Laboratory earthquake forecasting: A machine learning competition.

Earthquake prediction, the long-sought holy grail of earthquake science, continues to confound Earth scientists. Could we make advances by crowdsourcing, drawing from the vast knowledge and creativity of the machine learning (ML) community? We used Google's ML competition platform, Kaggle, to engage the worldwide ML community with a competition to develop and improve data analysis approaches on a forecasting problem that uses laboratory earthquake data. The competitors were tasked with predicting the time remaining before the next earthquake of successive laboratory quake events, based on only a small portion of the laboratory seismic data. The more than 4,500 participating teams created and shared more than 400 computer programs in openly accessible notebooks. Complementing the now well-known features of seismic data that map to fault criticality in the laboratory, the winning teams employed unexpected strategies based on rescaling failure times as a fraction of the seismic cycle and comparing input distribution of training and testing data. In addition to yielding scientific insights into fault processes in the laboratory and their relation with the evolution of the statistical properties of the associated seismic data, the competition serves as a pedagogical tool for teaching ML in geophysics. The approach may provide a model for other competitions in geosciences or other domains of study to help engage the ML community on problems of significance.

Dynamics of geologic CO2 storage and plume motion revealed by seismic coda waves.

Quantifying the dynamics of sequestered CO2 plumes is critical for safe long-term storage, providing guidance on plume extent, and detecting stratigraphic seal failure. However, existing seismic monitoring methods based on wave reflection or transmission probe a limited rock volume and their sensitivity decreases as CO2 saturation increases, decreasing their utility in quantitative plume mass estimation. Here we show that seismic scattering coda waves, acquired during continuous borehole monitoring, are able to illuminate details of the CO2 plume during a 74-h CO2 injection experiment at the Frio-II well Dayton, TX. Our study reveals a continuous velocity reduction during the dynamic injection of CO2, a result that augments and dramatically improves upon prior analyses based on P-wave arrival times. We show that velocity reduction is nonlinearly correlated with the injected cumulative CO2 mass and attribute this correlation to the fact that coda waves repeatedly sample the heterogeneous distribution of cumulative CO2 in the reservoir zone. Lastly, because our approach does not depend on P-wave arrival times or require well-constrained wave reflections it can be used with many source-receiver geometries including those external to the reservoir, which reduces the risk introduced by in-reservoir monitoring wells. Our results provide an approach for quantitative CO2 monitoring and plume evolution that increases safety and long-term planning for CO2 injection and storage.

Fault zone fabric and fault weakness.

Geological and geophysical evidence suggests that some crustal faults are weak compared to laboratory measurements of frictional strength. Explanations for fault weakness include the presence of weak minerals, high fluid pressures within the fault core and dynamic processes such as normal stress reduction, acoustic fluidization or extreme weakening at high slip velocity. Dynamic weakening mechanisms can explain some observations; however, creep and aseismic slip are thought to occur on weak faults, and quasi-static weakening mechanisms are required to initiate frictional slip on mis-oriented faults, at high angles to the tectonic stress field. Moreover, the maintenance of high fluid pressures requires specialized conditions and weak mineral phases are not present in sufficient abundance to satisfy weak fault models, so weak faults remain largely unexplained. Here we provide laboratory evidence for a brittle, frictional weakening mechanism based on common fault zone fabrics. We report on the frictional strength of intact fault rocks sheared in their in situ geometry. Samples with well-developed foliation are extremely weak compared to their powdered equivalents. Micro- and nano-structural studies show that frictional sliding occurs along very fine-grained foliations composed of phyllosilicates (talc and smectite). When the same rocks are powdered, frictional strength is high, consistent with cataclastic processes. Our data show that fault weakness can occur in cases where weak mineral phases constitute only a small percentage of the total fault rock and that low friction results from slip on a network of weak phyllosilicate-rich surfaces that define the rock fabric. The widespread documentation of foliated fault rocks along mature faults in different tectonic settings and from many different protoliths suggests that this mechanism could be a viable explanation for fault weakening in the brittle crust.

Effects of acoustic waves on stick-slip in granular media and implications for earthquakes.

It remains unknown how the small strains induced by seismic waves can trigger earthquakes at large distances, in some cases thousands of kilometres from the triggering earthquake, with failure often occurring long after the waves have passed. Earthquake nucleation is usually observed to take place at depths of 10-20 km, and so static overburden should be large enough to inhibit triggering by seismic-wave stress perturbations. To understand the physics of dynamic triggering better, as well as the influence of dynamic stressing on earthquake recurrence, we have conducted laboratory studies of stick-slip in granular media with and without applied acoustic vibration. Glass beads were used to simulate granular fault zone material, sheared under constant normal stress, and subject to transient or continuous perturbation by acoustic waves. Here we show that small-magnitude failure events, corresponding to triggered aftershocks, occur when applied sound-wave amplitudes exceed several microstrain. These events are frequently delayed or occur as part of a cascade of small events. Vibrations also cause large slip events to be disrupted in time relative to those without wave perturbation. The effects are observed for many large-event cycles after vibrations cease, indicating a strain memory in the granular material. Dynamic stressing of tectonic faults may play a similar role in determining the complexity of earthquake recurrence.

Frictional instabilities in clay illuminate the origin of slow earthquakes.

The shallowest regions of subduction megathrusts mainly deform aseismically, but they can sporadically host slow-slip events (SSEs) and tsunami earthquakes, thus representing a severe hazard. However, the mechanisms behind these remain enigmatic because the frictional properties of shallow subduction zones, usually rich in clay, do not allow earthquake slip according to standard friction theory. We present experimental data showing that clay-rich faults with bulk rate-strengthening behavior and null healing rate, typically associated with aseismic creep, can contemporaneously creep and nucleate SSE. Our experiments document slow ruptures occurring within thin shear zones, driven by structural and stress heterogeneities of the experimental faults. We propose that bulk rate-strengthening frictional behavior promotes long-term aseismic creep, whereas localized frictional shear allows slow rupture nucleation and quasi-dynamic propagation typical of rate-weakening behavior. Our results provide additional understanding of fault friction and illustrate the complex behavior of clay-rich faults, providing an alternative paradigm for interpretation of the spectrum of fault slip including SSEs and tsunami earthquakes.

Crustal permeability generated through microearthquakes is constrained by seismic moment.

We link changes in crustal permeability to informative features of microearthquakes (MEQs) using two field hydraulic stimulation experiments where both MEQs and permeability evolution are recorded simultaneously. The Bidirectional Long Short-Term Memory (Bi-LSTM) model effectively predicts permeability evolution and ultimate permeability increase. Our findings confirm the form of key features linking the MEQs to permeability, offering mechanistically consistent interpretations of this association. Transfer learning correctly predicts permeability evolution of one experiment from a model trained on an alternate dataset and locale, which further reinforces the innate interdependency of permeability-to-seismicity. Models representing permeability evolution on reactivated fractures in both shear and tension suggest scaling relationships in which changes in permeability ( Δ k ) are linearly related to the seismic moment ( M ) of individual MEQs as Δ k ∝ M . This scaling relation rationalizes our observation of the permeability-to-seismicity linkage, contributes to its predictive robustness and accentuates its potential in characterizing crustal permeability evolution using MEQs.

Earthquake energy dissipation in a fracture mechanics framework.

Earthquakes are rupture-like processes that propagate along tectonic faults and cause seismic waves. The propagation speed and final area of the rupture, which determine an earthquake's potential impact, are directly related to the nature and quantity of the energy dissipation involved in the rupture process. Here, we present the challenges associated with defining and measuring the energy dissipation in laboratory and natural earthquakes across many scales. We discuss the importance and implications of distinguishing between energy dissipation that occurs close to and far behind the rupture tip, and we identify open scientific questions related to a consistent modeling framework for earthquake physics that extends beyond classical Linear Elastic Fracture Mechanics.

Foreshock properties illuminate nucleation processes of slow and fast laboratory earthquakes.

Understanding the connection between seismic activity and the earthquake nucleation process is a fundamental goal in earthquake seismology with important implications for earthquake early warning systems and forecasting. We use high-resolution acoustic emission (AE) waveform measurements from laboratory stick-slip experiments that span a spectrum of slow to fast slip rates to probe spatiotemporal properties of laboratory foreshocks and nucleation processes. We measure waveform similarity and pairwise differential travel-times (DTT) between AEs throughout the seismic cycle. AEs broadcasted prior to slow labquakes have small DTT and high waveform similarity relative to fast labquakes. We show that during slow stick-slip, the fault never fully locks, and waveform similarity and pairwise differential travel times do not evolve throughout the seismic cycle. In contrast, fast laboratory earthquakes are preceded by a rapid increase in waveform similarity late in the seismic cycle and a reduction in differential travel times, indicating that AEs begin to coalesce as the fault slip velocity increases leading up to failure. These observations point to key differences in the nucleation process of slow and fast labquakes and suggest that the spatiotemporal evolution of laboratory foreshocks is linked to fault slip velocity.

Using a physics-informed neural network and fault zone acoustic monitoring to predict lab earthquakes.

Predicting failure in solids has broad applications including earthquake prediction which remains an unattainable goal. However, recent machine learning work shows that laboratory earthquakes can be predicted using micro-failure events and temporal evolution of fault zone elastic properties. Remarkably, these results come from purely data-driven models trained with large datasets. Such data are equivalent to centuries of fault motion rendering application to tectonic faulting unclear. In addition, the underlying physics of such predictions is poorly understood. Here, we address scalability using a novel Physics-Informed Neural Network (PINN). Our model encodes fault physics in the deep learning loss function using time-lapse ultrasonic data. PINN models outperform data-driven models and significantly improve transfer learning for small training datasets and conditions outside those used in training. Our work suggests that PINN offers a promising path for machine learning-based failure prediction and, ultimately for improving our understanding of earthquake physics and prediction.

The High-Frequency Signature of Slow and Fast Laboratory Earthquakes.

Tectonic faults fail through a spectrum of slip modes, ranging from slow aseismic creep to rapid slip during earthquakes. Understanding the seismic radiation emitted during these slip modes is key for advancing earthquake science and earthquake hazard assessment. In this work, we use laboratory friction experiments instrumented with ultrasonic sensors to document the seismic radiation properties of slow and fast laboratory earthquakes. Stick-slip experiments were conducted at a constant loading rate of 8 µm/s and the normal stress was systematically increased from 7 to 15 MPa. We produced a full spectrum of slip modes by modulating the loading stiffness in tandem with the fault zone normal stress. Acoustic emission data were recorded continuously at 5 MHz. We demonstrate that the full continuum of slip modes radiate measurable high-frequency energy between 100 and 500 kHz, including the slowest events that have peak fault slip rates <100 µm/s. The peak amplitude of the high-frequency time-domain signals scales systematically with fault slip velocity. Stable sliding experiments further support the connection between fault slip rate and high-frequency radiation. Experiments demonstrate that the origin of the high-frequency energy is fundamentally linked to changes in fault slip rate, shear strain, and breaking of contact junctions within the fault gouge. Our results suggest that having measurements close to the fault zone may be key for documenting seismic radiation properties and fully understanding the connection between different slip modes.

Creep fronts and complexity in laboratory earthquake sequences illuminate delayed earthquake triggering.

Earthquakes occur in clusters or sequences that arise from complex triggering mechanisms, but direct measurement of the slow subsurface slip responsible for delayed triggering is rarely possible. We investigate the origins of complexity and its relationship to heterogeneity using an experimental fault with two dominant seismic asperities. The fault is composed of quartz powder, a material common to natural faults, sandwiched between 760 mm long polymer blocks that deform the way 10 meters of rock would behave. We observe periodic repeating earthquakes that transition into aperiodic and complex sequences of fast and slow events. Neighboring earthquakes communicate via migrating slow slip, which resembles creep fronts observed in numerical simulations and on tectonic faults. Utilizing both local stress measurements and numerical simulations, we observe that the speed and strength of creep fronts are highly sensitive to fault stress levels left behind by previous earthquakes, and may serve as on-fault stress meters.

Frequency-Magnitude Statistics of Laboratory Foreshocks Vary With Shear Velocity, Fault Slip Rate, and Shear Stress.

Understanding the temporal evolution of foreshocks and their relation to earthquake nucleation is important for earthquake early warning systems, earthquake hazard assessment, and earthquake physics. Laboratory experiments on intact rock and rough fractures have demonstrated that the number and size of acoustic emission (AE) events increase and that the Gutenberg-Richter b-value decreases prior to coseismic failure. However, for lab fault zones of finite width, where shear occurs within gouge, the physical processes that dictate temporal variations in frequency-magnitude (F/M) statistics of lab foreshocks are unclear. Here, we report on a series of laboratory experiments to illuminate the physical processes that govern temporal variations in b-value and AE size. We record AE data continuously for hundreds of lab seismic cycles and report F/M statistics. Our foreshock catalogs include cases where F/M data are not exponentially distributed, but we retain the concept of b-value for comparison with other works. We find that b-value decreases as the fault approaches failure, consistent with previous works. We also find that b-value scales inversely with shear velocity and fault slip rate, suggesting that fault slip acceleration during earthquake nucleation could impact foreshock F/M statistics. We propose that fault zone dilation and grain mobilization have a strong influence on foreshock magnitude. Fault dilation at higher shearing rates increases porosity and results in larger foreshocks and smaller b-values. Our observations suggest that lab earthquakes are preceded by a preparatory nucleation phase with systematic variations in AE and fault zone properties.

Machine Learning Predicts the Timing and Shear Stress Evolution of Lab Earthquakes Using Active Seismic Monitoring of Fault Zone Processes.

Machine learning (ML) techniques have become increasingly important in seismology and earthquake science. Lab-based studies have used acoustic emission data to predict time-to-failure and stress state, and in a few cases, the same approach has been used for field data. However, the underlying physical mechanisms that allow lab earthquake prediction and seismic forecasting remain poorly resolved. Here, we address this knowledge gap by coupling active-source seismic data, which probe asperity-scale processes, with ML methods. We show that elastic waves passing through the lab fault zone contain information that can predict the full spectrum of labquakes from slow slip instabilities to highly aperiodic events. The ML methods utilize systematic changes in P-wave amplitude and velocity to accurately predict the timing and shear stress during labquakes. The ML predictions improve in accuracy closer to fault failure, demonstrating that the predictive power of the ultrasonic signals improves as the fault approaches failure. Our results demonstrate that the relationship between the ultrasonic parameters and fault slip rate, and in turn, the systematically evolving real area of contact and asperity stiffness allow the gradient boosting algorithm to "learn" about the state of the fault and its proximity to failure. Broadly, our results demonstrate the utility of physics-informed ML in forecasting the imminence of fault slip at the laboratory scale, which may have important implications for earthquake mechanics in nature.

The Role of Fabric in Frictional Properties of Phyllosilicate-Rich Tectonic Faults.

Many rock deformation experiments used to characterize the frictional properties of tectonic faults are performed on powdered fault rocks or on bare rock surfaces. These experiments have been fundamental to document the frictional properties of granular mineral phases and provide evidence for crustal faults characterized by high friction. However, they cannot entirely capture the frictional properties of faults rich in phyllosilicates. Numerous studies of natural faults have documented fluid-assisted reaction softening promoting the replacement of strong minerals with phyllosilicates that are distributed into continuous foliations. To study how these foliated fabrics influence the frictional properties of faults we have: 1) collected foliated phyllosilicate-rich rocks from natural faults; 2) cut the fault rock samples to obtain solid wafers 0.8-1.2 cm thick and 5 cm x 5 cm in area with the foliation parallel to the 5x5cm face of the wafer; 3) performed friction tests on both solid wafers sheared in their in situ geometry and powders, obtained by crushing and sieving and therefore disrupting the foliation of the same samples; 4) recovered the samples for microstructural studies from the post experiment rock samples; and 5) performed microstructural analyses via optical microscopy, scanning and transmission electron microscopy. Mechanical data show that the solid samples with well-developed foliation show significantly lower friction in comparison to their powdered equivalents. Micro- and nano-structural studies demonstrate that low friction results from sliding along the foliation surfaces composed of phyllosilicates. When the same rocks are powdered, frictional strength is high, because sliding is accommodated by fracturing, grain rotation, translation and associated dilation. Friction tests indicate that foliated fault rocks may have low friction even when phyllosilicates constitute only a small percentage of the total rock volume, implying that a significant number of crustal faults are weak.

Acoustic Energy Release During the Laboratory Seismic Cycle: Insights on Laboratory Earthquake Precursors and Prediction.

Machine learning can predict the timing and magnitude of laboratory earthquakes using statistics of acoustic emissions. The evolution of acoustic energy is critical for lab earthquake prediction; however, the connections between acoustic energy and fault zone processes leading to failure are poorly understood. Here, we document in detail the temporal evolution of acoustic energy during the laboratory seismic cycle. We report on friction experiments for a range of shearing velocities, normal stresses, and granular particle sizes. Acoustic emission data are recorded continuously throughout shear using broadband piezo-ceramic sensors. The coseismic acoustic energy release scales directly with stress drop and is consistent with concepts of frictional contact mechanics and time-dependent fault healing. Experiments conducted with larger grains (10.5 µm) show that the temporal evolution of acoustic energy scales directly with fault slip rate. In particular, the acoustic energy is low when the fault is locked and increases to a maximum during coseismic failure. Data from traditional slide-hold-slide friction tests confirm that acoustic energy release is closely linked to fault slip rate. Furthermore, variations in the true contact area of fault zone particles play a key role in the generation of acoustic energy. Our data show that acoustic radiation is related primarily to breaking/sliding of frictional contact junctions, which suggests that machine learning-based laboratory earthquake prediction derives from frictional weakening processes that begin very early in the seismic cycle and well before macroscopic failure.

Three-dimensional discrete element modeling of triggered slip in sheared granular media.

This paper reports results of a three-dimensional discrete element method modeling investigation of the role of boundary vibration in perturbing stick-slip dynamics in a sheared granular layer. The focus is on the influence of vibration within a range of amplitudes and on the fact that above a threshold early slip will be induced. We study the effects of triggering beyond the vibration interval and their origins. A series of perturbed simulations are performed for 30 large slip events selected from different reference runs, in the absence of vibration. For each of the perturbed simulations, vibration is applied either about the middle of the stick phase or slightly before the onset of a large expected slip event. For both cases, a suppression of energy release is on average observed in the perturbed simulations, within the short term following the vibration application. For cases where vibration is applied in the middle of the stick phase, a significant clock advance of the large slip event occurs. In the long term after vibration, there is a recovery period with higher-energy release and increased activity in the perturbed simulations, which compensates for the temporary suppression observed within the short term.