Circuit2Graph

This repository contains the source code for implementing graph classification in a graph neural network using PyTorch (CPU/GPU). The code transforms analog/digital circuits described in the papers into a graph network.

Publications

• Yusuke Yamakaji, Hayaru Shouno, and Kunihiko Fukushima, Circuit2Graph: Circuits With Graph Neural Networks, IEEE Access, vol. 12, pp. 51818-51827, 2024.
• Yusuke Yamakaji, Hayaru Shouno, and Kunihiko Fukushima, Circuit2Graph: Diodes as Asymmetric Directional Nodes, IEEE Access, vol. 12, pp. 168963-168974, 2024.
• Yusuke Yamakaji, Hayaru Shouno, and Kunihiko Fukushima, Equivalent Circuit for Single/Three Phase Magnetic Coupling With Graph Neural Networks, IEEE Transactions on Power Electronics, vol. 40, no. 2, pp. 3313-3325, Feb. 2025.

Environment Setup

Assume that LTspice XVII is installed in the "C:\LTspice" directory. If the installation path is different, please modify the "default=os.path.join('C:', 'LTspice')" in params.py. Additionally, LTspice XVII may have files like examples.zip and lib.zip in the "C:\LTspice" directory. Please unzip these files beforehand so that "C:\LTspice\examples" and "C:\LTspice\lib" are accessible.

Ensure that LTspice XVII is installed in the C:\LTspice directory. If the installation path is different, modify the default=os.path.join('C:', 'LTspice') in params.py. Additionally, unzip examples.zip and lib.zip in the C:\LTspice directory so that C:\LTspice\examples and C:\LTspice\lib are accessible.

For GPU environment

1. Install the NVIDIA driver and CUDA Toolkit according to your environment. This guide uses CUDA Toolkit version 11.8.

2. Install the required libraries:

   pip install torch --index-url https://download.pytorch.org/whl/cu118
   pip install numpy torch_geometric
   # Optional
   pip install matplotlib networkx 

This code has been tested with the following library versions:

• numpy: 2.2.1
• torch: 2.5.1
• torch_geometric: 2.6.1
• matplotlib: 3.10.0 (optional)
• networkx: 3.4.2 (optional)

For CPU environment

The calculation time is slower compared to a GPU environment, but the code supports training and inference on a CPU.

1. Install the required libraries:

   pip install numpy torch torch_geometric
   # Optional
   pip install matplotlib networkx 

Execution

When running this code for the first time, execute the batch file _make_netlist.bat created by the make_net_batch function in extract_netlist.py to extract netlists from sample circuits of LTspice.

To execute, run the following scripts in order:

python extract_netlist.py
python extract_node_edge.py
python data_loading.py
python train.py

Hyperparameters

The hyperparameters can be adjusted in params.py:

• --semiconductor: Defines how semiconductors are treated. Options are single, star, or complete.

  1. "single": Semiconductors are treated as single nodes.
  2. "star": Semiconductors are treated as a star graph.
  3. "complete": Semiconductors are treated as a complete graph.

• --mutual: Defines how the coupling coefficient of magnetic coupling is handled. Options are equiv or none.

  1. "none": The coupling coefficient is not held.
  2. "equiv": The coupling coefficient is held by transforming it into an equivalent circuit.