Quantum dot autotuning V2

Quantum dot autotuning with machine learning V2. First version available here (private repository).

The model training and the offline autotuning simulations use the QDSD dataset, which is generated using this repository.

Install

Required python >= 3.10 and pip

pip install -r requirements.txt

Then download the QDSD dataset and unzip in into a data folder at the root of this project. The mandatory files are:

• data/interpolated_csv.zip
• data/labels.json

Settings

Create a file settings.yaml to override settings documented in utils/settings.py

For example:

run_name: tmp
seed: 0
logger_console_level: info
show_images: False
model_type: CNN
trained_network_cache_path: out/cnn/best_network.pt
dropout: 0.6
nb_train_update: 30000

Start run

Line classification task (NN train & test)

python3 start_lines.py

Note: The QDSD dataset should be downloaded and extracted in the data folder.

Offline charge autotuning

python3 start_tuning_offline.py

Note: If the trained_network_cache_path setting is not set, the script will run the line classification task first to train a new model.

Online charge autotuning

python3 start_tuning_online.py

Note: The trained_network_cache_path need to be set.

A connectors need to be implemented to communicate with the experimental equipment. See connectors/py_hegel.py for an example.

Reproduce the results of the paper

python3 start_full_exp.py --seed 42000

Note: The --seed argument could be incremented to repeat the experiment with different random seeds.

Running this script can take several days (3 days with a GPU 3070Ti).

Files structure

Repository files

• autotuning/ : The different autotuning algorithm implementations
• circuit_simulation/ : Code to generate circuit description, run circuit simulation and benchmark the results
• classes/ : Custom classes and data structure definition
• connectors/ : Interface to connect with experimental measurement tools (for online diagrams tuning)
• datasets/ : Diagrams loading and datasets in pyTorch format
• models/ : Neural network definitions in pyTorch format and baseline models
• documentation/ : Documentation and process description
• plots/ : Code to generate figures
• runs/ : Code logic for the execution of the different tasks
• utils/ : Miscellaneous utility code (output handling, settings, etc.)
• start_full_exp.py: Script to run the complete experiment benchmark (line and autotuning tasks repeated with different meta-parameters)
• start_tasks_planner.py: Script to automatize several benchmarks with grid-search
• start_lines.py : Main file to start the line classification task
• start_tuning_[online|offline].py : Main files to start the charge state autotuning task (either online or offline)

Created files

• data/ : Contains QDSD diagrams data (**should be downloaded by the user **).
• out/ : Generated directory that contains run results log and plots if run_name setting field is defined. The outputs from the paper can be downloaded here.
• settings.yaml : Projet configuration file (should be created by the user)

Workflow

Credits

• PyTorch
• Blitz - Bayesian Layers in Torch Zoo

Contributors

• Victor Yon: Main developer
• Bastien Galaup: Contributor
• Yohan Finet: Integrate circuit simulation and hardware-aware methods