Domain-informed graph neural networks: a quantum chemistry case study
I would like to bring to attention that our (Jay Paul Morgan, Adeline Paiement, and Christian Klinke) paper 'Domain-informed graph neural networks: a quantum chemistry case study ' has been accepted for publication in the journal 'Neural Networks', while we are waiting for the article to become available, we may direct you to an archive version at: https://hal.science/hal-04142152
In this paper, we investigate strategies to incoporate domain knowledge of atomic interaction processes into the design of graph neural networks. These take form of two overall strategies: an indirect but easier to implementation version of multi-task method; and a more direct but more successful method of learning modulation parameters for different edge relations of chemical bonds.
We show how these designs lead to concrete implementations in various neural network archictectures, such as message-passing, or convolutional networks.
Finally, we evaluate these different implementations and demonstrate how prior domain knowledge can help neural networks to improve in accuracy, while providing some means of explainability in results, and being more robust to tasks adjacent to ones the neural network was trained on.