Algorithm Implementation and Statistical Evaluation of a Novel Failure Criterion Governing the Release of Slab Avalanches

Sammanfattning

Fracture-mechanics-based avalanche failure modeling offers a promising approach for predicting slope-specific snow stability, as opposed to broader regional avalanche danger levels. Understanding slab avalanche failure through weak layer collapse, governed by both material strength and toughness, is rooted in the coupled criterion proposed by Leguillion. This thesis implements a novel algorithm to evaluate the failure criterion originally developed by Rosendahl and Weißgraeber, leveraging the newly published fracture toughness envelope by Adam. For the first time, crack initiation in weak snow layers can be conclusively described from the fracture mechanical point of view, forming a practical implementation for avalanche initiation using the original coupled criterion proposed by Leguillion. The algorithm implementation of the coupled criterion displayed robust performance, as the standard snow profile parameter studies matched expectations from practitioner experience well. The predictive capabilities of the mechanical model were evaluated for 748 snow profiles and associated Rutschblock (RB) tests performed in Switzerland over 18 winter seasons from 2001/02 to 2018/19. To establish a baseline, traditional statistical models, including ordinary least squares (OLS), Huber regression, and Random Forest (RF), were applied to predict snow stability from raw snow profile data. Consistent with previous research, regression techniques struggled to predict stability across the full range of RB scores, though classification of extreme stability conditions was in general possible. No benchmark model could predict the full RB range, while binary classification was possible when using only snow profile raw data, and accurate when allowing local avalanche danger as an additional input parameter. The best performing RF classifier achieved accuracy of 0.920, precision of 0.911, recall of 0.901, F1 score of 0.905, ROC AUC of 0.959, and specificity of 0.934 on average. The mechanical model showed a positive correlation between critical skier weight and RB scores. However, its predictive capability could not be statistically validated. Due to large variance in the observed relationship between critical skier weight and RB scores, full-range prediction was not possible, though classification of stability extremes was comparable to the best-performing RF classifier using raw snow profile data. Compared to the stability index SK₃₈ᴹᴸ, the mechanical model demonstrated improved performance, though it was outperformed by the Random Forest classifier of Mayer et al. From a strictly statistical perspective, the mechanical model does not provide a significant predictive advantage. This may indicate either a fundamental limitation in its ability to model micro-scale failure mechanics or an insufficient dataset for proper validation. Dataset limitations are primarily attributed to variability in manually collected snow profile data and inconsistencies in RB test conditions. To validate the mechanical model effectively, future research should prioritize controlled laboratory experiments with high-precision snow profile measurements and well-defined force-loading conditions.