Releases · sdat2/PotentialHeight

26 Nov 13:01

sdat2

v0.1.3

ecfa6c9

v0.1.3 - thesis submission day release Latest

Latest

v0.1.3:

Created ability to transform data to required dual graph format for SurgeNet training in adforce/mesh.py.
Created training datasets needed to create the SurgeNet model, by forcing the ADCIRC model with IBTrACS storms from 1980-2024 in adforce/generate_training_data.py.
Created the potential height test set for SurgeNet training in adbo/create_test_set.py.

Assets 2

15 Oct 20:55

sdat2

v0.1.2

6e8547c

v0.1.2 - Final PS calculations

v0.1.2:

Improved logic for filling in temperature and humidity profile.
Actually managed to calculate potential sizes on ERA5 Aug/Feb data 1980-2024 for the whole grid 40S to 40N by burning 2000 node hours (128 CPUs per node) on Archer2. This shows parallelization works, but still very strong constraints. 99% of compute is still on the CLE15 profile calculation.
Improved the potential size calculation in w22.ps so that both potential sizes can be calculated together.
Improved the w22.stats2 module to generate lots of new tables to describe the CMIP6 results.
Added additional filters for IBTrACS data in tcpips.ibtracs to remove storms that are undergoing extratropical transition. Seems to support hypothesis that most supersize storms are either undergoing ET or have data quality issues.

Assets 2

10 Sep 08:39

sdat2

v0.1.1

d83cb02

v0.1.1

Added tcpips/ibtracs.py script to compare observations with potential sizes and potential intensity calculated from era5 monthly averages (all post 1980).
Added more dual graph functionality in adforce/dual_graph.py so that we can train different ML algorithms with ADCIRC output, fixing bugs in imp lementation. Is pretty well tested, but takes around a minute to convert each run of ADCIRC to a dual graph (5GB of data).
pyproj now an option for all distance calculations. It is a bit slow though, and now takes 6x longer to make the fort.22.nc input file for ADCIRC. Fairly well tested, but could be improved. Adds about 2 minutes to the run time of each adforce call on archer2 (to 5 minutes from 3 minutes).
Also added a sphere option, which will is much faster. Choice is controlled in adforce/config/grid/grid_fort22.yaml by geoid option. Now defaults to sphere option due to speed.
Added ability to choose between different GP kernels and data acquisition functions in adbo/exp.py.
Explored using different GP kernels in adbo/gp_exp.
Added 1D Bayesian optimization expeirment to adbo/exp_1d.py and adbo/gp_exp.py.
Added lc12 asymmetry option to adforce/fort22.py.
Added Ide et al. 2022 curved parabolic tropical cyclone tracks to adforce/fort22.py and adforce/geo.py.
Added new python implementation of calculating the CLE15 profile at w22/cle15.py, which is at least 10x faster, has trivial parallelization, has less artifacts, but is less numerically stable than original matlab implementation. In practice on archer2, it turns out to be about 10x faster than the matlab implementation, and is now the default.
Improved figure 2 in w22/plot.py to use figure_two() function, which is now more flexible and can be used for different places, and plots using ERA5 data as well as CMIP6 data.