Physically-Scaled Machine Learning for Tropical Cyclone Intensity Prediction

• Incorporating physical knowledge can improve machine learning predictions and ground them in physical reality. I am collaborating with the Data-Driven Atmospheric and Water Dynamics Group at the Université de Lausanne to develop and apply a physics-based scaling for tropical cyclone intensity prediction. We are exploring the ways in which we can leverage maximum potential energy theory to develop a physically consistent scaling for tropical cyclone intensity that will hopefully lead to more generalizable, flexible, and robust predictions of tropical cyclone intensity.

Forecasting Tropical Cyclone Intensity Using Machine Learning

• Tropical cyclone forecasts generally consist of two parts--the track (where the storm is headed) and the intensity (how strong the strongest winds will be). Tropical cyclone forecasters at institutes such as the National Hurricane Center regularly use simple statistical models to predict tropical cyclone intensity, in addition to more complex numerical weather prediction models. Thus, we at CIRA and at AI2ES are hopeful that machine learning models can provide useful forecasts of tropical cyclone intensity, and that we can develop machine learning-based tools that will help forecasters make even better predictions. In addition to having good prediction skill, we need to make sure our models are explainable, so the forecasters understand how the model makes predictions, and are able to apply their weather expertise to interpreting the model output. A paper on this topic will be coming shortly; in the meantime, check out a recent seminar I gave with Tom Beucler on artificial intelligence and tropical meteorology.

Uncertainty Quantification for Neural Networks in Geosciences

• Neural networks are becoming a popular analysis and prediction tool for many geoscience applications. However, in addition to making a prediction of the mean state, many geoscientists also want to be able to estimate the uncertainty of their predictions, in order to estimate how confident we can be in our predictions, as well as to better understand the full range of possible outcomes. Scientists have developed several novel methods for estimating the uncertainty of a neural network's prediction. We are comparing several common uncertainty quantification methods and their performance in geoscience tasks, such as making predictions from satellite imagery, to try to understand the pros and cons of various uncertainty quantification methods in geoscience applications [paper, code].

Predicting Extreme Sea Ice Loss on Day-to-Day Timescales

• Predictions of sea ice loss are often focused on longer timescales, but sea ice variability is also important on daily timescales. When extreme sea ice loss occurs over timescales of a few days, this sea ice loss can have large impacts on Arctic weather, as well as local communities that use sea ice for hunting and travel. Since there is some indication that these extreme sea ice loss events are driven by consistent atmospheric patterns, we are exploring whether or not weather forecasting models are able to skilfully predict these extreme events. For more details, check out my Github repository for this project, written up here, as well as Robin Clancy's related work on sea ice and Arctic cyclones. If you are excited about sea ice prediction in general, check out the Sea Ice Prediction Network.

Applying Causal Discovery Methods to Atmospheric and Climate Science

• As part of my Ph.D., I worked on making causality-based tools more accessible to atmospheric and climate scientists. I worked closely with colleagues in electrical and computer engineering to refine and apply causality-based methods and algorithms to climate science problems, detailed here. This collaboration also heavily influenced my work on two-way Arctic-midlatitude feedbacks (here).


• In addition to using causality tools to produce new scientific results, I have also emphasized clear and effective communication of these methods and approaches to make them more accessible to the atmospheric and climate science community. This desire to improve accessibilty was the primary driver behind my statistical discussion of the advantages of Granger causality in climate science research (here). I am currently working on making the analysis in this paper publicly available; in the meantime, I helped my colleague Savini Samarasinghe create an example based on my Monte Carlo simulations. We discuss this further over at the DATAS gateway. More recently, I've been working with Prof. Alex O. Gonzalez and his students on applying this analysis to understand the dynamical feedbacks in the Intertropical Convergence Zone (here ).
  

Understanding the Two-Way Nature of Arctic-Midlatitude Feedbacks

• Midlatitude weather and climate, including the jet streams, can be influenced by the weather and climate of the Arctic, and vice versa. Much of my Ph.D. work was focused on understanding the relationships between Arctic and midlatitude weather and climate using a two-way perspective. My colleague Savini Samarasinghe and I worked together using multiple different casual discovery-based methods to highlight the two-way nature of these connections (here). I expanded on this work by highlighting the important role that the midlatitude circulation plays in influencing Arctic weather on sub-monthly timescales (here). I also co-led a study with colleague Cory Baggett that highlighted how sea ice loss could impact midlatitude moisture transport into the Arctic, which further feeds back upon Arctic weather and climate(here).
  

Jet Variability and Climate Change

• The jet streams play a substantial role in the storms, precipitation, weather, and climate of the midlatitudes, and understanding how jet streams might change in a warmer atmosphere is key to understanding the future weather and climate. I used a set of idealized atmospheric model simulations to highlight the important role of seasonality in the jet repsonse to climate change (here). I used the same idealized model to look at how the relationship between the speed and position of the jet streams in these simple models is different in different seasons (here). Simple model experiments like these can help scientists develop a stronger understanding of the basic physics of the atmosphere, which we can then apply to studies of real-world observations or more complicated climate models.

Volcanic Eruptions and the Large-Scale Circulation

• Large volcanic eruptions can have a large impact on global weather and climate for as many as 3 years following the eruption; however, large eruptions are infrequent, and thus can be difficult to study. I used over 200 climate model simulations to fully characterize the variability of the large-scale atmospheric circulation response to large volcanic eruptions, and to analyze the role of large-scale climate forcings such as El Niño in this response (here).