Data-driven modeling of sea ice behavior using a stress-strain database and machine learning

Abstract

The decline in Arctic sea ice extent and thickness due to global warming has increased shipping and marine tourism, creating a critical need for offshore infrastructure capable of withstanding forces exerted by sea ice. Proper characterization of sea ice strength is essential for understanding its mechanical behavior and mitigating structural risks. This study integrates machine learning (ML) models with a comprehensive database of unconfined and triaxial compression tests to analyze peak and residual stresses in sea ice. In particular, residual stress, representing the ice's load-bearing capacity after failure, and residual strain, indicating ductility, were modeled using ML to understand the key parameters affecting their magnitudes. The analysis revealed that confinement, along with strain rate, average grain size, and test temperature significantly influence both peak and residual stresses, as well as transitions between ductile and brittle behavior. The results also demonstrated the effectiveness of ML models in capturing complex, nonlinear interactions among parameters, providing insights that traditional models cannot achieve. By addressing the limitations of conventional approaches, this work advances the understanding of sea ice mechanics and provides alternative frameworks for improving predictive modeling, ultimately informing the design of resilient Arctic infrastructure.