AI-Driven Temperature Inference
At PGSL, we're developing innovative methods to extract crucial information about basal properties of ice sheets from radar sounding data. This work contributes to improving our understanding of ice flow dynamics driven by basal slipperiness and enhancing the accuracy of sea-level rise predictions.
The Challenge of Basal Temperature
Understanding the temperature distribution at the ice-sheet base is critical for predicting ice flow. Warmer ice is more deformable, and the presence of water at the base can lead to faster sliding over the bedrock. However, directly measuring these basal temperatures across the vast expanse of Antarctica is nearly impossible. Our research aims to address this challenge by leveraging advanced radar techniques and artificial intelligence.
Our Approach
- Led by Ph.D. student Donglai Yang , our team is developing novel AI techniques to analyze radar sounding data and infer ice sheet properties:
- Radar Data Analysis:
- We use airborne radar sounding data collected by international polar programs to estimate ice temperatures at various depths. This involves analyzing how radar signals are attenuated (weakened) as they pass through the ice, which provides valuable clues about ice temperature and other properties.
- AI-Driven Temperature Inference:
- We're developing several AI methodologies that use principles from deep learning and statistical physics:
- Physics-Informed Neural Networks (PINN): These networks incorporate known physical laws into machine learning models. We use PINNs to capture the relationship between segmented attenuation rates and depth-varying ice temperature. This approach also allows us to incorporate different englacial temperature models into our analysis.
- Conditional Normalizing Flow (C-NF): This promising technique allows us to infer basal temperatures from radar-derived depth-averaged temperatures, producing probabilistic estimates of temperatures at the ice sheet base. We're working on using this approach to generate a comprehensive map of Antarctic basal temperatures.
By combining radar observations with these advanced AI techniques, we aim to create an improved map of Antarctic ice sheet temperatures. This work will provide crucial input for ice sheet models, contributing to our ability to predict future ice flow and sea-level rise.
Our ongoing work includes:
- to better account for complex ice properties such as crystal fabric and chemical composition.
- to include other ice sheet properties that influence flow, such as subglacial hydrology and basal roughness.
- to integrate our temperature estimates into large-scale ice flow simulations.
Integration with Ice Sheet Models
At PGSL, we are developing innovative approaches to integrate radar-derived observations into ice sheet models, aiming to enhance our ability to predict Antarctic ice sheet behavior and its contribution to sea-level rise. This important work is led by Dr. Eliza Dawson, a postdoctoral researcher in our lab.
Bridging the Gap Between Observations and Models
Understanding how ice sheets will respond to climate change is crucial for accurate sea-level rise projections. However, current ice sheet models are limited by uncertainties in subsurface conditions. At PGSL, we're working on methods to incorporate radar observations into numerical ice sheet models, with the goal of improving their accuracy and predictive power.
Our Approach
- Dr. Dawson's work focuses on integrating radar-derived subsurface observations into the Ice-sheet and Sea-level System Model (ISSM), a state-of-the-art ice flow model. This research involves three main components:
- Expanding Model Capabilities:
- We're collaborating with ISSM developers to enhance the model's inverse capabilities, including the thermomechanical solver. This allows us to assimilate radar-derived subsurface observations, constraining the model with real-world data throughout the ice column.
- Joint Inversion Techniques:
- We're implementing a novel joint inversion for basal friction and ice temperature using the new Automatic Differentiation (AD) feature in ISSM. This method has the potential to efficiently optimize complex functions without explicitly deriving the adjoint, which could be a significant advancement in ice sheet modeling.
- Radar Data Constraints:
- By incorporating radar-derived estimates of temperature throughout the ice sheet column, we're exploring a data-driven method to expand ISSM's inverse capabilities. This approach aims to better represent crucial factors like basal thermal state and ice rheology in our models.
