PROJECT I: ESTIMATING OCEAN CURRENTS FROM MEASUREMENTS OF SEA SURFACE HEIGHT (SSH)

Project Lead: Dhruv Balwada 

Project Description: One of the best ways to observe the Earth’s ocean is from space. Space-based observations allow for spatial and temporal coverage that is hard to achieve from in-situ observations, which are usually point measurements made by ships or other in-water platforms. While we routinely observe the sea level (also called sea surface height – SSH), temperature and color from space, other variables of interest, such as the ocean currents, are not directly observed. Estimating these currents is important since they transport materials, like heat, carbon, phytoplankton, and better knowledge of these currents can help to improve climate predictions, fisheries management, and oil spill monitoring. 

The LEAP momentum fellow will work on the task of estimating ocean currents from measurements of SSH. This will be in service of a new SSH measuring satellite, SWOT, which was recently launched to measure SSH at unprecedented spatial resolution. However, these new measurements come with a challenge – the conventional physics based approaches for estimating surface currents from SSH do not work at the finer scales that will be observed. The reason being that the SSH at these scales is usually dominated by internal waves, and the currents and SSH corresponding to these waves are not linked by simple mathematical relationships. Recent work has shown that ML based techniques can potentially help address this challenge (eg. Xiao, Balwada et al 2023, Wang et al 2022). 

However, important questions remain about the response of the ML models to wave amplitude, the generalization of the ML models trained on simulation data to real flows, the potential to apply transfer learning to improve predictions using in-situ observations, and the ability to estimate the errors in the predicted fields. The fellow will tackle some of these questions.

Learning Outcomes + Deliverables: A well organized github repository showcasing the jupyter notebooks and codes used during the internship. A final research paper written with the goal of submission to an academic journal.

***

PROJECT II: EQUIVARIANT GRAPH NEURAL NETWORKS FOR EFFICIENT EMULATION OF AEROSOL OPTICAL PROPERTIES

Project Leads: Kara Lamb and Carl Vondrick

Project Description: Forest fires and fossil fuel combustion are an important source of aerosols (small particulate matter) to the atmosphere. Aerosols from these sources strongly absorb sunlight, which can influence the atmospheric temperature profile, impact cloud formation, and cause snow to melt more quickly– all of which have important effects on the climate.

Understanding the role aerosols play in the climate remains a major challenge due to limitations in our modeling and observational capacities. Aerosols are a variety of sizes, shapes, and chemical compositions, all of which impact their optical properties. Atmospheric models and observational retrievals typically approximate these particles as spheres or ellipsoids, which leads to biases in determining how much sunlight aerosols absorb and scatter. Accurate methods (such as the Multiple Sphere T-Matrix Method, MSTM) to calculate aerosol optical properties for arbitrarily shaped particles are very computationally expensive, often requiring hours or days to compute the optical properties of single particles with complex shapes. 

Our recent work demonstrated that graph neural networks (GNN) are a promising approach for emulating expensive aerosol optical properties codes such as MSTM that may be capable of generalizing to new (and previously unseen) particle shapes. The goal of this summer project would be to extend and improve this GNN modeling approach by including physical constraints in the machine learning model architecture (such as equivariant neural network layers or spherical harmonic basis functions). Novel training methodologies for random orientation calculations of aerosol optical properties (calculating optical properties over all possible orientations and polarization states) will be explored. The Fellow will explore methods to improve the zero-shot performance (i.e. without re-training) of this GNN model.

Learning Outcomes + Deliverables: This research will lead to research papers submitted in scientific journals and/or in papers submitted to machine learning conferences. It would also contribute to the development of a code base to train and test machine learning methods for aerosol optical properties codes.

***

PROJECT III: INSTRUCTION + MENTORING FOR 2023 SUMMER REU (RESEARCH EXPERIENCE FOR UNDERGRADUATES) PROGRAM

Project Lead: Tian Zheng and Pierre Gentine

Project Description: Presented in partnership with the Summer at SEAS program, the DSI Scholars program and the Significant Opportunities in Atmospheric Research and Science (SOARS) program at UCAR, LEAP’s 2023 REU Program offers summer undergraduate research experiences (SUREs) on synergistic innovations in data science and climate science. The program hosts undergraduate researchers and offers a wide array of enrichment learning and networking opportunities. 

This summer, six (6) undergraduate students from around the United States will participate in the REU Program, starting with a 3-week-long Momentum Bootcamp during which they will gain familiarity and skills with the ClimSim dataset, baseline “ML Operations” such as quality controlling and data engineering, building training pipelines to perform machine learning, assessing goodness of fit, and exploring new algorithms. For those who already come with this basic foundation, advanced learning outcomes could include experimenting with stochastic and generative deep learning algorithms that attempt to fit the non-deterministic component of chaotic cloud dynamics. Momentum Bootcamp will be followed by 5 weeks of on-campus research at Columbia University, culminating in Final Research Presentations on July 27, 2023.

The Momentum Fellow will join Tian Zheng and Candace Agonafir on the Bootcamp’s team of instructors, and mentor the REU students through their on-campus research experiences.

Learning Outcomes + Deliverables: Measurable skill set in teaching and academic mentoring in the fields of climate science and machine learning. Maturation as a project manager and independent researcher with an eye toward establishment as a Principal Investigator capable of guiding the next generation of climate data science scholars.