Improved machine learning algorithm for predicting ground state properties

Install

conda create -n imp_qml python=3.9
pip install -r requirements.txt

How to use

Original Code

• All code implementing the new method, original methods (Dirichlet, Gaussian, NTK kernel), and plots are found in Code.ipynb.
• Changes to parameters (test size, shadow size, system size) can be made directly in the notebook.

Optimized Code:

• prepare dataset by executing python dataloader.py

• prepare kernels by executing python kernels.py

• to train the original methods (Dirichlet, Gaussian, NTK kernel) execute python train_kernel.py

  • you can specify the test-set fraction or shadow-size or grid size as follows python train_kernel.py --test-size 0.5 --shadow-size 500 --nrow 4
  • you can find more options in train.py@parse_args()

• to train the new method execute python train.py

  • you can again specify test-set fraction or shadow size or grid size as above: python train.py --test-size 0.5 --shadow-size 500 --nrow 4
  • to use the faster lasso library celer instead of sklearn execute: python train.py --test-size 0.5 --shadow-size 500 --nrow 4 --lasso-lib celer
  • you can find more options in train.py@parse_args()

• Look into Code_fast.ipynb, where we will use both methods to recreate the data to replicate the left plot in Fig.2

Other Files

• clean_results_old stores the results for running both new and original methods.
• heisenberg_data stores the training data used in Huang et al, 2022.
• new_data stores new training data generated with more samples (up to 500).
• visualization stores plots from running the plotting blocks of Code.ipynb.