This directory contains examples of how to use BubbleML that are much simpler than our experiment code. We use a downsampled version of several Subcooled Pool Boiling datasets. Each simulation has had its resolution downsampled to 1/8th size, so it is much smaller and faster to train models. These downsampled datasets have the exact same keys and metadata as the “true” datasets. The only difference is the domain resolution.
We provide two jupyter notebooks:
data_loading.ipynb shows how to load simulations using h5py, what the keys are for each hdf5 file,
and how to visualize different timesteps for each variable. It also shows how to use the distance
function to get point in liquid or vapor, or the phase interface.pytorch_training.ipynb uses the three sample datasets to train a Fourier Neural Operator.
We do not apply the training strategies used in our experiments, this is intended to serve as an
example of how to use BubbleML, not reproduce our results. This example shows how to setup a
PyTorch dataset for each HDF5 file and how to use a ConcatDataset to combine them. It then
shows how to build a Fourier Neural Operator, run a training loop, and visualize the results.The SciML conda environment includes jupyterlab as a dependency, so it can run the example notebooks. In the project root directory, you can create the conda environment with pytorch, neuraloperator, and jupyter.
conda env create -n bubble-sciml -f conda/pytorch-2.0.1-cuda-11.7.yaml