Predicting compressive strength and behavior of ice and analyzing feature importance with explainable machine learning models

Abstract

Building and using ice-related models is challenging due to the complexity of the material. A common issue, shared by both material models and (semi-)empirical approaches, is estimating unknown input parameters such as compressive strength. This is often done with additional empirical formulas which have drawbacks, e.g., they are based on a limited amount of data. Regarding material modeling, a strongly related problem is the prioritization of effects to include in the model. This is mostly done based on a subjective mix of knowledge, model purpose, and experimental studies limited to that purpose, which risks overlooking effects or interaction of effects, and limits transferability of material models to other scenarios. To tackle these issues, a hybrid approach of domain knowledge and explainable machine learning was used. A large ice test database was compiled to train machine learning models to predict compressive strength and behavior type. The machine learning models’ predictions were more accurate than existing empirical or analytical approaches and can thus be used as an alternative, though less straightforward, tool for such predictions. Further, the SHAP explainable AI method was applied to the predictions. Impact rankings of experimental parameters and interaction effects between parameters were analyzed and discussed in terms of ice mechanics. Top features were strain rate, triaxial stress state, temperature, and loading direction, but impact rankings were highly dependent on prediction target and type of ice. Few interaction effects were found. The approach adds objectivity to the prioritization of effects for material modeling and generated further insights into ice mechanics. It is also considered useful for other natural materials or generally when there is more data than knowledge.