Multiscale, multisensor analysis of scaled seismotectonic models: Bridging the Gap Between Laboratory and Nature through Machine Learning

Despite considerable progress in monitoring natural subduction zones, key aspects of megathrust seismicity remain puzzling, mainly due to the temporally incomplete and spatially fragmented available record. Scaled seismotectonic models yield valuable insights by spontaneously creating multiple stick-slip cycles in controlled, downscaled three-dimensional laboratory replicas. Here we report recent progress in analog modeling of the megathrust seismicity, particularly focusing on a meters-scale elasto-plastic model featuring a frictionally segmented, granular fault that mimics the subduction channel at natural subduction zones. We showcase how by employing analog materials under low-stress conditions, the potentialities of monitoring can be maximized using three diverse techniques: 1)  Precise monitoring of surface spatial deformation over time is achieved through digital image correlation techniques, mirroring a uniformly distributed dense geodetic network spanning land to trench in real subduction zones. 2) A Micro-Electro-Mechanical (MEMS) accelerometric network, emulating a seismic network, captures seismic wave propagation at the model surface. 3) Embedded piezoelectric sensors within the granular analog fault capture near-field acoustic signatures of frictional instabilities. These diverse monitoring techniques allow for investigating the consistency between continuous seismic activity and surface deformation data, offering insight into both micro and macroscopic features of analog seismic cycles. At the macroscopic level, the models' frictional behavior can be numerically reproduced via rate and state numerical simulations, considering earthquake fault slip as a nonlinear dynamical process dominated by a single slip plane. At smaller scales, the model accounts for complexities in fault slip emerging from grain interactions, reflecting nonlinearities that arise when considering faults as distributed three-dimensional volumes. These fundamental attributes, coupled with their capacity to create extensive catalogs of small labquakes, make scaled seismotectonic models exceptional apparati for employing Machine Learning (ML) in comprehending multi-scale spatiotemporal seismic processes. Cutting-edge Deep Learning methods are employed to predict the spatiotemporal evolution of surface deformation, where regression algorithms not only forecast timing but also the propagation and magnitude of analog earthquakes across diverse spatiotemporal scales. Given that one of the monitoring systems used in seismotectonic analog models mimics a geodetic-like network in nature (GNSS data-Global Navigation Satellite Systems), an attempt to generalize the promising outcomes achieved in the laboratory to natural subduction faults is proposed.  Such promising avenues emphasize the potential for ML to bridge the gap between laboratory experiments and real-world seismic events. These initial findings, combined with advancements in the instrumentation of fault laboratories in nature and expanding data reservoirs, reinforce the belief that ML can significantly augment our understanding of the multiscale behaviors of natural faults.