Evaluating LLM-based Workflows for Switched-Mode Power Supply Design

TL;DR

This work investigates the use of large language models (LLMs) to assist in switched-mode power supply (SMPS) circuit design for PCBs. It proposes a family of LLM-based workflows that combine reasoning, retrieval-augmented generation (RAG), and a custom SPICE-interaction toolkit to simulate circuit modifications and extract feedback. Across two benchmark experiments with 269 tasks, SPICE simulation feedback and LLM reasoning dramatically improve problem-solving capability, achieving up to a 91% solve rate, while topology adaptation remains challenging for complex nets due to text-based netlist limitations. The results suggest that while LLMs can substantially aid SMPS design, especially for parameter tuning, future work should focus on richer circuit representations and hybrid optimization to better explore design spaces and improve topology adaptation.

Abstract

Large language models (LLMs) have great potential to enhance productivity in many disciplines, such as software engineering. However, it is unclear to what extent they can assist in the design process of electronic circuits. This paper focuses on the application of LLMs to switched-mode power supply (SMPS) design for printed circuit boards (PCBs). We present multiple LLM-based workflows that combine reasoning, retrieval-augmented generation (RAG), and a custom toolkit that enables the LLM to interact with SPICE simulations to estimate the impact of circuit modifications. Two benchmark experiments are presented to analyze the performance of LLM-based assistants for different design tasks, including parameter tuning, topology adaption and optimization of SMPS circuits. Experiment results show that SPICE simulation feedback and current LLM advancements, such as reasoning, significantly increase the solve rate on 269 manually created benchmark tasks from 15% to 91%. Furthermore, our analysis reveals that most parameter tuning design tasks can be solved, while limits remain for certain topology adaption tasks. Our experiments offer insights for improving current concepts, for example by adapting text-based circuit representations

Figures (14)

• Figure 1: Required reasoning chain to solve the following exemplary SMPS design task: "Adjust the given netlist, such that the current ripple has the value 100 mA". The netlist and complete input prompt can be seen in Figure \ref{['fig:colored_prompt']}. The controller's datasheet and SPICE simulations are used as external information sources. $\Delta I_{L}$: ripple current, $f_s$: switching frequency, $V_{\textrm{out}}$: output voltage, $V_{\textrm{in}}$: input voltage, $R_\textrm{top}$ and $R_\textrm{bot}$: resistors in feedback path
• Figure 2: Example SPICE simulation signals from the case study
• Figure 3: Overview of the LLM-based workflows: The baseline workflow uses an instruction-tuned LLM (GPT-4o or o3) that receives a reference circuit and requirements as input to directly generate an adapted circuit netlist. The first extension enhances this workflow with information retrieval from datasheets using RAG. The second extension equips the LLM-based workflow with a comprehensive set of feature extraction tools to receive simulation feedback
• Figure 4: The three SMPS circuit types that serve as the basis for the benchmark, shown in order of increasing difficulty
• Figure 5: Illustration of a topology adaption task and the different parameter tuning tasks - $V_{\mathrm{out}}$, $\Delta I_{L}$, $\Delta V_{\mathrm{out}}$, $f_{\mathrm{sw}}$, $t_{\mathrm{start}}$. For the different task categories affected elements in the circuit excerpt (a) and in the simulation signal segments (b) are marked in the same color.