Current Research

Neural networks are conventionally thought to have limited interpretability, especially within the geosciences. We use tools developed by the computer science community, such as layer-wise relevance propagation, to understand how and why neural networks make their predictions. The need for neural network explainability is two-fold: understanding the decision-making process can 1) improve trust in machine learning models when the decisions made are consistent with physical intuition, and 2) allow us to investigate the complex relationships used by the neural network to make physical insights into the Earth system.

The climate system is notoriously difficult to predict several weeks to multiple years into the future. Across these subseasonal-to-decadal ("S2D") timescales, predictions are both too far into the future for weather models to show skill, and too short for climate model projections due to noisy climate variability. To work towards overcoming this challenge, we study several predictable modes of climate variability, like the Madden-Julian Oscillation, El Niño-Southern Oscillation, and Pacific Decadal Oscillation, and seek "windows of opportunity" for which there is greater S2D predictability.

The evolving behavior of the Earth system under global warming has massive implications for regional climate, atmospheric teleconnections, and weather extremes. Our research spans from exploring the mechanisms behind large-scale changes, such as shifts in the jet stream, to identifying indicators of forced climate change amidst the noise of internal variability and climate model uncertainty. 

Climate interventions (sometimes known as geoengineering) are potential methods to deliberately intervene in the Earth system to counteract climate change in concert with decarbonization. We study a range of topics related to hypothetical stratospheric aerosol injection (SAI) intervention methods, including: the design of scenarios for modeling, potential geophysical responses including human and ecological impacts, and detectability/perceptibility in the presence of climate noise. Our research questions frequently involve interdisciplinary work on topics such as social science and ecology.

Click [here] to watch a series of short videos about our research on climate intervention.

We are always looking for new methods to integrate into our research. There are many data science tools in computer science, electrical engineering, biology, etc. that have yet to be introduced to the atmospheric sciences. Beyond explainable AI, our recent interests include various causality approaches, such as Granger and Pearl, and neural network designs, such as abstention networks. We are motivated by creativity and strive to keep science fun.