Feature: 1st converging cloud microphysics model

[rouson]

This pull request exhibits nearly monotonic convergence as measured by the cost function decreasing 3 orders of magnitude in the first 120 epochs:

To reproduce this behavior, execute

./build/run-fpm.sh run -- --base training --epochs 120 --stride 720

with the present working directory containing the 29.7 GB training_input.nc and training_output.nc produced for the "Colorado benchmark simulation" using commit d7aa958 on the neural-net branch of https://github.com/berkeleylab/icar, which uses the simplest of ICAR's cloud microphysics models. The Inference-Engine run uses

• A single time instant (as determined by the above stride),
• A 30% retention rate of grid points where time derivatives vanish,
• Zero initial weights and biases,
• A batch size equal to the entire time instant,
• Gradient descent with no optimizer, and
• A single mini-batch.

The program shuffles the data set in order to facilitate stochastic gradient descent. However, because a single mini-batch is used, the cost function is computed across the entire data set, which negates the value of shuffling and thus presumably makes this gradient descent.

Because a single time instant is used, this case reflects the behavior that might be expected if Inference-Engine is integrated into ICAR and training happens during an ICAR run. In such a scenario, it might be desirable to iterate on each time instant as soon as the time step completes. Doing so might either be used to

• Pretrain the network to promote faster convergence during a subsequent training session after the ICAR run using data saved from the run or
• Obviate the need for saving large training data sets for subsequent training.