stopro (Stokesian Processes)

libraries for implementing Physics-Informed Gaussian Process Regression mainly on stokes flow problems

Requirements

GPU

• cuda: 11.x (5~)
• cudnn: 8.x.x (3.3~)
• nccl: 2.x.x
  Python
• python=3.11
• Summarized in requirements.txt
  • important part is "jax" versioning. 0.4.8~ will work, but currently these program doesn't use "jax.Array" so the latest jax may not work.

Installation

Clone this repository to the location where your Python can refer. In this example, we assume that this repository is clone to the directory $HOME/opt/.

cd ~/opt
git clone https://github.com/ogaken1104/stopro.git

Please make sure your python can refer to the directory ~/opt/stopro, by adding below script to your .bashrc or .zshrc.

export PYTHONPATH=$PYTHONPATH:~/opt/stopro

Structure of Program

We summarize the structure of the program. Details are described in docstrings of each script.

Modules for GP calculations

GP

• GP classes for handling all calculations about GP.
• gp.py: base class for all GP.
• gp_2D.py, gp_3D.py: base class for 2D and 3D GP. gp_3D.py inherits gp_2D.py.

solver

• optimization scheme.

sub_modules

Other modules

data_handler

• data_handle_module.py: class for loading and saving data.

data_generator

• classes for generating data for each system

data_preparer

• data_preparer.py: class for preparing data, using data_generator, data_handler

analyzer

• analysis.py: common functions for analyzing calculation results
• plot_xxx.py: plot functions for each system

Data

default_params:

• params_prepare.yaml: settings for preparing data
• params_main.yaml: settings for main calculation
• lbls.yaml: common labels for data

template_data:

• FEM or SPM references

Test

cd ~/opt/stopro
pytest ./test

By running the above command, you can test for

• flow between sinusoidal walls (forward problem)
• sin 1D naive
• Poiseuille flow
• sin 1D with Laplasian
• gaussian for 3D
• drag flow in 3D
• flow between sinusoidal walls (using explicit derivative of loss)

In tests_develop, tests code under development are stored.

Sample notebooks

In the ``sample notebooks'' directory, you can see two notebooks for detailed use of this library. (for sinusoidal one, it is only before training of the model because of the memory usage of the machine)