A large-scale validation of snowpack simulations in support of avalanche forecasting focusing on critical layers

Abstract. Avalanche warning services increasingly employ snow stratigraphy simulations to improve their current understanding of critical avalanche layers, a key ingredient of dry slab avalanche hazard. However, a lack of large-scale validation studies has limited the operational value of these simulations for regional avalanche forecasting. To address this knowledge gap, we present methods for meaningful comparisons between regional assessments of avalanche forecasters and distributed snowpack simulations. We applied these methods to operational data sets of 10 winter seasons and 3 forecast regions with different snow climate characteristics in western Canada to quantify the Canadian weather and snowpack model chain's ability to represent persistent critical avalanche layers. Using a recently developed statistical instability model as well as traditional process-based indices, we found that the overall probability of detecting a known critical layer can reach 75 % when accepting a probability of 40 % that any simulated layer is actually of operational concern in reality (i.e., precision) as well as a false alarm rate of 30 %. Peirce skill scores and F1 scores are capped at approximately 50 %. Faceted layers were captured well but also caused most false alarms (probability of detection up to 90 %, precision between 20 %–40 %, false alarm rate up to 30 %), whereas surface hoar layers, though less common, were mostly of operational concern when modeled (probability of detection up to 80 %, precision between 80 %–100 %, false alarm rate up to 5 %). Our results also show strong patterns related to forecast regions and elevation bands and reveal more subtle trends with conditional inference trees. Explorations into daily comparisons of layer characteristics generally indicate high variability between simulations and forecaster assessments with correlations rarely exceeding 50 %. We discuss in depth how the presented results can be interpreted in light of the validation data set, which inevitably contains human biases and inconsistencies. Overall, the simulations provide a valuable starting point for targeted field observations as well as a rich complementary information source that can help alert forecasters about the existence of critical layers and their instability. However, the existing model chain does not seem sufficiently reliable to generate assessments purely based on simulations. We conclude by presenting our vision of a real-time validation suite that can help forecasters develop a better understanding of the simulations' strengths and weaknesses by continuously comparing assessments and simulations.