Simulating the fractional quantum Hall effect (FQHE) with neural network variational Monte Carlo.

This repository contains the codebase for the paper Taming Landau level mixing in fractional quantum Hall states with deep learning. If you use this code in your work, please cite our paper.

Currently, DeepHall supports running simulations with spin-polarized electrons on a sphere and has been tested with 1/3 and 2/5 fillings.

Installation

DeepHall requires Python 3.11 or higher. It is highly recommended to install DeepHall in a separate virtual environment.

# Remember to activate your virtual environment
git clone https://github.com/bytedance/DeepHall
cd DeepHall
pip install -e .                  # Install CPU version
pip install -e ".[cuda12]"        # Download CUDA libraries from PyPI
pip install -e ".[cuda12_local]"  # Or, use local CUDA libraries

To customize JAX installation, please refer to the JAX documentation.

Performing Simulations

Command Line Invocation

You can use the deephall command to run FQHE simulations. The configurations can be passed to DeepHall using the key=value syntax (see OmegaConf). A simple example would be:

deephall 'system.nspins=[6,0]' system.flux=15 optim.iterations=100

In this example, we place 6 electrons on a sphere with a total flux $2Q=15$ through the spherical surface. The radius of the sphere is implicitly set as $\sqrt{Q}=\sqrt{15/2}$. This configuration corresponds to 1/3 filling. (Remember that the particle–flux relation on the sphere geometry is $2Q = N / \nu - \mathcal{S}$, where $\mathcal{S}=3$ for 1/3 filling.) The energy output includes only the kinetic part and the electron–electron interactions.

If you just want to test the installation, an even simpler example is the non-interacting case with a smaller network and batch size:

deephall 'system.nspins=[3,0]' system.flux=2 system.interaction_strength=0 optim.iterations=100 network.psiformer.num_layers=2 batch_size=100

Details of available settings are available at config.py.

Python API

You can also use DeepHall from your Python script. For example:

from deephall import Config, train

config = Config()
config.system.nspins = (3, 0)
config.system.flux = 2
config.system.interaction_strength = 0.0
config.optim.iterations = 100
config.network.psiformer.num_layers = 2
config.batch_size = 100

train(config)

Output

By default, the results directory is named like DeepHall_n3l2_xxxxxx_xx:xx:xx. You can configure the output location with the log.save_path config, which can be any writable path on the local machine or a remote path supported by universal_pathlib.

In the results directory, the file you will need most of the time is train_stats.csv, which contains the energy, angular momentum, and other useful quantities per step. The checkpoint files like ckpt_000099.npz store Monte Carlo walkers and neural network parameters so that the wavefunction can be analyzed, and the training can be resumed.

Wavefunction Analysis with NetObs

DeepHall contains a netobs_bridge module to calculate the pair correlation function, overlap with the Laughlin wavefunction, and the one-body reduced density matrix. With NetObs installed:

netobs deephall unused deephall@overlap --with steps=50 --net-restore save_path/ckpt_000099.npz --ckpt save_path/overlap

Citing Our Paper

If you use this code in your work, please cite the following paper:

@misc{qian_taming_2024,
  title = {Taming {{Landau}} Level Mixing in Fractional Quantum {{Hall}} States with Deep Learning},
  author = {Qian, Yubing and Zhao, Tongzhou and Zhang, Jianxiao and Xiang, Tao and Li, Xiang and Chen, Ji},
  year = {2024},
  month = dec,
  number = {arXiv:2412.14795},
  eprint = {2412.14795},
  primaryclass = {cond-mat},
  publisher = {arXiv},
  doi = {10.48550/arXiv.2412.14795},
  urldate = {2024-12-23},
  archiveprefix = {arXiv}
}