Fig. 5. The flow chart of the transfer learning for the energetic co-crystals.

The energetic cocrystal prediction model is also based on the CCGNet framework involving CCGBlocks, Multi-head Global attention and dense layers. The CC dataset is first applied to pretrain the model. Then the weights pretrained on CC dataset are served as initialization weights of CCGBlocks, Multi-Head Global Attention and part of dense layers (boxes surrounded by blue dotted-lines), which is called as weight transfer. Then, the last two dense layers are initialized randomly (Gray box). Finally, ECC dataset is used to finetune all the weights of the model.