🚀 AICircuit: A Multi-Level Dataset and Benchmark for AI-Driven Analog Integrated Circuit Design

AICircuit is a comprehensive multi-level dataset and benchmark for developing and evaluating ML algorithms in analog and radio-frequency circuit design. AICircuit comprises seven commonly used basic circuits and two complex wireless transceiver systems composed of multiple circuit blocks, encompassing a wide array of design scenarios encountered in real-world applications. We extensively evaluate various ML algorithms on the dataset, revealing the potential of ML algorithms in learning the mapping from the design specifications to the desired circuit parameters.

📣 Updates: We've added a detailed guide on running simulations and using the Docker container! Check it out here: SIM_GUIDE.md.

📣 Check out our new paper Supervised Learning for Analog and RF Circuit Design: Benchmarks and Comparative Insights.

🔧 Requirements

To install the requirements you can use:

pip install -r requirements.txt

💡 Usage

Run Model

To run the code, first change the path for your input data and output results in arguments of config files in Config/Circuits based on your needs. <br>

Second, generate your own train config file, or use our default one. For using the default config file:

python3 main.py

As an example, if you want to address your own config path and a special seed:

python3 main.py --path ./Config/train_config.yaml --seed 0

Run Simulation

To run the simulation, check SIM_GUIDE.md for detailed instructions and setup. Then, use the following code:

python simulation.py --circuit=CSVA --model=MultiLayerPerceptron --npoints=100

Description

Here is a description for each parameter in the train_config.yaml file:

model_config: a list of different model configs:
- model:
  - KNeighbors
  - RandomForest
  - SupportVector
  - MultiLayerPerceptron
  - Transformer
- other model args: add desired model arguments to the model_config (no ordering). if no model args added, uses the default ones. For now, you can add:
  - KNeighbors parameters: n_neighbors, weights
  - RandomForest parameters: n_estimators, criterion
  - Transformer parameters: dim_model, num_heads, num_encoder_layers, dim_hidden, dropout_p.
subset: a list of the fraction of data used for training.
circuits: a list of circuits from this list:
- CSVA: Common-Source Voltage Amplifier
- CVA: Cascode Voltage Amplifier
- TSVA: Two-Stage Voltage Amplifier
- LNA: Low-Noise Amplifier
- Mixer
- VCO: Voltage-Controlled Oscillator
- PA: Power Amplifier
- Transmitter
- Receiver
epochs: used for MLP and Transformer model, default is 100.
loss_per_epoch: for KNN, RF, or SVR is False, else can be True or False by config value. default is True for MLP and Transformer.
compare_method: if True, for each subset and circuit, plots different models' loss (MLP and Transformer) on the same figure.
kfold: if True, K fold cross validation is performed. K is defined based on subset value. if False, no K fold cross validation.
independent_kfold: if True, train and test data is selected independently from previous folds, if false, dataset is divided to k parts at first and each time one part is selected for test and the remaining for train.
save_format:
- csv
- numpy
- anything else: no saving

Results

In each run, find plots in the corresponding folder in out_plot folder and loss values in the corresponding folder in out_result folder. Also if you set the train_config.yaml in a way that it saves the predictions, you can find the prediction files in the mentioned directory at the circuit's yaml file.

📈 Visualization

Some of our visual results are as follows. For Two-Stage Voltage Amplifier(TSVA), we have:

For Transmitter:

📚 Paper

"AICircuit: A Multi-Level Dataset and Benchmark for AI-Driven Analog Integrated Circuit Design", by A. Mehradfar, X. Zhao, Y. Niu, S. Babakniya, M. Alesheikh, H. Aghasi, and S. Avestimehr, arxiv: 2407.18272, 2024.

"Supervised Learning for Analog and RF Circuit Design: Benchmarks and Comparative Insights", by A. Mehradfar, X. Zhao, Y. Niu, S. Babakniya, M. Alesheikh, H. Aghasi, and S. Avestimehr, arxiv: 2501.11839, 2025.

Paper

🎯 Citation

If you use AICircuit in a research paper, please cite our works as follows:

@article{Mehradfar2024AICircuit,
      title={AICircuit: A Multi-Level Dataset and Benchmark for AI-Driven Analog Integrated Circuit Design}, 
      author={Mehradfar, Asal and Zhao, Xuzhe and Niu, Yue and Babakniya, Sara and Alesheikh, Mahdi and Aghasi, Hamidreza and Avestimehr, Salman},
      journal={Machine Learning and the Physical Sciences Workshop @ NeurIPS},
      year={2024}
}

@article{Mehradfar2025Supervised,
      title={Supervised Learning for Analog and RF Circuit Design: Benchmarks and Comparative Insights},
      author={Mehradfar, Asal and Zhao, Xuzhe and Niu, Yue and Babakniya, Sara and Alesheikh, Mahdi and Aghasi, Hamidreza and Avestimehr, Salman},
      journal={arXiv preprint arXiv:2501.11839},
      year={2025}
}

❓ Where to Ask for Help

If you have any questions, feel free to open a Discussion and ask your question. You can also email mehradfa@usc.edu (Asal Mehradfar).

AICircuit

Install / Use

README