TLDR: A new study introduces a Physics-informed Neural Network (PINN) that integrates physical laws into machine learning models to predict Arctic sea ice velocity and concentration. This approach improves prediction accuracy, especially when limited training data is available, and offers better generalizability for future, rapidly changing sea ice conditions compared to purely data-driven methods.
The Arctic Ocean is undergoing dramatic changes, with significant declines in sea ice extent and thickness. Understanding and accurately predicting these changes, particularly sea ice velocity (SIV) and sea ice concentration (SIC), is crucial for comprehending the future of the Arctic and global climate. Traditionally, scientists have relied on complex numerical models based on physical laws or, more recently, on data-driven machine learning (ML) techniques.
However, both approaches have limitations. Physics-based models are computationally intensive and struggle to account for all the intricate variables influencing sea ice dynamics. Fully data-driven ML models, while less complex and faster, often lack generalizability and physical consistency. They depend heavily on the quantity and quality of training data, which can be problematic as the Arctic’s sea ice conditions are rapidly evolving, meaning historical data might not fully represent future states.
A recent study by Younghyun Koo and Maryam Rahnemoonfar addresses these challenges by developing a novel approach: a Physics-informed Neural Network (PINN). This innovative strategy integrates fundamental physical knowledge of sea ice directly into the machine learning model, aiming to produce more reliable and physically plausible predictions of SIV and SIC. The research paper, titled “Prediction of Sea Ice Velocity and Concentration in the Arctic Ocean using Physics-informed Neural Network,” details this method. You can read the full paper here.
Integrating Physics into Machine Learning
The core of this PINN model is built upon a Hierarchical Information-sharing U-net (HIS-Unet) architecture. This U-net is a type of deep learning model particularly effective for image data, which is relevant since sea ice data is often provided in grid formats from remote sensing. The HIS-Unet features separate branches for predicting SIV and SIC, but critically, these branches share information throughout the prediction process, enhancing overall accuracy.
To infuse physical knowledge, the researchers implemented two main strategies:
1. Physics Loss Functions: In addition to the standard data-driven loss function (which measures the difference between predictions and observations), two new physics-informed loss terms were introduced:
- A loss function that ensures sea ice velocity is zero in areas where sea ice concentration is very low (less than 15%), as SIV is not meaningfully defined in open water.
- A loss function based on the principle of mass conservation, which constrains the daily thermodynamic changes in sea ice concentration (freezing or melting) to realistic bounds, preventing physically impossible rapid changes.
2. Activation Function for SIC: A sigmoid activation function was added to the output layer of the SIC branch. This mathematically guarantees that the predicted sea ice concentration values always fall within the physically valid range of 0% to 100%.
Data and Experiments
The model was trained using daily SIV and SIC data collected from satellite observations between 2009 and 2022. Additional input variables included wind velocity and air temperature from the ERA5 climate reanalysis, along with geographical coordinates. The model used the previous three days’ data to predict the next day’s SIV and SIC.
To rigorously test the PINN’s robustness, experiments were conducted with varying amounts of training data (20%, 50%, and 100% of the 2009-2015 dataset) and different weights assigned to the physics loss functions. The models were then evaluated on data from 2016-2022, a period characterized by faster-moving sea ice and lower concentrations compared to the training period.
Also Read:
- Keeping Digital Replicas on Track: A New Approach for Simulating Complex Systems
- Unveiling Hidden Patterns in Directed Percolation with Deep Learning
Key Findings and Implications
The results demonstrated that the PINN approach significantly improved the accuracy of both SIV and SIC predictions compared to purely data-driven models. This improvement was particularly pronounced when the models were trained with smaller datasets (e.g., 20% of the available training samples). For instance, with limited training data, the PINN reduced SIV errors by up to 0.10 km/day and SIC errors by approximately 1.1%.
The benefits of the PINN were observed across different seasons, with SIV improvements being more notable in winter months and SIC improvements also significant during colder periods. Spatially, the PINN showed substantial improvements in regions like the central Arctic and near the northern Canadian Archipelago, areas critical for understanding Arctic dynamics.
These findings have profound implications for future Arctic sea ice prediction. As the Arctic continues to change rapidly, purely data-driven models trained on historical records may become less reliable. The PINN strategy, by embedding physical principles, offers a more robust and generalizable solution. It reduces the reliance on extensive historical data and can maintain consistent predictive performance even as multi-year ice coverage diminishes. This makes PINN a promising tool for understanding and forecasting sea ice conditions in a non-stationary climate.
