Search

Estimating an earthquake's magnitude and location may not be necessary to predict shaking in real time; instead, wavefield‐based approaches predict shaking with few assumptions about the seismic source. Here, we introduce GRAph Prediction of Earthquake Shaking (GRAPES), a deep learning model trained to characterize and propagate earthquake shaking across a seismic network. We show that GRAPES’ internal activations, which we call “seismic vectors”, correspond to the arrival of distinct seismic phases. GRAPES builds upon recent deep learning models applied to earthquake early warning by allowing for continuous ground motion prediction with seismic networks of all sizes. While trained on earthquakes recorded in Japan, we show that GRAPES, without modification, outperforms the ShakeAlert earthquake early warning system on the 2019 M7.1 Ridgecrest, CA earthquake. Have you ever heard something through the grapevine? It often takes you by surprise to hear a message from someone other than the original source. You might have felt an earthquake in a similar way: experiencing shaking (the message) at your location rather than movement along a fault (the source). We apply grapevine‐style communication to earthquake early warning (EEW). The goal of EEW is to warn people to prepare for earthquake shaking before damaging seismic waves arrive at their location. We build on recent work that used deep learning and large earthquake data sets to predict earthquake shaking. We developed a deep learning algorithm called GRAPES which predicts shaking in a manner similar to a game of seismic telephone: when a seismic sensor detects shaking, it sends a message to its neighboring sensors, warning them to expect shaking soon. These sensors then pass on the message to their more distant neighbors along the grapevine. We show that the messages GRAPES learned to send between sensors are like seismic status updates: “I'm seeing this type of seismic wave right now”. We applied GRAPES to the 2019 M7.1 Ridgecrest, CA earthquake and it predicted shaking accurately and quickly. A deep learning network trained to predict ground motion learned an internal representation of the seismic wavefieldIndividual neurons within the network activate with the arrival of P waves, S waves, surface waves, coda waves, and ambient noiseWhile trained on earthquakes in Japan, the model generalizes well to predicting ground motions for the 2019 Ridgecrest, CA earthquake A deep learning network trained to predict ground motion learned an internal representation of the seismic wavefield Individual neurons within the network activate with the arrival of P waves, S waves, surface waves, coda waves, and ambient noise While trained on earthquakes in Japan, the model generalizes well to predicting ground motions for the 2019 Ridgecrest, CA earthquake

During its ascent up Mount Sharp, the Mars Science Laboratory Curiosity rover traversed the Bagnold Dune Field. We model sand modal mineralogy and grain size at four locations near the rover traverse, using orbital shortwave infrared single‐scattering albedo spectra and a Markov chain Monte Carlo implementation of Hapke's radiative transfer theory to fully constrain uncertainties and permitted solutions. These predictions, evaluated against in situ measurements at one site from the Curiosity rover, show that X‐ray diffraction‐measured mineralogy of the basaltic sands is within the 95% confidence interval of model predictions. However, predictions are relatively insensitive to grain size and are nonunique, especially when modeling the composition of minerals with solid solutions. We find an overall basaltic mineralogy and show subtle spatial variations in composition in and around the Bagnold Dunes, consistent with a mafic enrichment of sands with cumulative aeolian‐transport distance by sorting of olivine, pyroxene, and plagioclase grains. Furthermore, the large variations in Fe and Mg abundances (~20 wt %) at the Bagnold Dunes suggest that compositional variability may be enhanced by local mixing of well‐sorted sand with proximal sand sources. Our estimates demonstrate a method for orbital quantification of composition with rigorous uncertainty determination and provide key constraints for interpreting in situ measurements of compositional variability within Martian aeolian sandstones. Spectral variability in the dunes arises from spatially variable mineral composition and qualitatively correlates with local sand fluxMineral abundance estimates from orbit are within <13 wt % of in situ measurements from CheMin in average, although uncertainties are largeOrbital observations of dune field mineral variations may be explained by both aeolian sorting and mixing of multiple sand sources

Reflected light from planetary surfaces provides information, including mineral/ice compositions and grain sizes, by study of albedo and absorption features as a function of wavelength. However, deconvolving the compositional signal in spectra is complicated by the nonuniqueness of the inverse problem. Trade‐offs between mineral abundances and grain sizes in setting reflectance, instrument noise, and systematic errors in the forward model are potential sources of uncertainty, which are often unquantified. Here we adopt a Bayesian implementation of the Hapke model to determine sets of acceptable‐fit mineral assemblages, as opposed to single best fit solutions. We quantify errors and uncertainties in mineral abundances and grain sizes that arise from instrument noise, compositional end members, optical constants, and systematic forward model errors for two suites of ternary mixtures (olivine‐enstatite‐anorthite and olivine‐nontronite‐basaltic glass) in a series of six experiments in the visible‐shortwave infrared (VSWIR) wavelength range. We show that grain sizes are generally poorly constrained from VSWIR spectroscopy. Abundance and grain size trade‐offs lead to typical abundance errors of ≤1 wt % (occasionally up to ~5 wt %), while ~3% noise in the data increases errors by up to ~2 wt %. Systematic errors further increase inaccuracies by a factor of 4. Finally, phases with low spectral contrast or inaccurate optical constants can further increase errors. Overall, typical errors in abundance are <10%, but sometimes significantly increase for specific mixtures, prone to abundance/grain‐size trade‐offs that lead to high unmixing uncertainties. These results highlight the need for probabilistic approaches to remote determination of planetary surface composition. We design and implement a Markov Chain Monte Carlo approach to spectral unmixing to test accuracies of the Hapke radiative transfer modelErrors/uncertainties in abundances/grain sizes are quantified for two sets of laboratory ternary mixtures, with and without noiseLarge uncertainties due to nonunique fits highlight the usefulness of probabilistic approaches to spectral unmixing of remote sensing data

The subduction zone in northern Chile is a well‐identified seismic gap that last ruptured in 1877. On 1 April 2014, this region was struck by a large earthquake following a two week long series of foreshocks. This study combines a wide range of observations, including geodetic, tsunami, and seismic data, to produce a reliable kinematic slip model of the Mw=8.1 main shock and a static slip model of the Mw=7.7 aftershock. We use a novel Bayesian modeling approach that accounts for uncertainty in the Green's functions, both static and dynamic, while avoiding nonphysical regularization. The results reveal a sharp slip zone, more compact than previously thought, located downdip of the foreshock sequence and updip of high‐frequency sources inferred by back‐projection analysis. Both the main shock and the Mw=7.7 aftershock did not rupture to the trench and left most of the seismic gap unbroken, leaving the possibility of a future large earthquake in the region. A Bayesian ensemble of kinematic slip models is constructed using geodetic, tsunami and seismic dataThe earthquake involved a sharp slip zone, more compact than previously thoughtThe main asperity is located downdip of the foreshock activity and updip of high‐frequency sources

We present an inversion strategy capable of using real‐time high‐rate GPS data to simultaneously solve for a distributed slip model and fault geometry in real time as a rupture unfolds. We employ Bayesian inference to find the optimal fault geometry and the distribution of possible slip models for that geometry using a simple analytical solution. By adopting an analytical Bayesian approach, we can solve this complex inversion problem (including calculating the uncertainties on our results) in real time. Furthermore, since the joint inversion for distributed slip and fault geometry can be computed in real time, the time required to obtain a source model of the earthquake does not depend on the computational cost. Instead, the time required is controlled by the duration of the rupture and the time required for information to propagate from the source to the receivers. We apply our modeling approach, called Bayesian Evidence‐based Fault Orientation and Real‐time Earthquake Slip, to the 2011 Tohoku‐oki earthquake, 2003 Tokachi‐oki earthquake, and a simulated Hayward fault earthquake. In all three cases, the inversion recovers the magnitude, spatial distribution of slip, and fault geometry in real time. Since our inversion relies on static offsets estimated from real‐time high‐rate GPS data, we also present performance tests of various approaches to estimating quasi‐static offsets in real time. We find that the raw high‐rate time series are the best data to use for determining the moment magnitude of the event, but slightly smoothing the raw time series helps stabilize the inversion for fault geometry. Inversions for distributed slip and fault geometry are computable in real timeHigh‐rate GPS data can estimate real‐time accumulated magnitude releaseComputed fault geometry is accurate once rupture has significant finite extent

The PLUM (Propagation of Local Undamped Motion) earthquake early warning (EEW) algorithm differs from typical source‐based EEW algorithms as it predicts shaking directly from observed shaking without first deriving earthquake source information (e.g., magnitude and epicenter). Here, we determine optimal PLUM event detection thresholds for U.S. West Coast earthquakes using two data sets: 558 M3.5+ earthquakes (California, Oregon, Washington; 2012–2017) and the ShakeAlert test suite of historic and problematic signals (1999–2015). PLUM computes Modified Mercalli Intensity (IMMI) using velocity and acceleration data, leveraging co‐located sensors to avoid problematic signals. An event detection is issued when the observed IMMIexceeds a given threshold(s). We find a two‐station detection method using IMMItrigger thresholds of 4.0 and 3.0 for the first and second stations, respectively, is optimal for detecting M4.5+ earthquakes. PLUM detected 79 events in the 2012–2017 data set, reporting (not including telemetry or alert dissemination) detection times on par, and sometimes faster than current EEW methods (mean 8 s; median 6 s). As expected, detection times were slower for the older 1999–2015 earthquakes (N= 21; mean 11 s; median 6 s) when station coverage was sparser. Of the 31 PLUM detected M5+ events (10 2012–2017; 21 1999–2015), theoretically 20 (∼65%) could provide timely warnings. PLUM issued no false detections and avoided issuing detections for all calibration/anomalous signals, regional and teleseismic events. We conclude PLUM can successfully identify IMMI4+ shaking from local earthquakes and could complement and enhance EEW in the U.S. Earthquake early warning (EEW) detection schemes require (1) ample seismic information to identify where large ground motions are underway; (2) determining if these ground motions are significant enough to issue a detection; and (3) detecting large ground motions in a timely fashion. Some EEW methods estimate earthquake source parameters like magnitude and location and then input those parameters into a ground‐motion prediction equation, while other methods use observations of the ground motions to directly forecast shaking. We explore the latter approach using the PLUM (Propagation of Local Undamped Motion) method to detect earthquakes that produce shaking above a target value. In this work, we test PLUM's ability to detect earthquakes using two data sets: 558 earthquakes magnitude 3.5 and above from California, Oregon, and Washington (2012–2017) and a test suite of historic and problematic signals (1999–2015) curated by ShakeAlert. We find a two‐station detection method is preferred over a one‐station method as two‐stations can greatly minimize false detections. The PLUM method is also 100% successful at avoiding non‐earthquake anomalous signals and can successfully differentiate ground shaking from local and distant earthquakes. We conclude that PLUM may be a promising candidate for integration into the U.S. EEW system. Propagation of Local Undamped Motion (PLUM) detects offshore events that produce at least moderate onshore ground motions, including an event problematic for other algorithmsPLUM's detection times are as timely, and sometimes faster, than other earthquake early warning detection timesPLUM correctly avoids erroneous detections for all teleseismic, calibration pulses, and anomalous signals in the test suite Propagation of Local Undamped Motion (PLUM) detects offshore events that produce at least moderate onshore ground motions, including an event problematic for other algorithms PLUM's detection times are as timely, and sometimes faster, than other earthquake early warning detection times PLUM correctly avoids erroneous detections for all teleseismic, calibration pulses, and anomalous signals in the test suite