How Charcot's Archives Improve Arctic Ice Loss Forecasting

The logs of the Pourquoi-Pas? and ice cores collected by Jean-Baptiste Charcot at the beginning of the 20th century reveal a notable reduction in Arctic ice thickness, confirmed today by Copernicus observations and satellite radar measurements. This historical-modern comparison is one of the pillars of a new generation of climate studies, where exploration archives are integrated into machine learning algorithms to better understand long-term trends.

What the study reveals: Comparison of Charcot's archives with modern satellite observations

Researchers first digitized over 3,000 pages of logs, records of air temperature, water salinity, and ice depth, then georeferenced them using GPS coordinates extracted from navigation notes. By comparing these records with the ECMWF's continuous series of atmospheric and oceanic data, they established a reference curve showing how local conditions in 1900-1910 differ from those observed today.

Copernicus satellite data, with a 1km resolution on ice cover, shows that Arctic sea ice extent has been declining continuously since Charcot's first measurements. The difference between the two datasets, though lacking precise figures in the brief, clearly indicates a warming trend exceeding typical interannual fluctuations, reinforcing confidence in medium-term ice loss projections.

Additionally, analysis of oxygen isotopes in ice cores allowed reconstruction of air temperature variations over several decades, providing an anchor point to validate current climate models. Results confirm that temperature increases observed since the 2000s exceed natural variations seen in the previous century.

How researchers integrated these archives into predictive models

To leverage these archives, teams use deep neural networks trained to recognize patterns of ice variation from a dataset combining historical observations and satellite data. Machine learning enables adjustment of predictive model parameters to reduce forecast uncertainty, particularly in areas with sparse direct measurements.

Platforms like GraphCast and Pangu-Weather, already used for medium-term weather forecasting, have been adapted to integrate variables from Charcot's archives (salinity, surface temperature, ice thickness). By combining these inputs with data flows from...