0a3b5a7a477d359746061d41c3a04fd6-Supplemental.pdf
–Neural Information Processing Systems
For a given elevation map, there is an infinite possible combinations of initial and boundary conditions that could potentially realize in future events. It is an interesting question how to automatically configure the most relevant initial and boundary conditions to train on, to getarepresentation that will beuseful inpotential future real-world scenarios. First, we find the lowest point on the boundary of the elevation map, and define this point as the outflux location. Then, the influx location is defined to be the middle of the opposite axis of the chosen outflux axis. We define it this way to get a large flood extent, by constraining the water tocross the elevation map before itcan flowout.
Neural Information Processing Systems
Feb-7-2026, 10:35:27 GMT