Abrupt frictional fault failure is normally accompanied by acoustic emission (AE)—impulsive elastic wave broadcast—with amplitude proportional to particle velocity. The cumulative sum of the fault particle velocities is a slip displacement. In laboratory shear experiments described here, measurements of a sequence of laboratory earthquakes includes local measurement of fault displacement and AE. Using these measurements we illuminate the connections between “cumulative sum of AE” and “slip displacement“. Additionally, the composition of the AE broadcasts reveals inhomogeneity in the fault mechanical structure from which they arise. This inhomogeneity, leading to a time invariant white AE component and an articulated AE, indicates that the articulated cumulative sum of the AE reveals a fault “state of the mechanical structure” diagnostic, that follows a distinctive pattern to frictional failure. This pattern explains why the continuous AE map to fault displacement as well as fault friction, shear stress, etc., as shown in many recent studies.

Machine learning models using seismic emissions can predict instantaneous fault characteristics such as displacement in laboratory experiments and slow slip in Earth. Here, we address whether the acoustic emission (AE) from laboratory experiments contains information about near-future frictional behavior. The approach uses a convolutional encoder-decoder containing a transformer layer. We use as input progressively larger AE input time windows and progressively larger output friction time windows. The attention map from the transformer is used to interpret which regions of the AE contain hidden information corresponding to future frictional behavior. We find that very near-term predictive information is indeed contained in the AE signal, but farther into the future the predictions are progressively worse. Notably, information for predicting near future frictional failure and recovery are found to be contained in the AE signal. This first effort predicting future fault frictional behavior with machine learning will guide efforts for applications in Earth.

Low frequency earthquakes (LFEs) originating below the central San Andreas Fault are associated with slow-slip within the more ductile portion of the crust beneath the seismogenic zone. Monitoring efforts over 15 years recorded >1 million LFEs with >70 per day. We apply machine learning (ML) to statistical features describing the seismic waveforms and estimate the LFE daily intensity. Using 4 years of independent data, the ML model produces a 0.68 correlation. The burst-like LFE behavior is reproduced and the largest misfit occurs during the low-amplitude daily undulations. The ability to continuously monitor LFE activity provides insight to when geodetic measurements of slow slip are possible, without the need for developing a computational-intensive template-matching catalog. Similarities are found between detecting LFEs and tremors, which provides evidence tremors are composed of LFEs. The approach reveals by ML the rich information contained in the features of continuous seismic waveforms.

The stress tensor is an important property for upper crustal studies such as those that involve pore fluids and earthquake hazards. At tectonic plate scale, plate boundary forces and mantle convection are the primary drivers of the stress field. In many local settings (10s to 100s of km and <10 km depth) in tectonic plate interiors, we can simplify by assuming a constant background stress field that is perturbed by local heterogeneity in density and elasticity. Local stress orientation and sometimes magnitude can be estimated from earthquake and borehole-based observations when available. Modeling of the local stress field often involves interpolating sparse observations. We present a new method to estimate the 3D stress field in the upper crust and demonstrate it for Oklahoma. We created a 3D material model by inverting multiple types of geophysical observations simultaneously. Integrating surface-wave dispersion, local travel times and gravity observations produces a model of P-wave velocity, S-wave velocity and density. The stress field can then be modeled using finite element simulations. The simulations are performed using our simplified view of the local stress field as the sum of a constant background stress field that is perturbed by local density and elasticity heterogeneity and gravitational body forces. An orientation of N82˚E, for the maximum compressive tectonic force, best agrees with previously observed stress orientations and faulting types in Oklahoma. The gravitational contribution of the horizontal stress field has a magnitude comparable to the tectonic contribution for the upper 5 km of the subsurface.

Convolutional Neural Networks (CNNs) can detect patterns that are otherwise difficult to identify and have been shown to excel in predicting fault characteristics in laboratory shear experiments and slow slip \emph{in situ}. Here we show that a suitably designed CNN can be trained to identify some precursory change in the seismic signal preceding some large natural earthquakes by up to a few hours, with a variable success rate. We use 65 $\textrm{M}_w\geq 6$ events in the NE pacific in and around Japan from 2012 to 2022. By repeating the training/testing cycle with variable random initial weights, we obtained up to 98\% in training accuracy and 96\% in testing accuracy in discriminating noise and precursor windows. In the $\sim 3$ hours preceding the earthquakes, the network identifies precursors progressively more frequently as earthquake time approaches. A final subset of more recent seismic events was used for a further verification, with mixed results. While the network appears to differentiate noise and precursor with a statistically positive incidence, the results are highly variable depending on the events that are analysed, with poor potential for generalisation. This may indicate that not all earthquakes in the catalog contain precursor signals, or at least no signal similar to the training subset. Discriminative features between precursor and noise windows appear most dominant over a frequency range of $\approx$ 0.1-0.9 Hz (in particular $\approx$0.16 and $\approx$0.21 Hz) broadly coinciding with observations made elsewhere of microseismic noise and broadband slow earthquake signal \cite{masuda_bridging_2020}.

Earthquake detection is critical for tracking fracture networks and fault zone deformation, particularly microseismicity that produces weak ground motions. We develop deep learning models to detect seismic phase arrivals and first motion polarities. The detection model is a convolutional encoder-decoder with a multi-head attention latent space that assigns a softmax value to each data point in continuous seismic records for classifying earthquake waveforms and the phase arrivals. The multi-output classification model utilizes weighted categorical cross entropy for the different softmax predictions to account for the unbalanced number of signal points compared to noise. The model training uses a benchmark data set of global seismic waveforms and the events are augmented using various techniques to reduce the signal-to-noise ratio, simulate multiple events arrivals, and channel failures. Detected p-waves are passed through a second model to obtain the first motion polarity. The phase arrivals, first motions, arrival waveforms, and additional metrics needed for catalog development are saved in a detection table. A neural network phase associator is used with the detection table to build an event arrival table. Locations are calculated and double difference locations are produced using correlation metrics from the waveforms retained in the detection table. The analysis is wrapped in a multiprocessing workflow to efficiently analyze large data sets. As a case study the workflow is applied to southern Kansas, a region with increased seismic activity related to hydrocarbon-production and waste water injection. The deep learning seismicity and focal mechanism catalogs show immensely more seismic activity than standard processing.