Investigate calling SURFEX fortran90 scripts in Python workflow.

rogerkuou commented 3 years ago

As a member of Team Atlas I would like to investigate how to call Fortran90 scripts in Python. This serves the data assimilation task in Global Vegetation. Where we have the DA process (as part of SURFEX) written in Fortran90, while neuron network trained by TensorFlow. We would like to:

[ ] Feed predicted ASCAT observations to Fortran90
[ ] Return updated LSM states from Fortran90 to Python
[ ] Avoid calling Python scripts in Fortran
[ ] Avoid disk IO during this process

rogerkuou commented 3 years ago

At this point the f2py from numpy seem to be a good option

rogerkuou commented 3 years ago

After discussion with Xu, we decided to use the FKB package to bind the Keras DNN to the Fortran software SURFEX to perform the data assimilation.

The main motivation to do this is to avoid rewriting the DA process, which is currently in SURFEX. FKB allows one to directly translate the DNN to a forwarding operator which can be taken directly by SURFEX.

Francesco helped Xu with the Makefile issue. Xu is able to perform the DA using SURFEX and the trained DNN now. We will discuss the results later.

NLeSC / team-atlas

Investigate calling SURFEX fortran90 scripts in Python workflow. #236