Advances in Deep Learning Applications for Slow Earthquake Research

Slow earthquakes represent an important mode of fault slip transitional between stable creep and dynamic rupture. Their occurrence is jointly controlled by mineral composition, pore-fluid pressure, effective normal stress, system stiffness, and microstructural evolution. Because the internal state of natural faults cannot be directly observed, studies of slow slip, tectonic tremor, and low-frequency earthquakes have long been challenged by weak signals, complex noise, and discrepancies in observational scales. Building on the physical foundations of rock friction and slip stability, this review summarizes recent applications of deep learning to laboratory friction and acoustic data, natural seismic waveforms, Global Navigation Satellite System (GNSS) observations, and strain measurements, with particular emphasis on event detection, fault-state estimation, rate-and-state friction parameter inversion, and forecasting of slip evolution. Existing studies have progressed from event identification to the reconstruction of shear stress, estimation of frictional parameters, and prediction of future fault states. Nevertheless, applications to natural faults remain dominated by event detection and catalog construction, whereas parameter inversion and forecasting still rely largely on laboratory experiments or synthetic data. Physics-informed neural networks, transfer learning, reduced-order modeling, and data assimilation provide promising pathways for integrating laboratory experiments, numerical simulations, and natural observations; however, their reliability remains limited by constitutive-model dependence, parameter non-uniqueness, domain shift, and insufficient independent validation. Future work should strengthen multi-observation integration, cross-region validation, and uncertainty quantification, while developing a bidirectional framework linking laboratory experiments, numerical simulations, and natural fault observations. At present, deep learning is better suited to fault-state characterization and probabilistic assessment of slip trends than to deterministic prediction of the exact timing of slow earthquakes.