We have applied a crystal graph convolutional neural network to conduct a high-throughput screening of most possible recipe for the HOIP candidate and by investigating the property of HOIP materials. The work has been adapted from the paper "Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties", and its code that can be found on GitHub. We use the model to predict band-gap properties for HOIP crystals that were obtained from this HOIP dataset. The original model has been adapted to the following tasks:

• Predict band-gap values using pre-trained CGCNN (found in original repository)
• Train CGCNN model from scratch using only HOIP data
• Apply transfer learning, using pre-trained model and HOIP data
• Test the performance of a different loss function

Our results show:
Training from scratch brings better results, but overfits.
Training with Cross Entropy is unstable and accuracy is limited with the the number of bins the model is initialised with.
Transfer learning with discriminative learning and differential learning rates has the best performance with stable steps and no overfitting.

Please download and extract the tarball provided from the dataset source. Use the hoip_band_gap.csv(renamed to id_prop.csv) and atom_init.json file by moving it into the folder with all the cif files.