Overview
This hackathon was hosted by the National Center for Supercomputing Applications(NCSA) at UIUC. The goal was to create a model to predict certain atmospheric properties given the concentrations of 100 chemicals. The main challenge was handling high dimensional temporal-spatial data. We were given time-series data where each entry was a 3D tensor (Chemical concentrations at each point in space). Since there were 100 chemicals, our data was of dimension (# of timesteps, # of chemicals, x, y, z), with structural correlation over the time and space dimensions.
In total we had 200GB of data, so the only viable way to train the model was using multiple GPUs and resources on clusters provisioned by NCSA.