Training a neural network surrogate for an atmospheric cloud microphysics model requires processing large
training data sets in which the vast majority of the data correspond to quiescent conditions representing clear air.
When the predicted model output variables include time derivatives of the potential temperature, specific humidity,
and the cloud, rain, and snow water content, greater than 99.9% of the values lie near zero. Moreover, the probability
distribution of these variables fall nine decades, mostly monotonically, over the range of predicted values. Filtering
the uninteresting, near-zero derivative values thus progressively leaves a filtered data set that still considerably
over-samples values near zero; whereas, not filtering greatly increases training costs and over-samples the least
interesting output values. This talk will explore the utility of a sampling strategy that trains on progressively larger
samples of the training data. The samples are chosen to flatten the phase-space in the manner of an
information-entropy-maximizing distribution. The talk will examine the extent to which this strategy accelerates
convergence toward an accurate neural-network cloud microphysics surrogate model for the Intermediate
Complexity Atmospheric Research (ICAR) model.