Modified workflow to construct a surrogate model for sensitivity analyses
We need a modified autosubmit workflow to produce N PISCES 1D simulations with parameter sets taken from a pre-computed matrix. The number of MEMBERS and CHUNKS doesn't matter, as long as MEMBERS x CHUNKS = N. Yet, the more CHUNKS we can run in parallel, the quicker the simulations will be completed.
The workflow should be similar to that used for the 2-parameter sensitivity tests in issue #3 (closed) . In particular:
- no genetic algorithm calculations are needed after each generation (MEMBER) because parameter values are read from a file.
- the output of each simulation has to be stored in a directory whose name can be easily matched to the values of the prescribed parameters.
I will provide the pre-computed matrix, which will have N rows (as many as simulations) and P columns (as many as PISCES parameters we are changing). We will start with N = 1000.
To save storage, we should reduce PISCES outputs. Current outputs for each variable have dimensions 3x3x75x(73x5), where 3x3 is the horizontal grid, 75 = vertical levels, and 73x5 = time dimension, 5 years at 5-day resolution. We want to keep only the central cell of the 3x3 grid, the top 47 levels, and the last year of the simulation (time dimension of 73).
We will keep output for the following 8 variables: PHY, PHY2, ZOO, ZOO2, POC, GOC, NCHL and DCHL.
Thanks!