Transfer learning using attentions across atomic systems with graph neural networks (TAAG)
Published in The Journal of Chemical Physics, 2022
Citation: Adeesh Kolluru, Nima Shoghi, Muhammed Shuaibi, Siddharth Goyal, Abhishek Das, C Lawrence Zitnick, Zachary Ulissi, The Journal of Chemical Physics 156 (18), 2022 https://pubs.aip.org/aip/jcp/article/156/18/184702/2841241
This research explores the use of transfer learning with Graph Neural Networks (GNNs) to generalize models across different domains in molecular and catalyst discovery. By pretraining a model on the Open Catalyst Dataset (OC20) and fine-tuning it on various datasets and tasks, including MD17, the CO adsorbate dataset, and OC20 itself, the study demonstrates that the initial layers of GNNs learn a more basic representation that can be effectively transferred to different domains, reducing the need for large, computationally expensive datasets in each domain.