This project will explore and build a Graph Neural Network Accelerator on the PYNQ platform.

PYNQ - Python productivity for Zynq

GNNs are now widely used in various graph-related tasks, this project will look at how we can make GNN inference run faster using custom hardware.

At first, we would focus on small graphs:

• QM7b and QM9: QM7b contains ~7K molecules and QM9 has around ~130K molecules

MoleculeNet: A Benchmark for Molecular Machine Learning

• MD17: A variety of ab-initio molecular dynamics trajectories from the authors of sGDML. For every trajectory, the dataset contains the Cartesian positions of atoms (in Angstrom), their atomic numbers, as well as the total energy (in kcal/mol) and forces (kcal/mol/Angstrom) on each atom. The latter two are the regression targets for this collection.
• ZINC: The ZINC dataset from the ZINC database  and the “Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules” paper, containing about 250,000 molecular graphs with up to 38 heavy atoms.

Project co-supervised by Dr. Shane Fleming from AMD Research