My research interests are centered around using statistical, data-driven, and machine learning models to study and model a variety of phenomena in weather and climate. I enjoy finding new signals in big, complex, noisy geospatial datasets, and I love thinking about how we as Earth scientists use these tools as much as I love applying them to exciting and relevant problems in weather and climate. Below, you can see some of the recurring themes that have turned up in my work throughout the years.

AI and Tropical Meteorology

Much of my work at CIRA has focused on developing machine learning models for various tropical cyclone forecasting tasks. Many of our models are developed with the end goal of producting operational forecast products for use at the National Hurricane Center or elsewhere at NOAA. Some of my work in this area includes:
• United Nations AI for Good Seminar: If you want an overview of my thoughts on AI and tropical meteorology, and many of the projects I’ve worked on, check out this United Nations AI for Good Seminar that I gave with Professor Tom Beucler in 2023.
• Tropical Cyclone Artificial Neural-network Error Model: We developed a neural network model that predicts tropical cyclone forecast errors for track and intensity. Our retrospective analyses have shown that our model has skill at track and intensity forecasting, and has the potential to reduce forecast uncertainty. This model is currently in the research to operations pipeline at the National Hurricane Center, and will hopefully be tested in real time in 2025. (Mark DeMaria's AMS Tropical Meteorology conference talk about machine learning estimates of TC track and intensity errors, Martin Fernandez's paper that is in press at Artificial Intelligence in the Earth Systems, real-time website with TCANE forecasts--should be live for the 2025 tropical cyclone season!)
• TC Structure in Synthetic Microwave Imagery: Kathy Haynes at CIRA led an effort to create synthetic passive microwave imagery from geostationary satellite imagery. Passive microwave imagers are on polar orbiting satellites, which only pass over a given location twice per day, at most; the synthetic passive microwave imagery provides synthetic microwave images at 10 minute intervals. We are now using this high time resolution synthetic passive microwave data to study tropical cyclone evolution and structure at high time resolutions that aren’t available with observed satellite imagery. (my recent talk at the AMS AI conference)
• TCBench: My collaborators at the Universite de Lausanne have led an effort to create a benchmarking dataset for tropical cyclone predictions in AI weather models. We are hoping to submit this benchmarking dataset to NeurIPS 2025. (Milton Gomez's poster at the AMS Tropical Meteorology conference in 2024).
• Tropical Cyclone Formation Probability: We are using machine learning models such as random forests and gradient boosted decision trees to create extended global tropical cyclogenesis forecasts (that is, we want to predict how likely a tropical cyclone is to form at a given location over the next 5-7 days). This project was started by Marshall Baldwin, a 2022 Hollings Scholar. (Marshall Baldwin's talk about machine learning and tropical cyclogenesis).
• Causal discovery for hurricane forecasting: Dr. Saranya Ganesh Sudheesh developed a causal discovery model that identified potential new predictors for statistical hurricane intensity forecasts. Her work identified 6 potential new predictors that are currently being tested in existing operational tropical cyclone intensity forecast models. (Tom Beucler's AMS Tropical talk about causal feature selection for tropical cyclone intensity forecasting.)

Climate Variability and Change

I started my career studying climate variability and change. While my interests have grown and expanded since then, I still enjoy studying spatial and temporal patterns of climate variability and understanding climate model scenarios. A few topics that I've worked on include:
• Jet variability in a changing climate: The jet stream plays a large role in weather and climate variability of North America, Europe, and eastern Asia. It often influences the strength and movement of large storms that regularly have wind damage, flooding, power outages, and other economic impacts. As the Earth warms, climate models predict that the jet stream will move towards the poles, which could change how these storms behave, and where they tend to have their largest impacts. We used idealized models (simple models that solve the equations of thermodynamics atmospheric motion on a rotating sphere) as well as climate models to try to better understand the physics of how and why the jet stream would respond to stimuli like volcanic eruptions or climate change. (Woollings et al. (2017), McGraw and Barnes (2016); McGraw et al. (2016))
• Extratropical–Polar Climate Variability: If you’ve ever watched a winter storm move from the Rocky Mountains towards the east, seen an atmospheric river travel across the Pacific to dump rain and snow on the West Coast, or experienced a Nor’easter, you know that the atmospheric patterns in one location can quickly impact those in another location. For some reason (a love of solitude? An Alaskan road trip the summer after college?), I’ve found myself drawn to extratropical-polar climate variability. I've studied these patterns of climate variability in climate model scenarios and reanalysis, and often used causal discovery modeling to identify feedback loops. (McGraw and Barnes (2020), Samarasinghe et al. (2019), McGraw et al. (2020))
• Polar Climate Variability and Sea Ice: My postdoc at the University of Washington was a great opportunity for me to try something new. I was lucky to get a position working with Prof. Cecilia Bitz, one of the foremost experts on sea ice and polar climate variability and change. I combined my weather and climate background with my love of statistics and data science to quantify weather models’ skill at predicting extreme sea ice loss events. While I’m not working on sea ice currently, I had a great time diving into a new-to-me aspect of our climate system with Prof. Bitz and her group, and I’m still hoping I can convince one of them to take me on a scientific expedition to Antarctica someday! (McGraw et al. (2022); Clancy et al. (2021).)
• Tropical Climate Variability: Tropical climate variability affects not just the climate of the tropics–it also affects the weather and climate patterns in the rest of the globe. I've worked on studies related to intertropical convergence zone (ITCZ) formation events in the tropical Pacific Ocean. The ITCZ has large impacts on the global circulation, and on where rain falls in the tropics, so understanding what governs changes in the ITCZ is important for weather and climate forecasting. I've also developed a prototype model for a climate-invariant tropical cyclone intensity prediction model that was scaled by maximum potential intensity (that is, the theoretical maximum intensity a tropical cyclone can reach given the background shear, humidity, and sea surface temperature). Incorporating physical knowledge can improve machine learning models of weather and climate phenomena, and make them more trustworthy. (AI for Tropical Meteorology: Challenges and Opportunities, AI for Good Seminar Series.
• Artificial Intelligence and Climate Variability: The use of artificial intelligence and machine learning in climate modeling has increased rapidly. My climate-invariant model of tropical cyclone intensity prediction uses physically-guided machine learning to try to develop a framework for tropical cyclone intensity prediction that will be robust in a changing, warming climate. I also participated in an exploratory work on using machine learning for bias correction of decadal climate predictions as part of the ClimateChangeAI 2022 Summer School. Neural network methods show promise for bias correction of climate model predictions due to their ability to learn both spatial and temporal information together--something that traditional methods can struggle to do. (AI for Tropical Meteorology: Challenges and Opportunities, AI for Good Seminar Series.

Causality, Interpretability, and Uncertainty

I've long been interested in designing data-driven models that provide insights on causality, interpretability, and uncertainty for weather and climate modeling tasks. These themes unite my work across a variety of weather and climate phenomena. While one or more of these themes is apparent in most of my work, some of my studies have focused more extensively on these topics.
• Aspen Global Change Institute Talk on Causality and Interpretability: For a brief overview of causal inference modeling and how we use it in climate science, check out a talk I gave at the Aspen Global Change Institute a couple of years ago. This talk was part of an AGCI workshop on machine learning and climate modeling. This workshop also produced an overview paper exploring the use of machine learning for climate modeling and climate science. I was one of the writers for a section on causality and interpretability.
• Using AI for Science Discovery: Kathy Haynes at CIRA led an effort to create synthetic passive microwave imagery from geostationary satellite imagery. Passive microwave imagers are on polar orbiting satellites, which only pass over a given location twice per day, at most; the synthetic passive microwave imagery provides synthetic microwave images at 10 minute intervals. We are now using this high time resolution synthetic passive microwave data to try to derive new insights on tropical cyclone evolution and structure. (my recent talk at the AMS AI conference).
• Uncertainty Quantification for Neural Networks in Earth Sciences: For most weather and climate predictions, scientists and decision makers don’t only want to see the best guess prediction–they also want to see the uncertainty around that prediction. The uncertainty is both a measure of how confident the prediction is, and an indicator of the range of possible outcomes. Some of us in the CIRA ML group wrote a paper to provide an overview of uncertainty quantification approaches for neural network models, and guidance on how to best apply these methods to Earth science tasks. (paper, and code repo).
• Causal Discovery and Climate Science: During my Ph.D., I used a variety of casual discovery models to study climate variability and teleconnection patterns (links between climate patterns in different geographic areas). I used a Monte Carlo simulation to highlight the advantages of Granger causality analysis, and then I went on to develop causal inference models for Arctic temperature and jet stream variability. More recently, I’ve participated in an effort led by Dr. Saranya Ganesh Sudheesh to use causal inference to identify new predictors for hurricane intensity forecast models. (My Ph.D. work: McGraw and Barnes (2020), Samarasinghe et al. (2019), McGraw and Barnes (2018); McGraw et al. (2020). Tom Beucler's AMS Tropical talk about causally-informed feature selection).

Understanding and Improving AI Models for Geosciences

Modeling weather and climate phenomena and trying to make weather and climate predictions are complicated, difficult tasks. These tasks require substantial domain knowledge, and working with complex and interconnected variables. I’m very interested in better adapting and improving AI models for geoscience tasks. Some problems I’ve worked on in this space include:
• Uncertainty quantification: Developing better uncertainty quantification strategies for neural network models in Earth sciences. I’ve given several talks about this topic as well, including as part of a summer school on trustworthy AI. (paper, and code repo; McGovern et al. (2023)
• TCBench: Assembling a benchmark dataset for tropical cyclone forecasts in AI weather prediction models. We hope this dataset will help AI model developers consistently evaluate how their models perform on tropical cyclone forecasting tasks that are relevant to meteorologists. (coming soon!)
• Bias in Artificial Intelligence and Machine Learning: Identifying and categorizing bias in AI models for Earth science tasks–I’m especially interested in the data and modeling aspects of this problem. (McGovern et al. (2024))
• Machine Learning and Climate Modeling: Identifying key roadblocks, issues, and opportunities for using AI and machine learning in climate modeling. (an overview paper of machine learning in climate modeling).

Extreme Event Analysis

Modeling extreme events is often a key aspect of modeling weather and climate. Extreme events are generally rare, very impactful, and can be difficult to model and analyze. Some of my work has focused on rare or extreme events, particularly for tropical cyclones and sea ice forecasting.
• Tropical Cyclone Rapid Intensification: much of my tropical cyclone forecasting work has incorporated rapid intensification forecasts–forecasts of big increases in tropical cyclone intensity that occur over a short period of time. I’ve developed a random forest model that forecasts rapid intensification in the National Hurricane Center operational guidelines and provided rapid intensification baseline models and evaluation guidelines for a tropical cyclone benchmarking dataset.
• Extreme sea ice loss: my postdoc at the University of Washington was focused on analyzing extreme sea ice loss events. We quantified the forecast skill of these extreme events in weather forecast models, with a focus on event timing, the strength and magnitude of the sea ice loss event, and the persistence of the forecast errors during and after the event. (McGraw et al. (2022))

AI and Geoscience Education

I enjoy teaching other scientists how to use machine learning, statistics, and data-driven modeling techniques in their own scientific research. Some of my efforts in this area include:
• Learning Journeys for the NOAA Committee on Artificial Intelligence: I have gotten to write a couple of Learning Journeys for the NOAA Committee for Artificial Intelligence (NCAI). These Learning Journeys are Jupyter notebooks that explore various topics at the intersection of NOAA scientific research and artificial intelligence. I’ve written notebooks on modeling tropical cyclone rapid intensification with machine learning; data preprocessing for machine learning studies; and using simple convolutional neural networks to create synthetic radar images. (rapid intensification learning journey).
• Teaching and Mentoring: Over the years, I've participated in several summer schools as well as one-on-one mentoring. Some highlights include lecturing and helping with a hackathon for the Trustworthy AI Summer School; and serving as a mentor for the ClimateChangeAI Summer School, where my team received a spotlight presentation at the 2022 "Tackling Climate Change with Machine Learning" Workshop at NeurIPS. I've also participated in one-on-one mentoring through several different programs, including the NSF REU program and the NOAA Hollings Scholar program. All three students I've mentored have gone on to pursue graduate degrees in atmospheric science, and are all currently employed in weather and climate science.