OVERALL WORKFLOW AND METHOD:

Autoencoder+MLP Framework:

With this code, one can perform the hybrid machine learning, unsupervised learnig with Autoencoder (AE) and predict storage capacity via a MLP for a hydrogen storage data given in the file train.dat. Here are the descriptions of the programs:

1. B1.py: Uses AE+MLP approach
2. MLP-B1.py: property predicts directly with the MLP without any feature transformation.
3. Corr.py: Studies the Pearson correlation between the features in the latent space and the real features.
4. train.dat: Training data for 1483 materials [First column target, next 36 columns features]

The directory Unknown-materials contains:

1. Program U.py to predict to the hydrogen storage capacity for the materials from only features:
2. set1.dat:[ TiAlN2, V2H2, Zr2TiAl, MgC, NLi]
3. set2.dat: [NMn2Ti,MgCHF, CAlB, MgCHF, MgMnVTi ] ->copy these files to unknown.dat and run U.py

The directory LLM contains:

1. Script to train the GPT-2 model and to save in a directory called TrainedModel. (python GPT-2.py)
2. Generate chemical formulas based on the loaded model (python Generators.py)
3. It should be noted that the generated materials depend on the parameters used.
4. To generate more materials at a time, change the parameters.