Predict Arctic Sea Ice Variability from Atmospheric River Activities and the Time of Arrival of Ice-free Arctic
PI: Laifang Li (Meteorology and Atmospheric Science, ICDS, and EESI)
Co-PI: Shujie Wang (Geography and EESI)
Project Description
The rapidly declining sea ice in the Arctic is one of the most concerning consequences of human activities. The loss of sea ice enables the ocean water originally underneath to absorb more heat and thus further promotes the melting. However, with Arctic water temperature still below freezing point, the kick-o mechanism for initial ice melting remains elusive. Research from the PI’s group found that ice melting in the Arctic occurs sporadically but usually accompanies the passage of atmospheric rivers (ARs; Zhang et al. 2023). The ARs are narrow moisture corridor, like river in the sky, which e ectively deliver water vapor into the Arctic circle. When arriving at the polar ocean, the strong wind accompanying the ARs mechanically breaks up the sea ice, and exposes more water surface to the atmosphere, which consequently allow the water vapor’s greenhouse e ect to further melt the sea ice. In addition, due to ice melting, AR events tend to last longer, and their spatial extent grows as fueled by the water vapor from the open ocean (Zhang et al. 2025). These ice melting processes associated with ARs can hardly be captured by the climate models, unless its horizontal resolution increases to kilometer scale (Chang et al. 2020). However, highresolution climate simulations are computationally expensive to perform, limiting our capability to assess Arctic sea ice variability and project its future changes. In particular, the occurrence of ARs are infrequent and long simulations must be executed to identify a handful of AR events and thus isolate their impacts on sea ice.
To overcome this limitation, we propose to apply deep-learning models (e.g., convolutional neural networks; CNN) to predict Arctic sea ice variability based on the life cycle of ARs. We will first train and validate the Artifical Intelligence/Machine Learning (AI/ML) models to predict Arctic sea ice extent using the historical Arctic AR and ice-melting events. These event logs have been archived in the PI’s research group. The predictive skills from the AI/ML will be evaluated against baseline models like anomaly persistence. The trained AI/ML models will then be applied to low-resolution model simulated AR parameters (e.g., moisture content, moisture flux, spatial extent, duration) to derive sea ice response with the passage of AR. The AI/ML derived sea ice response will be compared with the simulations by both the low-resolution and high-resolution versions of the model, so that to predictive skills from the AI/ML can be assessed. Finally, the AI/ML will be applied to predict the rate of sea ice decline in the future warmer climate when climate models unanimously project an increase in Arctic ARs (Henny and Kim 2025). Based on the AI/ML predicted sea ice decline rate, we will estimate the time of arrival of an ice-free Arctic and the associated uncertainty range.
Specific Areas of Computational and/or Data Science Expertise the current team is particularly interested in recruiting are: 1) proficient in Python programming and the ability to handle large data output from global climate models with di erent complexity; 2) ability to apply machine learning algorithms to identify patterns from observational based data sources, e.g., remote sensing of ice properties and atmospheric reanalysis products; 3) knowledge of basic statistical analysis and climate diagnostic tools.
Requirements and Expectations: The project looks for a highly motivated junior researcher with strong background training in Math, Physics and Climate Dynamics. Relevant coursework in cryosphere dynamics is desired but not required. The researcher is expected to work in a collaborative environment, bridging the existing expertise between the PI and Co-PI research groups. Research progress should be reported to the PIs in a timely and regular manner through written report or oral presentation during weekly group meetings.
Specific Objectives: The project outcome will serve as preliminary data for the junior researchers to secure fellowship grant, like the NASA FINESST and the NSF Postdoctoral fellowship. The preliminary results will also facilitate the PIs’ grant proposal for NASA MAP and NSF Polar Program. The AI/ML based sea ice prediction framework will also serve as a technical foundation to study cryosphere variability over the South Pole. We would expect one-to-two peer-reviewed journal articles from this study.
Medium to Long-Term Goals: 1) improve the forecast of polar ARs and long-lead prediction of Arctic sea ice using deep-learning models; 2) narrow down the uncertainties in estimating the time of arrival of an ice-free Arctic and quantify the role AR plays; 3) deliver an AI/MLbased sea ice prediction model that roots in the AR dynamics; 4) Secure external fundings to further the polar science research at PSU.
Connection to ICDS’s mission: The project will support the interdisciplinary collaboration between two junior faculty members by leveraging the applications of AI/ML in the intersection of climate dynamics and polar science. The project outcome will enhance our understanding of Earth system predictability, and thus our capacity to project future changes and inform science-based climate policies.
Recent and planned engagement with ICDS: The PI is an ICDS co-hire, and has actively participated in ICDS sponsored activities, like serving as poster judges for the ICDS symposium. Through this ICDS support, the PI, Co-PI and the selected junior researcher will plan to: 1) present the research findings at the upcoming ICDS symposium; and 2) host a US CLIVAR workshop on ARs at Penn State, through the joint sponsorship among ICDS, Department of Meteorology, Department of Geography, and EESI.