Collaborative Research: CAIG: Interpretable, Stable, Mass-Conserving AI for Air Pollution Modeling

Many geoscientific models—such as those used to study air pollution or climate—are computationally intensive to run, and this limits their usefulness. Often, limited availability of computational resources limits scientific progress; additionally, these models are not usable by scientists without access to high-performance computing clusters. This work aims to increase the computational speed of these models by creating simpler versions of each model component via machine-learned (ML) “surrogate models”.

This will allow improved tradeoffs to be made between accuracy and computational cost in geophysical modeling, resulting in more accurate and efficient virtual models of Earth. At the same time, results of the project will greatly decrease the computational expertise and resources required of new model users and developers, increasing the number of people able to engage with geoscientific modeling.

Removing barriers to model use in educational and policy settings will increase the fraction of the population familiar with the workings of geoscientific models, improving public trust and perception of the transparency of models and their outputs.

Project research will be organized in three Thrusts. Thrust 1 will develop surrogate models for atmospheric chemistry—the most computationally intensive component in models of atmospheric composition—and will also develop improved dimensionality reduction methods for these systems. Thrust 2 will use the same methods to develop ML models for wildfire plume rise, which is a key determinant of wildfire smoke transport (which is in turn an increasingly important determinant of public health) but is not well characterized in current models.

Thrust 3 of the proposed project will develop an “equation-based” platform for atmospheric chemical transport modeling which expands the state of the science in performance, modularity, and differentiability for geoscientific modeling, thus allowing improved integration between physics-based and ML modeling components. This platform will also remove barriers to the broader use of geoscientific models by making models easier to use, understand, and develop.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.