nikosavola / neuralnetworkquantumstate

Group work for Solid State physics course at Aalto University

Neural Network Quantum State

Group work for PHYS-E0421 - Solid State Physics course at Aalto University.

Check notebook.ipynb for running the model and results. Source code for the problem, model, and Ray Tune compability is in model.py.

Clusters with Slurm can be used to run the hyperparameter optimisation and parameter sweeps in parallel. See the .sbatch files for more information.

Installation

Linux and macOS

Install dependencies with:

pip install -r requirements.txt

For optional GPU support, install CUDA and cuDNN. Afterwards run:

pip install --upgrade "jax[cuda]" -f https://storage.googleapis.com/jax-releases/jax_releases.html

For optional but recommended MPI support for CPU and GPU, install netket with MPI deps.:

pip install --upgrade "netket[mpi]"

With GPU, this requires a correctly setup CUDA compiler, so you might have to do something like export CUDA_PATH=/usr/local/cuda/.

Windows

Not recommended natively. Use WSL instead and follow the Linux approach. If you really want you can install w/o mpi with an unofficial jaxlib:

pip install jaxlib -f https://whls.blob.core.windows.net/unstable/index.html
pip install -r requirements.txt

neuralnetworkquantumstate's Issues

Check proper way to implement AKLT Hamiltonian

Currently spin-1/2 Pauli Z-operators are used, should probably by something else. Additionally, is a sum rule needed?

Try RBMMultiVal

…and explain learning of translational symmetry in Overleaf

Settle on Neural network architecture

Currently, RBM for real and complex values is used. Other possibilities exist as well, but this might be okay.

Afterwards, hyperparameters should be swept, and plots or tables shown which show how the best combination of parameters was chosen.

Update Overleaf LaTeX theme

...to correspond to https://github.com/nikosavola/HomeworkTemplateLatex

Make parallel plot

Use HiPlot from facebook to create something like

Write more about J1-J2 problem

…in Overleaf

Try some new architecture

Probably NetKet GCCN or other convolutional