A Appendix

This means that we are free to choose any architecture for the three processes. In this section we investigate the choice of the "Forecast" architecture on the predictive performance as well as zero-shot super-resolution capabilities. Results are shown in Table 2. Increasing the number of spatial queries increases the predictive performance as expected. Moreover, having many queries also decreases the variance of the results. Figures 8 and 9 show the 1D models' predictions on each of the test set resolutions MAgNet[CNN] predictions visually match the ground-truth's For the 1D case, We use three of MPNN's PDE simulations (Brandstetter et al., 2022) as our experimental testbed.