Review demonstrates advances in deep learning for characterizing fault states and detecting slow earthquakes, suggesting probabilistic assessments outperform deterministic prediction.
Key Points
To review the progress, computational frameworks, and limitations of deep learning applications across laboratory and natural fault observations for slow earthquake research.
Synthesized deep learning implementations across laboratory friction and acoustic datasets, natural seismic waveforms, GNSS observations, and strain measurements.
Evaluated methodological frameworks including physics-informed neural networks, transfer learning, reduced-order modeling, and data assimilation across observational scales.
Deep learning effectively reconstructs shear stress, inverts rate-and-state friction parameters, and forecasts future fault slip in laboratory and synthetic settings.
Applications to natural fault systems remain largely confined to event detection and catalog construction due to domain shift, parameter non-uniqueness, and constitutive-model dependence.
Current machine learning methods reliably support fault-state characterization and probabilistic trend estimation rather than deterministic predictions of slow earthquake timing.