Forecasting earthquake rupture characteristics with deep learning: a proof of concept using analog laboratory foamquakes

In recent years, machine learning has been used to predict earthquake-like failures in various laboratory experiments. The predictions of these approaches have been framed with both regression and classification. Labquakes prediction in direct shear experiments has been achieved by predicting the time to failure of the sample (regression). Similarly, for laboratory analog subduction models, the time to failure has been successfully predicted. In the classification approach, demonstrated on analog models, a time window of “imminence” is predefined and the model determines if failure occurs within this time window or not. These previous approaches suffer from the problem of thresholding: in time-to-failure regression, there is the need to define a velocity or displacement that signals an event has occurred, in imminence classification the choice is the time window that we consider an event to be imminent. Here we remove this thresholding problem by taking a spatiotemporal regression framing that forecasts future surface velocity fields from past ones. In such a framing, the whole seismic cycle is forecast (i.e., interseismic, coseismic, and postseismic). We test this approach on Foamquake.

Foamquake is a novel 3D elastoplastic seismotectonic analog model mimicking the key features of the subduction megathrust seismic cycle in a scaled manner. Foamquake features a wedge-shaped elastic upper plate made of foam rubber. The analog megathrust includes a velocity weakening, rectangular patch embedded in a velocity neutral matrix. Plate convergence is imposed kinematically with a motor-driven belt (analog of the subducting plate) underthrusting the wedge. Foamquake experiences quasi-periodic cycles of stress accumulation and sudden drops through spontaneous nucleation of frictional instabilities. These labquakes are characterized by coseismic displacement of a few tens of meters when scaled to nature and source parameters (seismic moment-duration and moment-rupture area) scaling as real subduction interplate earthquakes. The 3D nature of Foamquake allows running models with two asperities along strike of the subduction zone divided by a barrier. This configuration generates sequences of full and partial ruptures, superimposed cycles, and nested rupture cascades: complex patterns similar to those inferred at natural megathrusts, representing the perfect testbed for developing new prediction strategies.

In particular, we step toward forecasting seismic cycle full surface velocity fields using deep-learning-based approaches from the Computer Vision field. This framing allows simultaneously to forecast the onset of a labquake and illuminate its space-time evolution at different prediction horizons. A variety of deep-learning algorithms have been tested and compared with Random Forest models (which we consider as a baseline machine learning model). We show that Convolutional Recurrent Neural Networks, with spatiotemporal sequences of surface velocities as input, perform the best in forecasting. Preliminary results suggest that the onset and the spatio-temporal propagation of individual lab-quakes can be predicted with relatively high accuracy at prediction horizons that are in the same order of labquake durations. Surface velocities at further horizons than labquake durations appear unpredictable. This study introduces an innovative framing of the earthquake forecasting problem which can open new perspectives for application to natural observations.