Research Interests
My current research uses explainable artificial intelligence to study the impacts of climate change and land-atmosphere interactions on extreme weather.
Extreme weather events (heatwaves, floods, etc.) can have destructive and costly impacts on human health, electrical infrastructure, food security, water supply, and ecosystem health. To prepare for future impacts, it’s critical that we understand (1) how extreme weather events will change in the coming decades, and (2) what physical processes may be responsible for driving these changes.
By leveraging explainable artificial intelligence techniques, we analyze large amounts of climate data to investigate how extreme weather events – and other relevant physical processes – may be impacted by climate change.
Recent Projects
Quantifying the impact of climate change on extreme weather events
On the right is a schematic showing how we use neural networks – trained on climate model simulations – to perform extreme event attribution. (1) We use historical weather data from an extreme heat event as inputs into our trained neural networks, and we make counterfactual temperature predictions under different levels of global mean temperature. (2) Then, we compare the counterfactual temperature predictions to estimate changes in the magnitude and frequency of the historical event at different levels of global mean temperature.
Using this approach, we find that global warming intensified the record-breaking 2023 Texas heatwave by up to 1.4℃ (2.5℉).
Exploring land-atmosphere interactions with machine-learning
On the right is a schematic showing how partial dependence analysis is used to derive the nonlinear relationship between soil moisture and temperature (SM-T) that has been learned by a neural network trained on historical weather data. (1) To isolate the SM-T relationship, we take a single daily 500 mbar geopotential height input map and (2) we combine this single map with every possible soil moisture anomaly input map sorted from driest to wettest. (3) We then use a trained neural network to predict temperature for these new input combinations. (4) We repeat steps (1–3) and average the behavior across all summertime GPH patterns to obtain the nonlinear SM-T coupling relationship.
