PowerCircuitAI: Agentic Generation of Schematics - The Case of a Three-Level NPC Inverter

Abstract

Power electronic schematic capture remains difficult to automate because a valid converter design must coordinate topology, device ratings, support circuitry, net connectivity. Large language models (LLMs) provide a useful interface for high-level design intent, but unconstrained generation can produce plausible text and code with electrically invalid nets or inconsistent pin assignments. This paper presents a constrained multi-agent workflow for translating natural-language power converter requirements into rule-checked schematic and executable simulations using an agentic pipeline supported by a library of standard topology templates. The proposed methodology selects a canonical circuit graph from the template library, validates graph integrity, and deterministically extracts structural layouts (input connections and phase legs) to generate KiCad schematics and bill-of-materials (BOM). Subsequently, an agentic SPICE generator builds a simulation deck that incorporates self-healing capabilities through automated solver relaxations and agentic repair loops. The method is evaluated on a single-sentence request for a 1.5 kW, 60 Hz, 540 V dc-bus, three-phase three-level neutral-point-clamped (NPC) inverter. The case study demonstrates that the pipeline successfully resolves the required components, derives the complete detailed schematic, executes SPICE transient simulations outputting the expected three-level pole voltages and multi-level line-to-line staircases, and produces a consistent artifact package. The results show that LLM agents can support the automation of power electronics schematics and simulations for a representative multilevel topology when subject to strict deterministic engineering checks.