pyPDAF - A Python interface to Parallel Data Assimilation Framework

pyPDAF is a Python interface to the Parallel Data Assimilation Framwork (PDAF) written in Fortran. The latest pyPDAF supports PDAF-V2.3.

As an interface to PDAF, pyPDAF supports all PDAF functionalities. You can use pyPDAF to construct a parallel ensemble data assimilation system in pure Python. The pyPDAF is designed as a framework that defines the workflow of given DA algorithms. Considering the versatility of the software, information on the model and observations are passed to the DA algorithms through user-supplied functions. With pyPDAF, all user-supplied functions can be implemented in Python. We expect that the coding of the user-supplied functions will be easier and more flexible than in Fortran due to the rich Python ecosystem.

pyPDAF can be used with two modes:

• online mode: DA is performed without interrupting the model program. Here, the model code is extended by calling PDAF functions to generate a single program. In online mode, the filtering gets the model state and distributes the analysis to model by in-memory exchange. This is the recommended mode for better efficiency.

• offline mode: DA is performed after the model program is finished. Here, a separate program is generated to perform DA. In offline mode, the filtering reads the model state from disk and writes the analysis to disk.

The potential applications of pyPDAF include:

• online DA systems with Python models, e.g., machine-learning models

• offline DA systems

This is a great tool for researchers who want to test and develop new DA systems. Compared to Fortran systems, the efficiency is decreased mainly from user-supplied functions and overhead for array conversions between Fortran and Python. The core DA algorithms are as efficient as in PDAF. Note that, for computational intensive user-supplied functions, the efficiency can be improved by using just-in-time compilation tools such as numba.

Here, we will provide a suite of examples and tutorials to demonstrate how to use pyPDAF. Currently, two examples are provided:

• A serial ensemble DA system using a simple wave model

• A parallel ensemble DA system using a simple wave model

Contents:

• Installation
  • Conda
  • Source code
• Parallelisation Strategy
  • Online mode
  • Offline mode
• Userguide
  • Naming conventions
  • Arguments conventions
• Design Details
  • pyPDAF wrapper to Fortran call
  • Treatment of arguments
  • Note
• API
  • Initialisation and finalisation
  • Fully parallel DA algorithms
  • Flexible DA algorithms
  • OMI functions
  • Local module functions
  • Utilities
  • Internal matrix operations
  • Internal callback functions
• API (legacy)
  • DA algorithms