Predicting cloud movement in 4D with SparseTensors

[NiekBongers]

Hello!

I'm working on predicting cloud movement in a 3D space over time (essentially a 4D problem, with time as the fourth dimension). My data is very sparse, as only a small percentage of the 3D atmosphere contains clouds, while the rest is filled with zeros.

To handle this sparsity, I'm using Minkowski's SparseTensors. However, I have encountered a significant challenge: Minkowski SparseTensors seem to predict only the coordinates present in the input tensor. This limitation makes it difficult to model and learn the patterns of cloud movement over time, as the output coordinates cannot evolve beyond the input.

Does someone have any suggestions on how to approach this? Are there any workarounds?

Thanks in advance!