Meteoclimatic drivers of rock mass plasticity before failures: Insights from an artificial neural network trained on monitoring data and weather forecasts

Detection and monitoring of strains on rock walls induced by environmental stressors, especially thermal forcing, is crucial for rock fall prediction. Machine-learning approaches, such as artificial neural networks (ANNs), can support the interpretation of long-term monitoring time series and help to identify conditions that lead to rock mass failure. This study uses data from a multi-parametric monitoring system, local meteorological observations, and weather forecasting to train ANNs to evaluate the role of meteoclimatic stressors in preparing rock masses for failure. The approach assumes that periodic or trending environmental conditions may induce progressive damage, promoting crack opening or propagation and eventually favouring instability. Two ANNs were designed to distinguish in-range elastic from out-of-range deformations, interpreted as proxies of plastic behaviour, and to identify the most relevant environmental stressors. Based on 5 years of monitoring and local weather data, ANN-1 achieved a mean area under the curve (AUC) of 0.93 on class-balanced test sets. ANN-2, trained on forecasted meteorological variables, reached an AUC of 0.83. Feature-importance analysis highlighted peak temperatures and 3-day thermal excursions for ANN-1, and forecasted precipitation probability, heating-phase thermal excursions, and multi-day thermal gradients for ANN-2. The weather-based ANN was validated at a similar site during a one-month forecast period, including an inventoried rock fall. The model returned an 86% probability of nonlinear behaviour on the failure day, anticipating the event by almost 24 hours. These results support the use of ANNs for forecasting environmentally induced nonlinear deformation in jointed rock masses.