Circuit-GNN: Graph Neural Networks for Distributed Circuit Design

Authors: Guo Zhang, Hao He, Dina Katabi

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compare our model with a commercial simulator showing that it reduces simulation time by four orders of magnitude. We also demonstrate the value of our model by using it to design a Terahertz channelizer, a difficult task that requires a specialized expert. The results show that our model produces a channelizer whose performance is as good as a manually optimized design, and can save the expert several weeks of topology and parameter optimization.
Researcher Affiliation Academia 1EECS, Massachusetts Institute of Technology, Cambridge, MA, USA.
Pseudocode No The paper does not contain any pseudocode or algorithm blocks. Methods are described in prose.
Open Source Code No For more dataset details, please see our project website: https://circuit-gnn.csail.mit.edu. This URL points to a project website for dataset details, not an explicit code repository for the methodology.
Open Datasets No To train our network, we generate labeled examples using the CST STUDIO SUIT (CST official website, 2018), a commercial EM simulator. We generate about 100,000 circuit samples made of 3 to 6 resonators on a distributed computing cluster with 800 virtual CPU cores. For more dataset details, please see our project website: https://circuit-gnn.csail.mit.edu. The dataset was generated by the authors, and the project website is stated for 'dataset details' rather than providing direct public access to download the dataset itself.
Dataset Splits Yes We train on 80% of the data with 4 and 5 resonators, and test on the rest, including the data with 3 and 6 resonators which are 100% reserved for testing.
Hardware Specification Yes In terms of the run-time for prediction, our model conducts one prediction in 50 milliseconds on a single NVIDIA 1080Ti GPU which is four orders of magnitude faster than running one simulation using CST on a modern desktop.
Software Dependencies No The paper mentions 'CST STUDIO SUIT' and 'Adam optimizer' but does not provide specific version numbers for these or any other software libraries or programming languages used (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes Training uses the Adam optimizer (Kingma & Ba, 2014) and a batch-size of 64. In total, the model is trained 500 epochs. The learning rate is initialized as 10 4 and decayed every 200 epochs by a factor of 0.5.