TLDR: SPICEAssistant is a new framework that enhances large language models (LLMs) like GPT-4o for designing switched-mode power supplies (SMPS). It addresses LLM limitations in interpreting simulation results, lacking domain knowledge, and handling multi-step design processes by providing tools for flexible interaction with SPICE simulations and datasheet information retrieval. Benchmarking shows SPICEAssistant significantly outperforms standalone LLMs, improving design capabilities by approximately 38% through iterative simulation feedback, particularly for parameter tuning tasks.
Large language models, or LLMs, have shown incredible capabilities across many fields, including software development. However, their application in electronic design automation (EDA), particularly for complex tasks like designing switched-mode power supplies (SMPS) on printed circuit boards (PCBs), has presented unique challenges. A new framework called SPICEAssistant aims to bridge this gap by empowering LLMs with specialized tools to interact with simulation environments and access critical design information.
The Hurdles for LLMs in Circuit Design
Traditional LLMs face several significant obstacles when it comes to electronic circuit design. Firstly, they struggle to accurately interpret results from key simulation tools like SPICE. For instance, a case study revealed that GPT-4o, a state-of-the-art LLM, could not reliably extract information like ripple peak-to-peak values from time-series data, whether presented as raw numbers or images. Secondly, LLMs often lack specific domain knowledge for SMPS circuits, which frequently involve complex integrated circuits (ICs) and their peripheral components. Publicly available training data for these specific power controllers is scarce, and datasheets, which are crucial information sources, are often multimodal (containing text, images, and graphs) and difficult for LLMs to interpret. Thirdly, the SMPS design process is multi-step and complex, requiring sophisticated reasoning that LLMs find hard to mimic. Lastly, LLMs can sometimes produce unreliable or ‘hallucinated’ answers, a problem exacerbated when they lack domain-specific knowledge or deal with intricate reasoning tasks.
Introducing SPICEAssistant: An LLM-Powered Design Aid
To overcome these challenges, researchers have developed SPICEAssistant. This framework integrates an LLM, specifically GPT-4o, with a comprehensive set of tools. These tools act as an interface to the LTSpice simulator, allowing the LLM to interact flexibly with simulations. This interaction enables the LLM to estimate the impact of its circuit modifications and receive reliable feedback on simulated signals, such as mean output voltage, ripple, or switching frequency. Additionally, SPICEAssistant enhances the LLM’s circuit-specific knowledge by enabling it to interact with datasheets through Retrieval Augmented Generation (RAG).
This combination allows the LLM to blend its general knowledge and reasoning abilities with precise insights into the physical functionality of circuits. By going through multiple iterations with the simulator, the LLM can learn the effects of circuit changes, significantly improving its ability to mimic the complex SMPS design process. The simulation feedback also serves as a crucial verification step, boosting the reliability of the LLM’s responses.
Benchmarking SPICEAssistant’s Performance
To evaluate SPICEAssistant, a benchmark was created consisting of 256 questions. These questions test the LLM’s ability to adapt circuit netlists to fulfill various SMPS design tasks. The benchmark includes three types of circuits, ranging in difficulty from a simple general buck converter to more complex designs using the LTC3419 and LTC7802 controllers. Questions fall into two categories: parameter tuning (e.g., adjusting supply voltage or output voltage) and topology adaptation (e.g., selecting a specific operating mode).
The results are compelling. SPICEAssistant significantly outperforms a standalone GPT-4o across the entire benchmark, achieving an approximately 38% higher solve rate. The median absolute percentage error (APE) also dramatically improved from 64% to 4%. This indicates that simulation feedback is the primary driver of performance gain. While RAG for datasheet information showed a slight improvement for more complex circuits (LTC3419 and LTC7802), its impact was less pronounced for the general buck converter, likely due to GPT-4o’s existing knowledge of this common topology.
The study also observed that performance generally declined with increasing circuit complexity, which is expected given the more intricate netlists and functions involved. SPICEAssistant performed better on parameter tuning tasks compared to topology adaptation tasks, especially for the more complex circuits. This suggests that the semantic structure of SPICE netlists, which encode topology in a tabular format rather than a language-like description, might be a limiting factor for LLMs in topology adaptation.
Furthermore, the research showed that the performance of SPICEAssistant improved with an increasing number of interactions between the LLM and its simulation tools, plateauing after about five iterations. This highlights the importance of iterative feedback in complex design processes.
Also Read:
- AI Agents Reshaping Conceptual Engineering Design with Structured Language Models
- CodeAssistBench: A New Standard for Evaluating Programming AI
Future Outlook
SPICEAssistant represents a promising advancement in applying LLMs to electronic design. By providing LLMs with the ability to interact with SPICE simulations and retrieve datasheet information, it significantly enhances their capacity to understand, adapt, and dimension electronic circuits. While it excels in parameter tuning, future research could explore alternative text-based circuit representations to improve performance on topology adaptation tasks. For more details, you can refer to the full research paper available at arXiv.org.
