You are here:
The Well-Tempered Deposition Algorithm: Theme and Variation on Physically-based Machine Learning
Citation:
Lee, C., P. Makar, K. Toyota, C. Hogrefe, O. Clifton, M. Coyle, E. Fredj, O. Gazetas, I. Goded, L. Horváth, Q. Li, I. Mammarella, G. Manca, W. Munger, R. Staebler, E. Tas, T. Vesala, T. Weidinger, Z. Wu, AND L. Zhang. The Well-Tempered Deposition Algorithm: Theme and Variation on Physically-based Machine Learning. 22nd Annual CMAS Conference, Chapel Hill, NC, October 16 - 18, 2023.
Impact/Purpose:
This abstract and eventual presentation are based on the use of a machine learning approach to refine the representation of ozone dry deposition in Environment and Climate Change Canada's online weather and air quality model, GEM-MACH. The study uses the observational ozone dry deposition flux datasets compiled during Phase 4 of the Air Quality Model Evaluation International Initiative (AQMEII). AQMEII is a collaboration between North American and European regional air quality modelers and is being co-chaired by EPA and the European Commission Joint Research Centre. AQMEII is currently conducting its fourth phase of research with a focus on atmospheric deposition. Results from the activity are expected to help guide future development of dry deposition representation in photochemical models and inform the use of modeled deposition fields for impact assessments.
Description:
Modern machine learning approaches such as deep learning and random forests have shown remarkable progress and promising results in atmospheric modeling, but there remain valid concerns which have largely prevented their deployment in operational forecast settings. Physically-based machine learning is a recent approach where physical insights are coded into the machine learning model and taken into consideration during the training process. Providing these kinds of constraints has been shown to actually improve model representation of observational data compared to unconstrained data-based approaches in some situations. In this study, we extract the Wesely dry-deposition scheme written in Fortran from Environment and Climate Change Canada's online weather and air quality model, GEM-MACH, and translate it to TensorFlow, a modern deep learning framework which uses python as the programming language. From here we use physically-based machine learning, applying various different physical constraints to our machine learning model, in order to best reproduce ozone dry-deposition velocity data derived from field measurements at eight terrestrial sites with different land cover types (peat bog, grass, temperate mixed forest, deciduous broadleaf forest, Evergreen needleleaf forest and shrub) under different climate conditions. This dataset, compiled as part of the Activity 2 of the Air Quality Model Evaluation International Initiative Phase 4 (AQMEII-4) study, includes a wealth of ancillary data such as leaf area index, soil moisture conditions, CO2 concentrations, meteorological and micrometeorological quantities observed simultaneously at the sites. Using this training data, we are able to improve the model’s ability to represent the observations while still maintaining the overall physical structure of the model, which should make it easier to justify the use of such a model in an operational setting.