Applications of machine learning for modelling of ice flexural strength

The design of marine vessels and structures operating in regions where ice is present, must consider the loads transferred to the structure upon impact with an ice feature. The flexural strength of ice is an important material property and can have significant impact on the loads transferred to a structure. Flexural strength is generally considered to be dependent on the size or scale of the sample (often reported as beam volume), ice temperature and brine volume (in the case of sea ice), however the influence of temperature and beam volume have been debated in the literature. Conventionally flexural strength was often modelled as a constant (i.e. average strength), or was modelled as a single parameter or dual parameter (sea ice only) empirical relationship. Employing an extensive database of flexural strength measurements, with over 2000 freshwater and 2800 sea ice measurements, machine learning (ML) algorithms were utilized to define a relationship between these ice parameters and the measured flexural strength. The implementation of ML algorithms was able to highlight a link between freshwater flexural strength and ice temperature, a relationship often ignored or not perceivable in existing models. When considering sea ice, the use of ML algorithms were able to highlight a dependence of flexural strength on scale, brine volume and temperature. These findings have the potential to impact the design of ice strengthened structures, and highlights the importance of accurately recording these parameters when performing tests in the either the field or laboratory.