ML based Qubit Design

Contents

Contents

• Features
• Contributions
• Quick Start
  • Environment Setup
  • Fermilab Elastic Analysis Facility (EAF) setup
  • Large Files
  • Notebooks
  • Potential optimization strategy
• Structure
• Within each folder there are three scripts
• Completed
• Desired Flow

Features

This repo has tools/options to experiment with different configurations of a Multi-Layer Perceptron (MLP) for Qiskit Metal parameter prediction, when given a desired set of hamiltonian parameters. You can use the contained notebooks to explore the following options:

1. Hyperparameter Optimization

   • Compare using Keras Tuner to automatically search for and return the best hyperparameters vs. using a predefined set.

2. Feature Scaling (0–1 normalization)

   • Option to scale all features to the range [0, 1] or leave them unscaled.
   • I strongly recommended scaling because input values can vary drastically in magnitude for these Qiskit metal design parameters, and proper scaling often improves prediction accuracy across all parameters.
   • Scaling amounts are saved, allowing you to safely reverse the scaling later and recover accurate/real-world predicted values.

3. Categorical Data Encoding

   • Compare one-hot encoding vs. linear encoding for categorical Qiskit Metal parameters.
   • Since an MLP predicts numeric values rather than categories directly, categorical features must be converted into numeric form. These experiments help determine which encoding method gives better performance.
   • Initial experimentation says one-hot encoding is better across the board

Contributions

Please contact Olivia Seidel at olivias@fnal.gov with any questions or comments.

Contributors:

• Sara Sussman (Fermilab): Ideas Contributor
• Giuseppe Di Guglielmo (Fermilab): ML expert and advice giver
• Firas Abouzahr (Northwestern): Code tester/bug finder

Quick Start

Environment Setup

Do this first. See the ENV_SETUP documentation.

Fermilab Elastic Analysis Facility (EAF) setup

A GPU is reccomended. If you have a Fermilab services account, you can use it to run jupyter notebooks on the EAF using the following instructions:

1. Navigate to https://eafdocs.fnal.gov/master/index.html and sign in with your credentials
2. Follow the three steps under Quickstart. When navigating to the link in step two, enter your credentials again
3. Click on Add New Server
4. Scroll to the bottom of server options and chose the one that says Fermilab generic notebooks
5. Select the middle of the three options to choose a GPU, and from the drop down menue chose the largest slot
6. Clone the repo there

Please feel free to contribute instructions if you use a GPU elsewhere.

Large Files

If you aren't planning on rerunning all of the notebooks and parsing data/retraining things, you will need some additional files and saved models that were already trained.

Some model checkpoints and intermediate datasets are too big for GitHub. You can grab everything you need from this Google Drive folder here.

The drive contains three directories, each containing a zip file—simply unzip them into each respective folder of your cloned repo. They’re already in .gitignore, so you won’t accidentally commit them after adding them, and you can dive straight into the notebooks without rerunning model training or the data-processing steps.

If you have any access issues please send me a quick note at olivias@fnal.gov.

Notebooks

If you have set up the environment and data, then move to the data analysis notebook (00_), then the model training (01_), etc.

To open your JupyterLab and then the notebooks:

conda activate qubit-design-env
jupyter-lab

Potential optimization strategy:

In the parameters.py file start off with KERAS_TUNER=True, and after doing an automated hyperparameter search, grab the best hyperparameter values and update the parameters.py file. Then rerun with KERAS_TUNER=False to see how the model learns over epoch number. You can also add more epochs here and fine tune hyperparameters.

Structure

The following three folders contain scripts to use Machine Learning to predict Qiskit design parameters based on target hamiltonian values for each specified design part:

• model_predict_cavity_claw_RouteMeander_eigenmode

• model_predict_coupler_NCap_cap_matrix

• model_predict_qubit_TransmonCross_cap_matrix

Within each folder there are three scripts:

• ml_00_data_analysis:loads the data and parses it into a format to use in the next script

• ml_01_train_keras: Trains the model using an MLP

• ml_hyperparameter_search_analysis: provides plots to visualize the hyperparameter search, so the user can easily see the best values

• ml_02_print_results: Loads the model and makes predictions with it

Completed

• Three models have been trained with optimized hyperparameters from keras-tuner, each model predicting Qiskit metal parameters for various parts of a transmon cross chip/resonator design
• Different encoding values were tested and optimized for the categorical y parameters in Qiskit Metal
• Different scaling techniques were implemented, where both X and Y are scaled
• Training and validation sets were explicitly seperated

Desired Flow

Below is the desired flow for this project. All three models will be stitched together to predict a Qiskit Metal design when given a set of desired "Top_Level_X" Hamiltonian values. This will be done using the "X_2.0" values that are simulated from the 'y' values predicted with each of the 3 individual models: