Internal Processes Driving the Slow-to-Fast Transition of a Rockslide

Sibashish Dash, Michael Dietze, Qi Zhou, Jens M. Turowski, Peter Makus, Fabian Walter, Marcel Fulde, Niels Hovius

Landslides may creep slowly for decades to centuries under external influences such as rainfall or seismic shaking. Predicting when and how they transition into catastrophic acceleration remains a major challenge because the internal processes driving failure occur at depth and are often not evident from surface observations alone. Here, we combine local seismic and geodetic measurements to investigate landslide dynamics during the final 55 days preceding the collapse of the 2-million-m³ Brienz/Brinzauls rockslide (Switzerland) in summer 2023. Using a supervised machine-learning approach applied to continuous seismic data, we detect and classify subsurface cracking and rockfalls within the accelerating slope and track their temporal evolution. Initially, both cracking and rockfall activity were primarily rainfall-driven. However, 37 days before collapse, we identify the onset of runaway acceleration, marked by an increase in subsurface cracking despite the absence of rainfall. Seismic observations further indicate a progressive increase in the efficiency of basal sliding that persisted until failure. In contrast, rockfall activity responded about 8 days later, reflecting progressive degradation of the mechanical integrity of the moving mass, with the surface-to-subsurface event ratio showing a pronounced increase during the final three days before failure. These observations resolve the coupled evolution of the basal shear zone and the deforming rock mass leading up to catastrophic failure. Our results demonstrate that internal damage processes alone can sustain runaway acceleration for more than a month, even in the absence of significant rainfall, revealing fundamental slow-to-fast transition mechanics previously inferred mainly from laboratory experiments and physical models, opening new avenues for advancing future early warning approaches.