there’s still some gaps that exist, some open questions… and ideas…?

• It seems that increasing the depth of the GNN model has consistently improved its performance. What are the limitations to this approach, and how far can we go in terms of model depth for the OC20 dataset?
• Our best-performing models do not incorporate physical biases. Could incorporating such biases improve the models' performance? Previous experiments with physically inspired embeddings did not show an advantage over random initializations, so are there other ways to incorporate this type of information into the models?
• Uncertainty estimation will be important for large-scale screening in later stages of the project. How can we obtain reliable uncertainty estimates from large-scale GNNs?
• Are message-passing GNNs the only type of architecture we can use, or could alternative architectures potentially offer similar or better performance?
• Trajectories are just sequences of data points. How can we use sequential modeling techniques to model the full trajectory?