ABB Researchers Demonstrate LLM-Based Multi-Agent Framework for Autonomous Power Electronic Circuit Design
Key Takeaways
- ▸Multi-agent LLM framework successfully generates complete power electronic circuits from natural language specifications without fine-tuning
- ▸System achieves 100% logical connectivity and automated routing, solving a critical bottleneck in circuit design automation
- ▸Physics-aware layout optimization and manufacturing compliance (creepage, clearance, thermal management) remain open challenges requiring deeper EDA tool integration
Summary
Researchers at ABB Inc. have published a novel approach to automating power electronic circuit design using Large Language Models (LLMs) orchestrated through a multi-agent framework. The system translates natural language specifications directly into manufacturing-ready PCB designs without requiring fine-tuning of underlying LLMs, addressing a long-standing challenge in automating the traditionally manual process of power converter design.
The framework employs specialized agents for device specification, component selection, netlist generation using SKiDL, and layout completion, operating within a constrained environment that bridges the gap between generative AI's stochastic nature and the rigid requirements of Electronic Design Automation (EDA) tools. The team successfully demonstrated the approach by autonomously generating a 400V 3-phase converter for a variable frequency drive, achieving 100% logical connectivity and automated routing.
While the framework solves the connectivity problem, the researchers identified physics-aware optimization as the next frontier, noting that meeting real-world manufacturing requirements for creepage, clearance, commutation loop inductance, and thermal management will require further integration of physics-based feedback loops into the agent architecture.
- Approach demonstrates the viability of pairing generative AI with deterministic routing tools for complex hardware design domains
Editorial Opinion
This work represents a significant step forward in applying LLM-based agents to highly constrained hardware design problems. The key insight—that LLMs excel at capturing high-level design intent but require specialized agents and deterministic tools to enforce domain constraints—offers a promising template for other complex engineering domains. However, the gap between logical correctness and physics-aware manufacturing readiness underscores an important limitation: generative AI alone cannot replace the deep domain knowledge embedded in EDA tools, and successful automation requires careful orchestration between stochastic and deterministic systems.
