Using scientific workflows to calibrate an Australian land surface model (AWRA-L)

The AWRA-L landscape hydrology model is one of three model components which form the Australian Water Resources Assessment (AWRA) system that aims to produce interpretable water balance estimates for Australia which (as much as possible) should agree with observations such as point gauging data and satellite observations. It is a 5 km grid-scale model, driven by interpolated climate inputs, that produces continental surfaces representing water stores and fluxes across the landscape, as well as energy and vegetation dynamics at daily timestep. The system is jointly developed by CSIRO and the Bureau of Meteorology (BoM) with model improvements proposed regularly, requiring re-calibration. This regular execution makes scientific workflows an attractive solution for engaging in our large computation and data transformation tasks whilst also providing HPC scheduling, repeatability, traceability and monitoring. Initially, various software tooling was explored to assist with the calibration of AWRA-L. These included the Catchment Water Yield Water Estimation Tool (CWYET), and various third party products such as PEST, UCODE and PGO. These tooling products were not quite suitable for use with calibrating AWRA-L largely due to the complex calibration requirements (such as the formulation of the AWRA-L objective function) and the lacking flexibility to evolve these products for future AWRA-L calibration requirements. Instead, two prototype systems were developed in Matlab: one with a targeted execution environment on a desktop; and the other with a targeted execution environment on a cluster. From these two prototypes a toolset was then developed based upon the Metaheuristics (Perraud et al, 2012) Application Programming Interface (API), which is a loosely coupled framework for assisting with model optimisation. Scientific workflow software was then used to orchestrate the flow of information between the model calibration processes. This paper describes the process undertaken to calibrate proposed changes to AWRA-L, design objectives, and current state of the suite of workflow activities developed known as the 'AWRA Calibration Tools'. Trident workbench was used to develop automated workflows for processing including: including: parameter optimisation; model simulation; and benchmarking of results. These tools allow domain experts to execute and share workflows, enabling AWRA-L to be steered in a direction that either improves the model's predictability or is considered to have an improved physical representation without significant degradation in predictions. The paper concludes by identifying key challenges that have emerged, and suggests some improvements for the future.