III: Medium: Advancing Deep Learning for Inverse Modeling

In scientific disciplines, such as earth and environmental sciences and engineering, researchers use models to understand complex systems and make predictions about future states or behaviors. For example, hydrology models are used for the prediction of streamflow in the river basin and for understanding water cycles, predicting floods and droughts, and making operational decisions such as reservoir release.

The models of these physical systems often depend on a number of parameters (characteristics) that describe the system. In the streamflow example, for instance, the important characteristics include slope, land cover, and soil type. However, for a wide variety of reasons, these parameters are often not known or poorly approximated.

Inverse modeling is a type of scientific method that involves working backward from observations of a physical system to estimate the parameters and identify the underlying processes or mechanisms that could have produced the observed behavior. Existing approaches for inverse modeling used in the physical science community often take too much time to compute and are unable to effectively leverage increasing amounts of data becoming available at continental and global scales.

The goal of this project is to develop a new generation of machine learning algorithms for inverse modeling in environmental science applications that can leverage large datasets to provide improved prediction and uncertainty estimation at decision-relevant scales while significantly reducing the amount of computation required. These methods will have wide applicability in disciplines as diverse as health, environment, agriculture, and engineering, and thus have the potential to address major societal challenges.

This project aims to develop an embedding-based machine learning framework for inverse modeling that is applicable to a wide range of scientific problems where the goal is to identify explicit or implicit characteristics of a system given its drivers and response data. This proposed methodology will introduce innovations to address challenges such as data sparsity, spatial heterogeneity, the need to handle uncertainty in data, and the ability to work with data at different scales and fidelity.

Method advancements will be made to the family of neural process methods so that they can model physical systems involving multiple interacting processes with multiple inputs and outputs. These neural process models will also incorporate scientific knowledge, such as conservation laws, as well as knowledge implicit in process-based models to generalize to out-of-sample scenarios.

A new approach will be developed to improve the process-level understanding of physical systems via a generative model based on deep latent variable methods. A graph neural network approach will be developed to learn the affinity among entities based on their inherent characteristics while also injecting scientific domain knowledge into the relationships.

Method advancements will be made to estimate and mitigate uncertainty using Bayesian deep learning for better explainability and for obtaining better distributional recovery of the input embeddings.

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.