}$—far exceeding the complexity of games like Go. Traditional machine learning approaches often struggle because they require manual parameter abstraction, which limits the AI's ability to explore non-intuitive but effective structural configurations.

Methodology: The Relational Induction Neural Network

To overcome the dimensionality curse, we introduce the Relational Induction Neural Network (RINN). The core innovation lies in its ability to "induct" the relationship between physical structure and electromagnetic function without human intervention.

Overview of the RINN framework, including clustering and the Deep RL model.

The RINN architecture consists of two primary stages:

  1. Unsupervised Clustering (Induction): Instead of exploring every possible vertex movement, the system uses a $k$-means clustering algorithm to group mesh changes that produce similar effects on the circuit's S-parameters. This step compresses the vast action space into a manageable set of functionally significant "typical actions."
  2. Deep Reinforcement Learning (Design): An Asynchronous Advantage Actor-Critic (A3C) agent utilizes these clusters to navigate the design space. The agent observes the current mesh state and S-parameters, selecting actions that maximize a reward function based on target specifications like center frequency, passband range, and return loss.

Results and Emergent Engineering

The RINN agent was tested on various filter and antenna design tasks, demonstrating remarkable efficiency. A key highlight is the speed of convergence: RINN reached optimal designs in roughly 10,000 steps, whereas traditional Deep Reinforcement Learning (DRL) models showed no convergence even after 1,000,000 steps—a speedup of four orders of magnitude.

Optimization steps for various tasks and reward curves comparing RINN with traditional DRL.

Beyond performance metrics, the agent exhibited "emergent engineering" behavior. Without being taught explicit filter theory, it discovered that adjusting resonator lengths controls frequency and modifying gaps between them affects coupling coefficients. Interestingly, RINN-designed circuits often feature irregular, "organic" shapes that differ significantly from the geometric simplicity favored by human engineers, yet they meet or exceed all technical benchmarks.

Conclusion

The Relational Induction Neural Network bridges the gap between raw structural manipulation and high-level functional design in MWICs. By automating the discovery of electromagnetic relationships, RINN enables the rapid design of complex components that challenge human intuition. This approach holds promise for other fields governed by wave physics, including acoustics and structural mechanics.

Summary updated

Abstract

The automation design of microwave integrated circuits (MWIC) has long been viewed as a fundamental challenge for artificial intelligence owing to its larger solution space and structural complexity than Go. Here, we developed a novel artificial agent, termed Relational Induction Neural Network, that can lead to an automotive design of MWIC and avoid brute-force computing to examine every possible solution, which is a significant breakthrough in the field of electronics. Through the experiments on microwave transmission line circuit, filter circuit and antenna circuit design tasks, strongly competitive results are obtained respectively. Compared with the traditional reinforcement learning method, the learning curve shows that the proposed architecture is able to quickly converge to the pre-designed MWIC model and the convergence rate is up to four orders of magnitude. This is the first study which has been shown that an agent through training or learning to automatically induct the relationship between MWIC's structures without incorporating any of the additional prior knowledge. Notably, the relationship can be explained in terms of the MWIC theory and electromagnetic field distribution. Our work bridges the divide between artificial intelligence and MWIC and can extend to mechanical wave, mechanics and other related fields.
BibTeX
@misc{liu2019microwave,
  author = {Jie Liu and Zhi-Xi Chen and Wenhui Dong and Xiao Wang and Jia Shi and Hong-Liang Teng and Xi‐Wang Dai and Stephen S.‐T. Yau and Chang-Hong Liang and Pingfa Feng},
  title = {{Microwave Integrated Circuits Design with Relational Induction Neural Network}},
  year = {2019},
  eprint = {1901.02069},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1901.02069},
}

Loading related content...

Figure