Leveraging Predictive Models for Adaptive Sampling of Spatiotemporal Fluid Processes