Fabio Corbi

Earthquakes occurrence is currently unpredictable in terms of exact time, location, and size. Fault dynamics exhibits many slip behaviors though, ranging from regular earthquakes to slow slip events (SSEs). Like regular earthquakes, SSEs are the result of frictional instabilities and exhibit cyclic but non-periodic behavior, with complex spatio-temporal patterns. We run a prospective forecasting experiment to forecast SSEs along the Cascadia subduction zone within a real-time framework. Forecasts based on naturally occurring analogs of seismo-geodetic observables outperform standard baselines (persistence, expectation, and autoregressive models) and a renewal model. This indicates the existence of structured, state-dependent dynamics and is consistent with a hierarchic multi-scale nonlinear dynamical system rather than a self-organized critical one. These results show the feasibility of real-time probabilistic SSEs forecasting.

Understanding the factors that influence the spatial and temporal rupture history of large subduction earthquakes is essential for seismic hazard assessment. In this study, we focus on one candidate: the interaction between neighboring asperities of different sizes and normal loads. We use scaled seismotectonic models to generate multi‐cycle records of analog earthquakes. We quantify the coefficient of variation CoV of the recurrence time Rt at different points along the strike of the model. The lateral continuity of CoV measurements reveals agreement between analog earthquake periodicity and frictional segmentation. We observe quasi‐periodic recurrence as well as various types of supercycles, including rupture cascades, clustered ruptures, and multi‐asperity events, similar to observed behavior in subduction zones. Under certain experimental conditions, the behavior of our models shows similarity to observations in south‐central Chile, where variations in normal stress, potentially detectable through gravity anomalies, can provide a basis for seismic segmentation and hazard assessment.

Geodetic observations along convergent margins have achieved unprecedented resolution in detailing deformation associated with earthquake cycles. A comprehensive understanding of how to best interpret these data for forecasting remains crucial. Here, we combined analogue seismo‐tectonic models of a megathrust and Explainable Artificial Intelligence (XAI) to characterize the link between earthquakes and deformation. We utilized deformation features to train convolutional neural networks (CNN) that forecast the time left before a laboratory earthquake. We then used Integrated Gradients (IG), an XAI technique, to identify areas and features contributing to model forecasts. CNNs perform better compared to decision trees utilizing sparse point‐wise features highlighting the importance of spatial patterns. IG reveals the significance of trench‐perpendicular deformation downdip the rupturing asperity, trench‐parallel deformation inland, and local deformation curl, in forecasting rupture timing. These emphasize the need for dense networks to monitor deformation and suggest patterns that may signify rupture imminence in convergent margins.

Modeling the seismic cycle requires multiple assumptions and parameters. Providing a quantitative assessment of the model behavior is pivotal for determining the degree of similarity between different scales and modeling strategies and for exploring dependencies with respect to selected parameters. Here we compare stick-slip ruptures nucleating spontaneously in scaled seismotectonic models (i.e., laboratory experiments capturing the first-order physics of the seismic cycle of subduction megathrusts) with slow earthquakes in nature. We rely on two non-dimensional parameters, namely the Ruina number (Ru) and system dimension (D) to quantify model behavior. Ru is proportional to the ratio of the asperity size to the critical nucleation size. Within the rate- and state friction framework, for velocity weakening asperities Ru controls the behavior of the system, which can be either periodic or not, and it can exhibit both slow and fast ruptures. D measures how complicated the system evolution is. D reveals how many variables are required to describe the seismic cycle because it tells us the minimum dimension needed to embed the observed dynamics. By coupling the Simulated Annealing algorithm and quasi-dynamic numerical simulations, we retrieve rate and state friction parameters characterizing single asperity models with different lateral extent of the velocity weakening patch. Similarly to slow earthquakes, we found optimal rate and state parameters indicative of low (< 4) Ru. We also found a direct proportionality between Ru and the lateral extent of the asperity. Next, we implement tools from non-linear time-series analysis and Extreme Value Theory to compute D from models of different sizes, materials, deformation rates and frictional configurations (single or twin asperities along strike). Our analysis supports the existence of a low dimensional attractor (D

We investigate the physics of laboratory earthquakes in scaled seismotectonic models of megathrust seismic cycles. We study models of different sizes, materials, deformation rates, and frictional configurations. We use nonlinear time-series analysis tools to characterize the dynamics of the models. Observations are described, on average, by a low-dimension (<5), similar to slow slip episodes in nature and friction experiments performed with quartz powder. Results seem insensitive to the along-strike frictional segmentation of the megathrust. Using displacement as an input variable, the instantaneous dimension and the instantaneous extremal index vary through the seismic cycles. We notice the highest values of the instantaneous dimension associated with slip phases. Under specific circumstances, clear drops of the instantaneous extremal index can serve as an early indicator of slip episodes. Prediction horizons in the order of slip duration mirror similar predictability as for slow slip episodes in nature. We conclude that seismotectonic models are effective tools to study frictional physics despite their different spatio-temporal scales.

Accurate assessment of the rate and state friction parameters of rocks is essential for producing realistic earthquake rupture scenarios and, in turn, for seismic hazard analysis. Those parameters can be directly measured on samples, or indirectly based on inversion of coseismic or postseismic slip evolution. However, both direct and indirect approaches require assumptions that might bias the results. Aiming to reduce the potential sources of bias, we take advantage of a downscaled analog model reproducing megathrust earthquakes. We couple the simulated annealing algorithm with quasi‐dynamic numerical models to retrieve rate and state parameters reproducing the recurrence time, rupture duration and slip of the analog model, in the ensemble. Then, we focus on how the asperity size and the neighboring segments' properties control the seismic cycle characteristics and the corresponding variability of rate and state parameters. We identify a tradeoff between ( a – b ) of the asperity and ( a – b ) of neighboring creeping segments, with multiple parameter combinations that allow mimicking the analog model behavior. Tuning of rate and state parameters is required to fit laboratory experiments with different asperity lengths. Poorly constrained frictional properties of neighboring segments are responsible for uncertainties of ( a – b ) of the asperity in the order of per mille. Roughly one order of magnitude larger uncertainties derive from asperity size. Those results provide a glimpse of the variability that rate and state friction estimates might have when used as a constraint to model fault slip behavior in nature.

Subducting seamounts are recognized as one of the key features influencing megathrust earthquakes. However, whether they trigger or arrest ruptures remains debated. Here, we use analog models to study the influence of a single seamount on megathrust earthquakes, separating the effect of topography from that of friction. Four different model configurations have been developed (i.e., flat interface, high and low friction seamount, low friction patch). In our models, the seamount reduces recurrence time, interseismic coupling, and fault strength, suggesting that it acts as a barrier: 80% of the ruptures concentrate in flat regions that surround the seamount and only smaller magnitude earthquakes nucleate above it. The low‐friction zone, which mimics the fluid accumulation or the establishment of fracture systems in natural cases, seems to be the most efficient in arresting rupture propagation in our experimental setting.

Accurate assessment of rate and state friction parameters is essential for producing realistic rupture scenarios and, in turn, for seismic hazard analysis. Those parameters can be directly measured in the laboratory, with experimental apparati that reproduce fault conditions in nature. Alternatively, indirect estimates (i.e., inversion) of rate and state parameters are based on postseismic slip evolution studies and numerical modeling. Both direct and indirect approaches require a series of assumptions that might bias the results.Here we take advantage of a downscaled analog model reproducing experimentally megathrust earthquakes. The analog model shares many characteristics of real subduction zones, although being intentionally oversimplified with respect to nature. This allows reducing the number of potential sources of bias (e.g., fault geometry and asperity size). We perform five analog models with a single, rectangular asperity of different lengths embedded in a nearly velocity neutral matrix. We focus on two different physical conditions, namely the along-strike asperity length and the asperity to neighboring segments length ratio, and study systematically how they tune the model seismic behavior. Then, by coupling quasi-dynamic numerical models with the simulated annealing algorithm, we retrieve rate and state parameters that allow reproducing both the recurrence time, rupture duration and slip amplitude of the analog model, in ensemble. We identify a tradeoff between (a-b) of the asperity and (a-b) of neighboring creeping segments, with multiple combinations that allow mimicking the analog model behavior and variability. We also identify a negative correlation between (a-b) of the asperity and asperity size, with Dc remaining relatively constant within the investigated asperity size range. When estimating (a-b), poorly constrained properties of neighboring segments are responsible for uncertainties in the order of per mille. Roughly one order of magnitude larger uncertainties derive from asperity size. Those results provide a first order assessment of the variability that rate and state friction estimates retrieved for nature conditions might have when used as constraint to model fault slip behavior.

Recent observations from Global Navigation Satellite Systems (GNSS) displacement time-series have demonstrated that the motion of tectonic plates is ubiquitously not steady-state. GNSS displacement time-series can be described as the sum of expected motions such as long term tectonic interseismic motion or annually and semi-annually seasonal oscillations, and unexpected “transients”, such as slow-slip events or volcanic deformations. Although the number and quality of globally available GNSS time-series have increased dramatically in recent years, detecting and modeling these signals remains challenging, because of their elusive nature, masked by the presence of noise. The most popular approach for filtering such noise is the application of common-mode-filters (CMFs) that use weighted averages of residuals after fitting individual time-series with trajectory models. CMFs exploit the spatial coherence of higher frequency noise in both Precise Point Positioning and network double difference solutions to systematically reduce the noise of each time-series. However, the application of CMFs is limited in the case of sparsely distributed GNSS stations. Additionally, CMFs can potentially map local transients into noise when the original trajectory model fits are made suboptimally.Here we propose an alternative method to CMFs by exploiting Deep Learning (DL) techniques. Our supervised learning regression method aims to remove the noise present in each GNSS time-series regardless of its geographical location on a station-by-station basis. Our dataset consists of nine thousand time-series: all GNSS time-series available on the Nevada GNSS repository with at least 4 years of contiguous data. Our approach is defined by two subroutines. Firstly, the Greedy Automatic Signal Decomposition (GrAtSiD) algorithm is used to fit each time-series and leave behind a residual. GrAtSiD is a sequential greedy linear algorithm that decomposes the time-series into a minimum number of transient basis functions and some permanent functions. The residual time-series after signal decomposition is defined as noise, without qualifying its nature. Once this noise is identified, a DL model is trained to recognize this noise from the raw time-series, without the need for any trajectory modeling. The supervised learning regression model predicts what the residual to a trajectory model would be. Although GrAtSiD is very effective in isolating the high frequency noise of GNSS time-series, its fitting is dependent on the temporal length of input time-series. Additionally, GrAtSiD needs to set arbitrary thresholds which control the convergence of the inversion routine to avoid under- or over-fitting input data. In this context, our DL approach proposes a generalization of GrAtSiD solutions, by exploiting a weakly supervised training based on millions of examples - essentially, the model generalizes an optimal fit having seen many examples of GrAtSiD trajectory fits. This generalization allows the DL model to preserve apparent transient features of the time-series. A multitude of DL architectures are tested in both sequence-to-single and sequence-to-sequence regression framings. This exploration allows us to identify the best framing, architecture, and related hyperparameters for our method to be successful. With the best performing models, we demonstrate the effects of the DL high frequency noise removal and compare it to the CMF approach.

The interplay between tectonics and surface processes defines the evolution of mountain belts. However, correlating these processes through the evolution of natural orogens represents a scientific challenge. Analogue models can be used for analyzing and interpreting the effect of such interaction. We use nine analog models characterized by different combinations of imposed regional slope and rainfall rates to investigate how a small scale orogen evolves in response to tectonics and surface processes. We show how the combination of these parameters control the development of drainage networks and erosional processes. We quantify the different morphological expression of analogue landscapes in terms of a proposed Ae number that accounts for both observables and boundary conditions. We find few differences between analogue models and natural prototypes, in terms of parametrization of the detachment-limited stream power law. We observe a threshold in the development of channelization, modulated by a tradeoff between applied boundary conditions.