Principal Investigator

Yuliya Marchetti, yuliya.marchetti@jpl.nasa.gov

Description

A machine learning framework in combination with a physical modeling system is used to provide global predictions of systematic errors, i.e. biases, between the observed values of surface ozone and physical model estimates. The main goal of the machine learning framework is to diagnose and improve a complex physical system.

Accurate global estimates of air pollutants are essential to evaluate the global public health burden of disease associated with air quality exposure, which in turn will help environmental policy making to reduce risks from associated disease and to save human lives. The experimental machine learning framework uses NASA JPL’s Multi-mOdel Multi-cOnstituent Chemical data assimilation (MOMO-Chem) estimates of ozone and the respective observed values of surface ozone from Tropospheric Ozone Assessment Report (TOAR) network to model the differences in their values.

The machine learning framework is based on the ensemble tree method, Quantile Random Forest, incorporates various methods of explainability, e.g. conditional feature contributions, SHAP, permutation importance, provides uncertainty estimates and allows to expand to new potential bias drivers through external datasets. Unsupervised learning algorithm can optionally be applied to estimate various distinct regions of interest. An integrated machine learning and analysis workflow and pipeline was implemented to scale and streamline the experimentation.

Publications and External Links

Leveraging Deep Learning for Physical Model Bias of Global Air Quality Estimates
Kelsey Doerksen, Yuliya Marchetti, Kazuyuki Miyazaki, Kevin Bowman, Steven Lu, James Montgomery, Yarin Gal, Freddie Kalaitzis
Machine Learning and the Physical Sciences Workshop, NeurIPS 2023 [link]

ML Pipeline Github repo:
https://github.com/JPLMLIA/SUDSAQ

Other sources:
https://www.jpl.nasa.gov/go/suds/seminars