PCBSchemaGen

LLM-based PCB schematic generation with automated topology verification.

Installation

PCBSchemaGen requires:

• Python >= 3.10
• KiCad 9 (with pcbnew Python bindings)
• OpenAI-compatible API access

Step 1: Install KiCad 9

Follow the official KiCad installation guide for your platform:

https://www.kicad.org/download/

After installation, verify KiCad and pcbnew:

kicad-cli --version
python3 -c "import pcbnew; print(pcbnew.GetBuildVersion())"

Step 2: Install Python Dependencies

pip install openai pandas skidl networkx cairosvg Pillow

Step 3: Environment Check

cd "sample design"
XDG_DATA_HOME=$(pwd)/../.xdg python3 run_samples_test.py

All 17 tasks should print [PASS]. If not, check your KiCad installation.

Quick Start

cd task
XDG_DATA_HOME=$(pwd)/.. python3 run.py \
  --task_id 1 \
  --model gpt-4o \
  --api_key "YOUR_API_KEY" \
  --base_url https://openrouter.ai/api/v1

This generates a circuit for Task 1 (voltage divider).

API Keys

Any OpenAI-compatible API works. We recommend OpenRouter:

1. Create account at https://openrouter.ai
2. Get API key from dashboard
3. Use with --base_url https://openrouter.ai/api/v1

Benchmark

• Task definitions: benchmark.tsv (23 tasks across 3 difficulty levels)
• Reference designs: sample design/p*.py
• Component library: component.json, kg_component.json
• KiCad symbols/footprints: library/

Task Difficulty Levels

Level	Tasks	Description
Easy	P1-P6	Voltage dividers, LDOs, sensing circuits
Medium	P7-P16	Half-bridge stages, gate drivers, isolated DC-DC
Hard	P17-P23	Multi-switch converters (sync buck, DAB, LLC, 3-phase)

Example Output

The example_output/ directory contains a successful run of Task 17 (Synchronous Buck Converter):

File	Description
attempt_1_skidl.py	LLM-generated SKiDL code
attempt_1.svg	Circuit schematic visualization
task_17_stats.json	Run statistics (tokens, time, status)
task_17_output.txt	Full LLM output with chain-of-thought
extracted_task_17.*	Final artifacts (netlist, KiCad project)

Project Structure

PCBSchemaGen/
├── task/
│   ├── run.py                  # Single task runner
│   ├── run_feedback_trials.py  # Batch experiments
│   ├── topo/                   # Verification pipeline
│   └── prompt_template*.md     # LLM prompts
├── sample design/              # Ground truth implementations
├── library/                    # KiCad symbols & footprints
├── example_output/             # Example Task 17 output
├── benchmark.tsv               # Task definitions
├── component.json              # Component pin definitions
└── kg_component.json           # Component constraints

Batch Experiments

To run multiple tasks with multiple trials:

cd task
XDG_DATA_HOME=$(pwd)/.. python3 run_feedback_trials.py \
  --task-range 1-16 \
  --trials 15 \
  --model gpt-4o \
  --api_key "YOUR_API_KEY" \
  --base_url https://openrouter.ai/api/v1 \
  --feedbacks full \
  --parallel-threads 8

Results are saved to task/feedback_runs/.