BirdAttractor BirdAttractor RevA

PCBench: A Dataset for Printed Circuit Board Routing

PCBench is a dataset for PCB routing task, it includes a dataset consisting of 164 printed circuit boards (PCB), a data augmentation script to expand the dataset for supervised learning (SL) tasks, and a reinforcement learning (RL) environment.

Dataset

Folder Structure

All the PCB designs in our dataset are stored in the folder PCBs, where each PCB design is stored in a subfolder named {author/org}_{PCB_name}_{design_version(optional)}. Below is the folder structure of the dataset.

├── PCBs
│   ├── master_metadata.json
│   ├── subfolder1
│   │   ├── raw.kicad_pcb
│   │   ├── processed.kicad_pcb
│   │   ├── final.json
│   │   ├── metadata.json
│   │   ├── visual.png
│   ├── subfolder2
│   │   ├── raw.kicad_pcb
│   │   ├── processed.kicad_pcb
│   │   ├── final.json
│   │   ├── metadata.json
│   │   ├── visual.png
│   ├── ...

The raw.kicad_pcb is the raw PCB design file, process.kicad_pcb is the file after cleaning up, final.json stores only routing-related information used for ML tasks. metadata.json stores the meta information for each single PCB design. visual.png is a visualization of each single PCB design. master_meta.json contains metadata of all PCBs and global-level information.

PCB Routing Description Language (PCB-RDL)

We propose a JSON specification (final.json for each PCB design) called the PCB Routing Description Language (PCB-RDL) that expresses a PCB routing problem and its solutions intuitively using basic concepts in order to facilitate research into automated PCB routing using ML. Following figure shows is an example of PCB with marked components in PCB-RDL (Left) and it corrsponding PCB-RDL (Right).

Data Augmentation

Run

The augmentation data can be generated with the script Scripts/Data_augmentation/augmentation.py. To generate 3 samples by randomly extracting 50% nets from the PCB 1Bitsy_1bitsy, please run the following command

cd Scripts/Data_augmentation/
python augmentation.py --num_samples 3 --net_ratio 0.5 --pcb_name 1Bitsy_1bitsy

Output

All the generated PCBs will be stored in the folder named augmented_data under Scripts/Data_augmentation/. In augmented_data, all the generated samples will be stored in a subfolder with the name of selected PCB. Each generated sample has the name sample_index.json. For example, the above command will generate the following file under Scripts/Data_augmentation/

├── augmented_data
│   ├── 1Bitsy_1bitsy
│   │   ├── 1.json
│   │   ├── 2.json
│   │   ├── 3.json

RL environment

Installation

The RL environment requires Python >= 3.7. You can simply install the environment with the following commands:

git clone https://github.com/PCBench/PCBench.git
cd PCBench
python setup.py install

After installation, open a Python console and type

from RLEnv.EnvLayer.PCBRoutingEnv import PCBRoutingEnv

If no error occurs, you have successfully installed the PCB routing environment.

API Usage

The following content shows an example of API usage of PCBRoutingEnv.

from RLEnv.EnvLayer.PCBRoutingEnv import PCBRoutingEnv
import random

resolution = [0.5, 0.5]
pcb_folder = '../PCBs/'
pcb_names = ["1Bitsy_1bitsy"]
iters = 30
env = PCBRoutingEnv(resolution=resolution, pcb_folder=pcb_folder, pcb_names=pcb_names)
obs, info = env.reset()
for _ in range(iters):
    act = random.randint(0,5)
    obs, rew, terminal, _, info = env.step(act)
    if terminal:
        env.reset()

You can find more examples of customizing functions of reward and state observation from RLEnv/examples.ipynb.