PSST (Polymer Solution Scaling Theory) Library

The psst library allows users to study and implement our deep learning scaling theory methods and test similar approaches.

psst Module

The novel parts of the code are grouped into the module psst, which will be added to PyPI in the near future. For now, the library can be installed using the command pip install . in the head directory (the directory containing pyproject.toml). Dependencies will be handled by pip. Users may wish to create a new virtual environment first.

• The models submodule contains two convolutional neural network (CNN) models we used in our initial research, models.Inception3 and models.Vgg13.
• psst.configuration defines the reading and structure of the system configuration, as specified in a YAML or JSON file.
• psst.surface_generator contains the SurfaceGenerator class that is used to procedurally generate viscosity curves as functions of concentration (phi) and chain degree of polymerization (Nw). The submodule also contains normalize and unnormalize functions. Normalizing transforms the true values, optionally on a log scale, to values between 0 and 1.
• psst.training contains the train and validate functions, as well as checkpointing functionality for the model and optimizer.

NOTE: The normalize/unnormalize functions and the checkpointing functionality may move to different submodules, perhaps new files.

Other Directories

The examples directory contains scripts to optimize and train networks, and one to evaluate experimental data. These are similar to the scripts used during our research. Details are in examples/README.md.

The deprecated directory contains deprecated code that may be useful in the future.

The doc directory contains a markdown file of the mathematical derivation transforming the details in the original publications to the representations used in the code.

The img directory will contain images used in this README and in the derivations.md file.