How predictable are laboratory earthquakes? Insights from dense fault instrumentation and graph neural networks

Earthquake prediction remains one of the most challenging problems in Earth science. Recent advances in physics-based fault modelling, high-resolution laboratory observations, and deep-learning frameworks have opened new opportunities to assess how predictable seismic processes may be in controlled environments.

Here, we present the analysis of more than 1,000 laboratory earthquakes produced in a biaxial apparatus hosting a 400 × 100 mm PMMA fault interface, allowing two-dimensional rupture propagation analogous to natural faults. The experimental setup is instrumented with 38 strain gauges distributed within the fault interface, 20 accelerometers located along both fault surfaces, and 14 acoustic emission (AE) sensors positioned at varying distances from the fault. The experiments were conducted under constant loading rate and normal stresses ranging from 50 to 250 bar. This dense instrumentation enables us to reconstruct, for each laboratory earthquake, the nucleation location, initiation time, rupture evolution, and final event magnitude. The resulting catalog spans nearly three orders of magnitude in seismic moment (from M_w=-6.5 for small ruptures to  M_w=-3.8 for the largest events).

Building on this comprehensive dataset, we explore the potential of Graph Neural Networks to predict the spatial and temporal occurrence of laboratory seismicity. The models are trained on a subset of experiments and tested on independent experiments not included in the training phase. We focus in particular on identifying the minimal set of observational features required for successful prediction, and on assessing the level of physical complexity that machine-learning algorithms trained on homogeneous laboratory faults can capture.