Physics-informed machine learning for glass forming ability

Supporting datasets and scripts for the paper "Evaluation of GlassNet for physics-informed machine learning of glass stability and glass forming ability", Allec et al. A preprint of the manuscript is available on arXiv.

Jupyter notebooks for reproducing the figures of the paper are in the notebooks folder. Supporting datasets are in the data folder.

Python environment setup

After activating your python environment that you will be using to run these notebooks (e.g., conda activate piml_gfa), use pip to install the requirements:

pip install -r requirements.txt

Note that installation of glasspy requires pytorch (for pytorch installation, see here).

Also, different notebooks require different glasspy versions in order to reproduce the results in the paper exactly. Please see the top cell of each notebook to see if a specific version is recommended.

Running the notebooks

All of the notebooks can be ran independently of each other; however, certain lines are time consuming and can be reduced by saving data locally. Two examples of this are i) the loading of the GlassNet training and test datasets and ii) the regression of the MYEGA equation for these datasets.

1. Loading GlassNet data

Currently, in each notebook, we use glasspy to load the GlassNet data like this:

# Import GlassNet
from glasspy.predict.models import GlassNet

# Load the model
glassnet = GlassNet()

# Then load data
glassnet_test_df = glassnet.get_test_dataset()
glassnet_train_df = glassnet.get_training_dataset()

Instead, you can do this once on your machine, and then save these dataframes as csv files and read them in later:

# Save dataframes to csv files
glassnet_test_df.to_csv('glassnet_test_df.csv')
glassnet_train_df.to_csv('glassnet_train_df.csv')

# Load dataframes from csv files
glassnet_test_df = pd.read_csv('glassnet_test_df.csv', index_col=0, header=[0, 1])
glassnet_train_df = pd.read_csv('glassnet_train_df.csv', index_col=0, header=[0, 1])

2. MYEGA regression

Currently, in any notebook that requires $\eta(T_l)$, we use GlassNet to regress the MYEGA equation:

# Import GlassNet
from glasspy.predict.models import GlassNet

# Load the model
glassnet = GlassNet()

# Read training data from csv file (this can only be done after completing 1. above)
glassnet_train_df = pd.read_csv('glassnet_train_df.csv', index_col=0, header=[0, 1])

X = glassnet_train_df.elements
Tl = glassnet_train_df.property.Tl
visc_at_Tl = glassnet.predict_log10_viscosity( T = Tl, composition = X )

Similar to the example above, you can do this once and then save the viscosity in a csv file to load later:

# Save series to csv file
visc_at_Tl.to_csv('viscosity_glassnet_train.csv')

# Load series from csv file
visc_at_Tl = pd.read_csv('viscosity_glassnet_train.csv', index_col=0)

Other notes

• For the ternary plots, the only systems for which the true glass forming region can be plotted are Na2O-Fe2O3-P2O5 and Na2O-B2O3-SiO2. This is hard-coded. If you'd like to add your own system, you will have to edit plot_ternary of notebooks/utils.py.