Making Machine Learning-based Snow Water Equivalent Forecasting Research Productive and Reusable by Geoweaver

SWE (snow water equivalent) is a key indicator for monitoring and forecasting how the snow cover changes. There are many research groups exploring new methods to improve SWE forecasting, and machine learning is one of the widely studied methods. However, due to the randomness of ML, high complexity of full stack ML workflow, and poor management of experiment history, many research groups experience struggles of low productivity and the workflows are either non-reusable or non-reproducible by other scientists. To address the challenge and improve our productivity, we employed Geoweaver, a novel lightweight workflow management tool, to our team's daily routine. Geoweaver can automatically record all the history versions and logging output, and the entire workflow with history can be exported as a plain package file which can be shared freely without losing critical steps or important experiment run history. We used Geoweaver to construct the full-stack SWE forecasting workflow and the experiences so far prove that Geoweaver can significantly help in preserving full-stack ML workflows, and making them easily shareable and reproducible among team members. Geoweaver also allows us to use all the available computing resources to run the SWE forecasting while managing all the code in one place. By the end of the day, sharing Geoweaver workflow package saves a big amount of time for team members to synchronize on the progress, e.g., new ML models tested and their performance on the same testing dataset. In addition, it is very convenient to pass the Geoweaver workflow package to other colleagues in its entirety so the researchers outside our team can easily reuse or reproduce our experiments.