Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userPredicting Fault Slip via Transfer Learning
195
0
0.0
(
0
)
Ask ChatGPT about the research
No Arabic abstract
Data-driven machine-learning for predicting instantaneous and future fault-slip in laboratory experiments has recently progressed markedly due to large training data sets. In Earth however, earthquake interevent times range from 10s-100s of years and geophysical data typically exist for only a portion of an earthquake cycle. Sparse data presents a serious challenge to training machine learning models. Here we describe a transfer learning approach using numerical simulations to train a convolutional encoder-decoder that predicts fault-slip behavior in laboratory experiments. The model learns a mapping between acoustic emission histories and fault-slip from numerical simulations, and generalizes to produce accurate results using laboratory data. Notably slip-predictions markedly improve using the simulation-data trained-model and training the latent space using a portion of a single laboratory earthquake-cycle. The transfer learning results elucidate the potential of using models trained on numerical simulations and fine-tuned with small geophysical data sets for potential applications to faults in Earth.
rate research
Read More
Segment lengths along major strike-slip faults exhibit a size dependency related to the brittle crust thickness. These segments result in the formation of the localized P-shear deformation crossing and connecting the initial Riedels structures (i.e. en-echelon fault structures) which formed during the genesis stage of the fault zone. Mechanical models show that at all scales, the geometrical characteristics of the Riedels exhibit dependency on the thickness of the brittle layer. Combining the results of our mechanical discrete element model with several analogue experiments using sand, clay and gypsum, we have formulated a relationship between the orientation and spacing of Riedels and the thickness of the brittle layer. From this relationship, we derive that for a pure strike-slip mode, the maximum spacing between the Riedels are close to three times the thickness. For a transtensional mode, as the extensive component becomes predominant, the spacing distance at the surface become much smaller than the thickness. Applying this relationship to several well-characterized strike-slip faults on Earth, we show that the predicted brittle thickness is consistent with the seismogenic depth. Supposing the ubiquity of this phenomenon, we extent this relationship to characterize en-echelon structures observed on Mars, in the Memnonia region located West of Tharsis. Assuming that the outer ice shells of Ganymede, Enceladus and Europa, exhibit a brittle behavior, we suggest values of the corresponding apparent brittle thicknesses.
The surface roughness of a recently exhumed strikeslip fault plane has been measured by three independent 3D portable laser scanners. Digital elevation models of several fault surface areas, from 1 m2 to 600 m2, have been measured at a resolution ranging from 5 mm to 80 mm. Out of plane height fluctuations are described by non-Gaussian distribution with exponential long range tails. Statistical scaling analyses show that the striated fault surface exhibits self-affine scaling invariance with a small but significant directional morphological anisotropy that can be described by two scaling roughness exponents, H1 = 0.7 in the direction of slip and H2 = 0.8 perpendicular to the direction of slip.
Seismogenic plate boundaries are presumed to behave in a similar manner to a densely packed granular medium, where fault and blocks systems rapidly rearrange the distribution of forces within themselves, as particles do in slowly sheared granular systems. We use machine learning and show that statistical features of velocity signals from individual particles in a simulated sheared granular fault contain information regarding the instantaneous global state of intermittent frictional stick-slip dynamics. We demonstrate that combining features built from the signals of more particles can improve the accuracy of the global model, and discuss the physical basis behind decrease in error. We show that the statistical features such as median and higher moments of the signals that represent the particle displacement in the direction of shearing are among the best predictive features. Our work provides novel insights into the applications of machine learning in studying frictional processes that take place in geophysical systems.
Many fast-flowing glaciers and ice streams move over beds consisting of reworked sediments and erosional products, commonly referred to as till. The complex interplay between ice, meltwater, and till at the subglacial bed connects several fundamental problems in glaciology including the debate about rapid mass loss from the ice sheets, the formation and evolution of subglacial landforms, and the storage and transport of subglacial water. In-situ measurements have probed the subglacial bed, but provide surprisingly variable and seemingly inconsistent evidence of the depth where deformation occurs, even at a given field site. These observations suggest that subglacial beds are inherently dynamic. The goal of this paper is to advance our understanding of the physical processes that contribute to the dynamics of subglacial beds as reflected in existing observations of basal deformation. We build on recent advances in modeling dense, granular flows to derive a new numerical model for water-saturated till. Our model demonstrates that changes in the force balance or temporal variations in water pressure can shift slip away from the ice-bed interface and far into the bed, causing episodes of significantly enhanced till transport because the entire till layer above the deep slip interface becomes mobilized. We compare our model results against observations of basal deformation from both mountain glacier and ice-stream settings in the past and present. We also present an analytical solution to assess the variability of till transport for different glacial settings and hydraulic properties.
Achieving desirable receiver sampling in ocean bottom acquisition is often not possible because of cost considerations. Assuming adequate source sampling is available, which is achievable by virtue of reciprocity and the use of modern randomized (simultaneous-source) marine acquisition technology, we are in a position to train convolutional neural networks (CNNs) to bring the receiver sampling to the same spatial grid as the dense source sampling. To accomplish this task, we form training pairs consisting of densely sampled data and artificially subsampled data using a reciprocity argument and the assumption that the source-site sampling is dense. While this approach has successfully been used on the recovery monochromatic frequency slices, its application in practice calls for wavefield reconstruction of time-domain data. Despite having the option to parallelize, the overall costs of this approach can become prohibitive if we decide to carry out the training and recovery independently for each frequency. Because different frequency slices share information, we propose the use the method of transfer training to make our approach computationally more efficient by warm starting the training with CNN weights obtained from a neighboring frequency slices. If the two neighboring frequency slices share information, we would expect the training to improve and converge faster. Our aim is to prove this principle by carrying a series of carefully selected experiments on a relatively large-scale five-dimensional data synthetic data volume associated with wide-azimuth 3D ocean bottom node acquisition. From these experiments, we observe that by transfer training we are able t significantly speedup in the training, specially at relatively higher frequencies where consecutive frequency slices are more correlated.
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
