Jordan Aaron

Rock avalanches are composed of large volumes of rapidly flowing rock fragments that can reach velocities exceeding 50 m/s and devastate extensive areas, posing a severe threat to people and infrastructure. Numerical models are essential for forecasting the associated hazards and risks; however, the runtime of most existing models limits their ability. While Deep Learning (DL) models can potentially address this limitation, the efficiency of most numerical models limits the data available for training DL. In the present work, we have overcome this limitation using the GPU-based numerical model ORIN-3D (parallel simulatOR of landslIde motioN over 3D terrain), which provides large quantities of training data for DL. We compare LinkNet and Fourier Neural Operator (FNO) for forecasting the runout of rock avalanches, which accounts for the effects of topography, initial conditions, and path materials. Utilizing ORIN-3D simulations derived from our database comprising 20 case histories and 100 synthetic cases. Both architectures demonstrate strong parameter interpolation on known terrains (especially FNO with a True Positive Rate (TPR): 96.3%, False Positive Rate (FPR): 0.7%). While cross-terrain generalization remains challenging, FNO significantly outperforms U-Net and achieves better generalization performance when temporal dimension is included in the architecture (TPR: 86.0%, FPR: 5.1%). Both architectures achieve significantly higher forecast speeds than ORIN-3D, with FNO being roughly 1000 times faster. This dramatic improvement is crucial for applications such as landslide risk analysis, where large-scale rapid forecasts are essential.

Rock avalanches are extremely rapid, flowlike landslides that can threaten people and infrastructure, making accurate forecasts of their motion using numerical models critical for risk assessment. These forecasts are a primary input into land use and evacuation decisions; however, a comprehensive analysis of the effectiveness of various forecasting approaches applied to a wide variety of rock avalanche types has never been done. This is primarily due to a lack of field data, as well as the prohibitive computational cost. In this work, we address both limitations by applying a Bayesian posterior analysis and leave-one-out cross validation strategy to a dataset of 22 case histories. This enables us to calibrate the GPU (Graphics Processing Unit)-accelerated runout model ORIN-3D (parallel simulatOR of landslIde motioN over 3D terrain) for three rheological parameterizations: (1) Frictional, (2) Voellmy, and (3) a combined Frictional-Voellmy rheology. The analysis involves around 400,000 simulations, which is a computational effort that would be impractical with CPU (Central Processing Unit)-based models. Probabilistic forecasts for each case history are conducted using calibration results across eight forecast sets and evaluated with the Tversky loss function (TLF). The results show that the best performing model parameterization uses a volume-dependent friction angle in the source zone (Frictional rheology) and Voellmy rheology in the path zone. This work thus provides systematically validated, probabilistic forecasts of a database of 22 rock avalanche case histories, enabling improved hazard and risk assessments of this damaging phenomenon.

Susceptibility mapping is critical in assessing shallow landslide hazard and sediment transport potential. Advancements in modelling techniques and the availability of high-resolution spatial data have continuously improved the performance of landslide susceptibility maps. Nevertheless, discrepancies between predicted susceptibility and observed landslide occurrence remain. In addition to shortcomings in model design and the incompleteness of landslide inventories, the accuracy and transferability of susceptibility models are critically limited by hidden variables, such as site-specific variability in soil development, that control the triggering process but are rarely available in inventories. Here we developed an extensive case study framework, and apply it to two uniquely detailed inventories in order to quantify the role of hidden variables, as well the effects of incomplete landslide inventories. The first inventory is a comprehensive regional dataset containing over 24,000 mapped landslides across 5,939 km², and the second is a field-validated dataset of 734 landslides which includes detailed documentation of hidden variables. We trained two Random Forest machine learning models using a wide range of explanatory variables, including topography, land cover, soil properties, and climate. The first model was optimized for the first dataset, and achieved high predictive performance within its training domain (mean cross-validation of the area under the curve, AUC = 0.89). However, its accuracy decreased significantly (AUC = 0.74) when applied to the second dataset, highlighting limitations in transferability. The second model was optimized for the second dataset (AUC = 0.79). A comparison of the two models revealed that regional climatic and geologic data hindered transferability to remote regions because the relationship between available and hidden variables is not properly captured by the susceptibility model. We further analysed the predicted susceptibility values as a function of the site-specific information collected in the second database, to quantitatively explore the role of hidden variables. The analysis suggested that variables related to (i) subsurface heterogeneity and (ii) vegetation complexity govern landslide initiation, but are rarely accounted for in susceptibility models. Specifically, the models underestimated susceptibility in poorly developed soils and areas with uniform forest layering. This study underscores the necessity of a process-based understanding grounded in field observations to capture the full complexity of landslide failure mechanisms, relevant to landslide susceptibility modelling.

Susceptibility mapping is critical in assessing shallow landslide hazard and sediment transport potential. Advancements in modelling techniques and the availability of high-resolution spatial data have continuously improved the performance of landslide susceptibility maps. Nevertheless, discrepancies between predicted susceptibility and observed landslide occurrence remain. In addition to shortcomings in model design and the incompleteness of landslide inventories, the accuracy and transferability of susceptibility models are critically limited by hidden variables, such as site-specific variability in soil development, that control the triggering process but are rarely available in inventories. Here we developed an extensive case study framework, and apply it to two uniquely detailed inventories in order to quantify the role of hidden variables, as well the effects of incomplete landslide inventories. The first inventory is a comprehensive regional dataset containing over 24,000 mapped landslides across 5,939 km², and the second is a field-validated dataset of 734 landslides which includes detailed documentation of hidden variables. We trained two Random Forest machine learning models using a wide range of explanatory variables, including topography, land cover, soil properties, and climate. The first model was optimized for the first dataset, and achieved high predictive performance within its training domain (mean cross-validation of the area under the curve, AUC = 0.89). However, its accuracy decreased significantly (AUC = 0.74) when applied to the second dataset, highlighting limitations in transferability. The second model was optimized for the second dataset (AUC = 0.79). A comparison of the two models revealed that regional climatic and geologic data hindered transferability to remote regions because the relationship between available and hidden variables is not properly captured by the susceptibility model. We further analysed the predicted susceptibility values as a function of the site-specific information collected in the second database, to quantitatively explore the role of hidden variables. The analysis suggested that variables related to (i) subsurface heterogeneity and (ii) vegetation complexity govern landslide initiation, but are rarely accounted for in susceptibility models. Specifically, the models underestimated susceptibility in poorly developed soils and areas with uniform forest layering. This study underscores the necessity of a process-based understanding grounded in field observations to capture the full complexity of landslide failure mechanisms, relevant to landslide susceptibility modelling.

Susceptibility mapping is critical in assessing shallow landslide hazard and sediment transport potential. Advancements in modelling techniques and the availability of high-resolution spatial data have continuously improved the performance of landslide susceptibility maps. Nevertheless, discrepancies between predicted susceptibility and observed landslide occurrence remain. In addition to shortcomings in model design and the incompleteness of landslide inventories, the accuracy and transferability of susceptibility models are critically limited by hidden variables, such as site-specific variability in soil development, that control the triggering process but are rarely available in inventories. Here we developed an extensive case study framework, and apply it to two uniquely detailed inventories in order to quantify the role of hidden variables, as well the effects of incomplete landslide inventories. The first inventory is a comprehensive regional dataset containing over 24,000 mapped landslides across 5,939 km², and the second is a field-validated dataset of 734 landslides which includes detailed documentation of hidden variables. We trained two Random Forest machine learning models using a wide range of explanatory variables, including topography, land cover, soil properties, and climate. The first model was optimized for the first dataset, and achieved high predictive performance within its training domain (mean cross-validation of the area under the curve, AUC = 0.89). However, its accuracy decreased significantly (AUC = 0.74) when applied to the second dataset, highlighting limitations in transferability. The second model was optimized for the second dataset (AUC = 0.79). A comparison of the two models revealed that regional climatic and geologic data hindered transferability to remote regions because the relationship between available and hidden variables is not properly captured by the susceptibility model. We further analysed the predicted susceptibility values as a function of the site-specific information collected in the second database, to quantitatively explore the role of hidden variables. The analysis suggested that variables related to (i) subsurface heterogeneity and (ii) vegetation complexity govern landslide initiation, but are rarely accounted for in susceptibility models. Specifically, the models underestimated susceptibility in poorly developed soils and areas with uniform forest layering. This study underscores the necessity of a process-based understanding grounded in field observations to capture the full complexity of landslide failure mechanisms, relevant to landslide susceptibility modelling.

The behaviour of slowly moving rock slope instabilities in alpine regions is governed by the interaction between inherited structural discontinuities and externally imposed environmental forcing, yet the mechanisms linking these controls across scales remain poorly constrained due to the lack of continuous subsurface observations. Rock slope toppling represents a typical form of such structure-controlled deformation. Here, we present a comprehensive investigation of a structurally complex toppling rock slope in pre-Variscan metamorphic units in the Bedretto Valley (Swiss Alps). This landslide is intersected by the unlined Bedretto Tunnel over a length of approximately 450 m. The presence of the tunnel and the associated facilities of the Bedretto Underground Laboratory provides a unique opportunity to resolve spatially variable deformation.To understand the controls on the motion of this landslide, we installed a multi-parameter monitoring network, which integrates surface and subsurface measurements. It combines meteorological stations covering the toppling crown and toe, sectional groundwater pressure monitoring, and a 2-km-long distributed fiber optic sensing (DFOS) cable anchored along the Bedretto Tunnel. This configuration provides a unique internal view of deformation within the rockmass, enabling continuous, decimeter-scale observations of microstrain and temperature beneath up to 1500 m of overburden.The measurements reveal two distinct deformation responses of the landslide. First, reversible, centimeter-scale strain oscillations correlate with surface temperature fluctuations but exhibit anomalously high amplitudes and penetration depths, which cannot be explained solely by conductive heat transfer. This points to a non-local thermoelastic response, whereby far-field thermal stresses are generated and anisotropically transmitted through the fracture network within the rockmass. Superimposed on this cyclic thermoelastic background, the data reveal discrete, irreversible strain steps near critical fracture zones. These steps temporally coincide with major seasonal hydrologic events including sustained snowmelt and intense rainfall when a two-layer bucket model predicts corresponding peaks in groundwater storage and pressure transients. This correlation provides direct evidence for hydro-mechanically driven, progressive damage within the fracture network of the slope.Multivariate decomposition of the deformation time series isolates not only the dominant thermo-hydraulically driven cyclic signal, but also residual components characterized by spatially variable, and locally opposing, monotonic strain trends. These opposing trends are partially explained by the mechanical and geometrical heterogeneity of the fracture network and reveal the accumulation of progressive inelastic deformation. Beyond direct thermal or hydraulic forcing, such components suggest a creep-like weakening mechanism of the rockmass under quasi-static gravitational stress.These findings reveal the dynamics of a coupled thermo-hydro-mechanical system, in which seasonal forcing drives both reversible deformation and irreversible damage. The study thus highlights the critical role of discontinuities in controlling slope behavior, showing how transient hydrology and thermal cycling progressively degrade rockmass strength along pre-existing fractures and joints, ultimately weakening large rock slope failures.

In late May 2025, a series of large rock failures from Kleines Nesthorn in the Swiss Lötschental (Lötschen valley) fell directly onto the Birchgletscher (Birch Glacier), loading the latter with around 4 million m3 of rock. On May 28, following several days of acceleration, Birchgletscher collapsed in its entirety, claiming one life and causing the near-total destruction of the historic village of Blatten (which at this point was completely evacuated). Totaling more than 9 million m3 of rock and glacier ice (with a ratio of about 3:1), the rock-ice avalanche dammed the river Lonza and led to the formation of a lake that damaged additional parts of the village.To reconstruct and understand the physical processes that controlled this remarkable hazard cascade, we used aerial topographic surveys, radar and time-lapse images, direct field observations, eyewitness accounts, meteorological data, and numerical modeling. From these data we 1) determined the precise chronology of the event, including the failure and deposition volumes and geomorphologic event traces; 2) reconstructed the pre-event (1946-2023) history of Kleines Nesthorn and Birchgletscher, including the substantial mass loss of the latter and its recent surge-type acceleration; 3) analyzed the kinematics of the rock instability on Kleines Nesthorn and the resulting rock failures that loaded the glacier; 4) used the 3-D finite element model Elmer/Ice to reconstruct the effect of the rock loading on the force balance of Birchgletscher and its relevance for the observed acceleration and collapse; and 5) processed data from several long-term weather stations, satellite data and climate models to evaluate the relevance of human-caused climate change on Birchgletscher, snow-cover, permafrost and the entire process chain. Our results highlight the complexity of the Nesthorn-Birchgletscher hazard cascade and provide valuable insights for the assessment and management of glacier-related hazards in high mountains.