The article presents a new approach to geometric neural networks (GNNs), inspired by quantum physics. Traditional GNNs, which model physical systems as dynamic 3D many-body point clouds, are limited in their ability to capture complex relationships within these geometric graphs. To overcome this, the authors introduce a new method based on tensor networks and an equivariant Matrix Product State (MPS)-based message-passing strategy. This method is able to model complex many-body relationships and capture symmetries within geometric graphs. The approach has been validated on benchmark tasks, showing superior accuracy.
Publication date: 3 Jan 2024
Project Page: https://arxiv.org/abs/2401.01801v1
Paper: https://arxiv.org/pdf/2401.01801
