i-Melt

News

All properties can now be computed using only a single forward pass in the neural network using the model.predict_all() function. This limits the computing time and is much more efficient and thus frugal, particularly when doing MC Dropout. The old API still can be used, but it is encouraged to use predict_all(), which returns a dictionary that contains all the latent variables and observed properties.
the predict() function API has changed: several properties can now be asked passing a list to the function, for the argument methods. The old argument method still works too. Therefore, when interest in several properties (latent or observed), please use the new ``methods``argument. It will do only one forward pass in the neural network to make all the predictions. This results in better efficiency and significantly less computing time when making multiple query.
Code has been cleaned and some functions not needed for i-Melt were removed.
the Streamlit app has been updated and is now much much faster, thanks to the new predict() behavior.
Training a neural network is also faster as we can group viscosity predictions: we gained a factor 2 approximately.
All example notebooks have been updated accordingly.

Updated minimal Python version to 3.10 due to issues with older releases and dependencies.

Removed unnecessary dependencies that were causing installation issues unrelated to i-Melt.

i-Melt is now a Python package: install with pip install imelt!
New examples provided in the examples folder of the repository.
The database has been moved to ./src/imelt/data.
New helper functions available: generate_query_single and generate_query_range to simplify queries.
The results notebook has been renamed to ./examples/Replicate_2023_paper.ipynb.

Updated the results notebook Results_paper.ipynb following manuscript reviews.

Version submitted after 2nd round of minor revisions at Geochimica et Cosmochimica Acta (GCA).
Various fixes implemented.
Fixed the requirements.txt file.
Added automatic creation of the ./model/candidates/ folder in Training_single.ipynb.
Added important clarifications to README.md.

Version submitted after first round of revisions for GCA manuscript.
Various fixes and performance improvements.
Added a class to store weights of different loss functions.
Activation function type can be passed as an optional argument to the model class.
Training function now includes a “save switch” to disable model saving if desired.
Validation loss calculation during training now excludes gradient computation, reducing memory usage.
New training2() function performs K-fold dataset splitting to avoid memory issues on small GPUs (slower, but much smaller memory footprint).
New function R_Raman() allows automatic calculation of the R_Raman parameter.
Notebooks for experiments have been replaced with Python scripts for easier cluster execution.
Improved result display notebooks.
A “best” architecture has been selected for candidate training (4 layers, 300 neurons per layer, dropout 0.01) and selection (10 best networks kept for predictions).
A simple prediction notebook for single melt compositions is available: Prediction_simple.ipynb.