Session 2 | Edge-to-end data workflows

Session 2

Edge-to-end data workflows
Convener: Josep de la Puente | Co-coveners: Rui Ferreira, Tiziana Rossetto
Orals
| Wed, 24 May, 14:30–18:00
Poster
| Attendance Wed, 24 May, 18:00–19:00
Orals |
Wed, 14:30
Wed, 18:00
Exascale is not only about applications and capability runs, but entails also workflow managers (orchestrators) and huge amounts of data streams and processes (eventually involving artificial intelligence), from decentralized edge (sensor level) computing, HPC applications (centralized or cloud-based) to the final post-processed results collided to the end- user.

Orals: Wed, 24 May | Sala d'Actes

14:30–15:30
|
GC11-solidearth-53
|
solicited
Scott Callaghan, Philip Maechling, Karan Vahi, Ewa Deelman, Fabio Silva, Kevin Milner, Kim Olsen, Robert Graves, Thomas Jordan, and Yehuda Ben-Zion

Scientific workflows are key to supporting the execution of large-scale simulations in many scientific domains, including solid earth geophysics. Although many different workflow tools exist, they share common features, enabling application developers to express their simulations as a series of linked software elements with data dependencies and then execute the workflow efficiently on distributed resources.

To illustrate the use and benefits of scientific workflows in seismic applications, this talk will describe CyberShake, a probabilistic seismic hazard analysis (PSHA) platform developed by the Southern California Earthquake Center (SCEC). CyberShake uses 3D physics-based wave propagation simulations with reciprocity to calculate ground motions for events from an earthquake rupture forecast (ERF). Typically, CyberShake considers over 500,000 events per site of interest, and then combines the individual ground motions with probabilities from the ERF to produce site-specific PSHA curves. CyberShake has integrated modules from another SCEC workflow application, the Broadband Platform (BBP), enabling CyberShake simulations to include both low-frequency deterministic and high-frequency stochastic content. This talk will discuss the workflow framework that CyberShake utilizes to support campaigns requiring hundreds of thousands of node-hours over months of wall clock time, and the lessons learned through 15 years of CyberShake simulations.

This talk will also reflect on the growth and development of workflow-based simulations and explore the challenges faced by applications in the exascale era, such as managing massive volumes of data, taking full advantage of exascale systems, and the emergence of AI-informed simulations. The talk will discuss ways in which workflow technologies may help mitigate these challenges as we move our science forward.

How to cite: Callaghan, S., Maechling, P., Vahi, K., Deelman, E., Silva, F., Milner, K., Olsen, K., Graves, R., Jordan, T., and Ben-Zion, Y.: Preparing Seismic Applications for Exascale Using Scientific Workflows, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-53, https://doi.org/10.5194/egusphere-gc11-solidearth-53, 2023.

15:30–15:45
|
GC11-solidearth-7
Natalia Poiata, Javier Conejero, Rosa M. Badia, and Jean-Pierre Vilotte

In this work we present a scalable parallelization with PyCOMPSs (Tejedor et al., 2017; the Python binding of COMPSs) of the Python-based workflow BackTrackBB (Poiata et al., 2016) for the automatic detection and location of seismic sources using continuous waveform data recorded by regular to large seismic networks. PyCOMPSs is a task-based programming model for Python applications that relies in a powerful runtime able to extract dynamically the parallelism among tasks and execute them in distributed environments (e.g., HPC Clusters, Cloud infrastructures, etc.) transparently to the users. BackTrackBB with PyCOMPSs implementation allows to fully parallelize the seismic source detection and location process making it efficient and portable in terms of the use of available HPC resources.

We provide details of the BackTrackBB workflow implementation with PyCOMPSs and discuss its performance by presenting the results of the scalability tests and memory usage analysis. All the tests have been performed on the MareNostrum4 High-Performance computer of the Barcelona Supercomputing Centre. The first version of the BackTrackBB with PyCOMPSs workflow was developed in the context of the European Centre Of Excellence (CoE) ChEESE for Exascale computing in solid earth sciences. The initial workflow developments and performance tests made use of a simplified synthetic dataset emulating a large-scale seismic network deployment in a seismically active area and corresponding to 100 vertical sensors recording a month of continuous waveforms at a sampling rate of 100 sps. In the following testing step, the workflow was applied to the real-case two-month long dataset from Vrancea seismic region in Romania (corresponding to the 60-190 km deep earthquakes activity). Real seismic data scenario proved to present some challenges in terms of the data-quality control, that often occurs in the case of continuous waveforms recorded by the seismic observatories. This issue have been resolved and corresponding modifications were included in the final version of BackTrackBB with PyCOMPSs. The real dataset tests showed that the workflow allows improved detection and location of seismic events through the efficient processing of the large continuous seismic data with important performance and scalability improvements.

We show that BackTrackBB with PyCOMPSs workflow enables generation of fully reproducible, seismic catalogues (or seismic catalogues realizations) through the analysis of the continuous large (in terms of the number of seismic stations, data record length and covered area) seismic data-sets. Such implementations making use of advances full-waveform detection and location methods are currently highly-challenging or, some-times, impossible due to the amount of required main memory or unfeasible time to solution. PyCOMPSs has demonstrated to be able to deal with both issues successfully allowing to explore in greater depth the usage with BackTrackBB method. Workflows such as BackTrackBB with PyCOMPSs has the ability to significantly improve the detections and location process that is currently in place at seismological observatories or network operation centres, providing fully reproducing detailed catalogues in the seismically-active regions and allowing multiple input parameters testing (e.g., station configuration, velocity models).

How to cite: Poiata, N., Conejero, J., Badia, R. M., and Vilotte, J.-P.: BackTrackBB workflow for seismic source detection and location with PyCOMPSs parallel computational framework, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-7, https://doi.org/10.5194/egusphere-gc11-solidearth-7, 2023.

15:45–16:00
|
GC11-solidearth-10
Steven Gibbons, Erlend Storrøsten, and Finn Løvholt

Local Probabilistic Tsunami Hazard Analysis (PTHA) aims to quantify the likelihood of a given metric of tsunami inundation at a given location over a given time interval. Seismic PTHA can require the simulation of thousands to tens of thousands of earthquake scenarios and can become computationally intractable when inundation over high-resolution grids is required. The numerical tsunami simulations write out time-series at offshore locations to simulate the wave height that would be recorded on tide gauges at selected locations. The offshore time-series can be calculated at a fraction of the cost of the full inundation calculations. For a stretch of the coast of Eastern Sicily, we explore the extent to which a machine learning procedure trained on a small fraction of the total number of scenarios can predict the inundation map associated with a given offshore time-series. We exploit a set of over 30000 numerical tsunami simulations to train and evaluate the ML-procedure. The ML-based inundation predictions for locations close to the water's edge, which are flooded in many of the scenarios, show excellent correspondence with the numerical simulation results. Predicting inundation at locations further inland, which are flooded in only a small number of the simulations, is more challenging. Mitigating this shortcoming is a priority in the ongoing study.

How to cite: Gibbons, S., Storrøsten, E., and Løvholt, F.: ML Emulation of High Resolution Inundation Maps for Probabilistic Tsunami Hazard Analysis, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-10, https://doi.org/10.5194/egusphere-gc11-solidearth-10, 2023.

16:00–16:15
|
GC11-solidearth-12
|
ECS
Silvia Massaro, Fabio Dioguardi, Alejandra Guerrero, Antonio Costa, Arnau Folch, Roberto Sulpizio, Giovanni Macedonio, and Leonardo Mingari

The atmospheric dispersion of gases (of natural or industrial origins) can be very hazardous to life and the environment if the concentration of some gas species overcome specie-specific thresholds. In this context, the natural variability associated to the natural phenomena has to be explored to provide robust probabilistic gas dispersion hazard assessments. 

VIGIL-1.3 (automatic probabilistic VolcanIc Gas dIspersion modeLling) is a Python simulation tool born to automatize the complex and time-consuming simulation workflow required to process a large number of gas dispersion numerical simulations. It is interfaced with two models: a dilute (DISGAS) and a dense gas (TWODEE-2) dispersion model. The former is used when the density of the gas plume at the source is lower than the atmospheric density (e.g. fumaroles), the latter when the gas density is higher than the atmosphere and the gas accumulates on the ground and may flow due to the density contrast with the atmosphere to form a gravity current (e.g. cold CO2 flows).

In the enhancement of the code towards a higher-scale computing, here we present the ongoing improvements aimed to extend some code functionalities such as memory management, modularity revision, and full-ensemble uncertainty on gas dispersal scenarios (e.g. sampling techniques for gas fluxes and source locations).

Optimizations are also provided in terms of tracking errors, redesignation of the input file, validation of data provided by the users, and addition of the Latin hypercube sampling (LHS) for the post-processing of model outputs.

All these new features will be issued in the future release of the code (VIGIL-2.0) in order to facilitate the users which could run VIGIL on laptops or large supercomputer, and to widen the spectrum of model applications from routinely operational forecast of volcanic gas to long-term hazard and/or risk assessments purposes.

How to cite: Massaro, S., Dioguardi, F., Guerrero, A., Costa, A., Folch, A., Sulpizio, R., Macedonio, G., and Mingari, L.: Improving Probabilistic Gas Hazard Assessment through HPC: Unveiling VIGIL-2.0, an automatic Python workflow for probabilistic gas dispersion modelling, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-12, https://doi.org/10.5194/egusphere-gc11-solidearth-12, 2023.

16:15–16:30
|
GC11-solidearth-42
|
ECS
Farnaz Bayat, Milad Kowsari, Otilio Rojas, Marisol Monterrubio-Velasco, Josep de la Puente, and Benedikt Halldorsson

The strongest earthquakes in Iceland take place in its two large transform zones, the largest being up to magnitude 7.1. As a result, the earthquake hazard in Iceland is the highest in the transform zone. The capital region along with multiple small towns are either in close proximity or on top of the Southwest Iceland transform zone. As a result, the seismic risk is the highest in this region. A new physical 3D finite-fault system model has been developed that model strike-slip faulting in the transform zone as occurring on an array of north-south, near-vertical, dextral strike-slip faults and distributed along the entire transform zone with inter-fault distances ranging from 0.5-5 km. It is well-established that for near-vertical strike-slip faults, large-amplitude and long-period velocity pulses are found in the direction parallel and normal to the fault strike, respectively. The former is due to permanent tectonic displacement as a result of fault slip, and the latter is due to directivity effects. While the former is concentrated in close proximity to the fault and in particular the location of largest subevent of slip on the fault, the directivity pulse is found close to the fault ends and further away along the strike direction, either away from one end or both depending on if the fault rupture is uni- or bilateral, respectively. The forward directivity effect is generally considered to be the most damaging feature of the ground motions, particularly for long-period structures in the near-fault region. The recorded near-fault data in Iceland, however, is relatively sparse, making it difficult to accurately capture the physical characteristics of near-fault ground motions. However, in the ChEESE project we have implemented the new 3D finite-fault system into the CyberShake simulation platform and applied in the kinematic rupture modelling and the corresponding ground motion time history simulation. As a result, we have produced a vast dataset of synthetic ground motion time histories for Southwest Iceland. The synthetic dataset now contains near all possible permutations of near-fault effects and will now be parametrized to reveal the scaling of key near-fault ground motion parameters (e.g., amplitude of pseudo-acceleration spectral, peak ground velocity, and the period of the near-fault pulses) associated with the source (fault slip distribution, and fault plane geometry). This parametrization will increase our understanding of near-fault ground motion and allow the development of simple, but physically realistic near-fault GMM that find practical application in physics-based PSHA.

How to cite: Bayat, F., Kowsari, M., Rojas, O., Monterrubio-Velasco, M., de la Puente, J., and Halldorsson, B.: A first look at the calibration of near-fault motion models to synthetic big data from CyberShake’s application to the Southwest Iceland transform zone, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-42, https://doi.org/10.5194/egusphere-gc11-solidearth-42, 2023.

16:30–17:30
17:30–18:00

Poster: Wed, 24 May, 18:00–19:00 | Poster area (V217)

P7
|
GC11-solidearth-4
Frank Bell

California wild and forest fires in 2017 resulted in over 100 fatalities. While WEA alerts were transmitted to mobiles in selected areas, the power and network outages limited their delivery. WEA is similar to the SMS Broadcast system used elsewhere. It does not require subscription, and can be geotargeted, usually by map polygon. There are already available and in development other alerting technologies. The Emergency Alert System on radio and TV in the U.S. has been in use for many years. It is a broadcast break-in system the overrides program content. This was used in one location for the wildfires, but not elsewhere as geotargeting is not possible with this system. It is and analog broadcast technology architecture. AM and FM Broadcast in the U.S. now has HD Radio that is mixed analog and digital. A limited data message can be carried and used for selective delivery of messages. DAB, DAB+ and DRM also can carry a message payload, which can be used for a selective delivery mechanism when the receiver has location position. This may be in a vehicle radio/navigation system. The current digital television system in the U.S. and some other countries is now being replaced by ATSC 3.0. This provided a superior modulation format, Layered Division Multiplexing (LDM) for delivery of program content and alerts to suitable mobiles.  An IC for UHF reception and prototype mobiles have been developed. No external antenna is required. Bothe of these new technologies are tested as delivering alerts independently of the mobile network. Within the limitations of radio and TV propagation, such capabilities would provide technology redundancy. The television signal propagation may be limited in rural areas, but ATSC 3.0 is capable of having on frequency repeaters to make a single frequency network for improved coverage of program content and alerting. Multilingual alerts based on the CAP Event Terms list with Message Formats are being provided for.

How to cite: Bell, F.: Alerting technologies to save lives in Forest Fires are effective with technology redundancies and multilingual CAP Event Terms basis. Design for worldwide consumer electronics adoption is preferable to reduce costs., Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-4, https://doi.org/10.5194/egusphere-gc11-solidearth-4, 2023.

P8
|
GC11-solidearth-8
|
ECS
Sebastian Noe, Dirk-Philip van Herwaarden, Sölvi Thrastarson, and Andreas Fichtner

We present the second generation of the Collaborative Seismic Earth Model (CSEM), a multi-scale global tomographic Earth model that continuously evolves via successive regional and global-scale refinements. Given finite computational resources, a systematic community effort enables the Earth model construction within the CSEM-architecture. It thereby takes advantage of the distributed human and computing power within the seismological community. The basic update methodology utilizes the current version of the CSEM as the initial model for regional tomographies. This setup allows to consistently incorporate previously accumulated knowledge into each new iteration of the CSEM. The latest generation of the CSEM includes 21 regional refinements from full seismic waveform inversion, ranging from several tens of kilometers to the entire globe. Some noticeable changes since the first generation include detailed local waveform inversions for the Central Andes, Iran, South-east Asia and the Western United States, continental-scale refinements for Africa and Asia and a global long-period tomography in areas that are not included in any of the submodels. Across all regional refinements in the current CSEM, three-component waveform data from 1,637 events and over 700,000 unique source-receiver pairs are utilized to resolve subsurface structure. Minimum periods of models range between 8 and 55 seconds. Using this model as a starting point, a global full-waveform inversion over multiple period bands down to periods of 50 seconds is deployed to ensure that the regional updates predict waveforms and that whole-Earth structure is honored. In this contribution, we will present the CSEM updating scheme and its parameterization, as well as the current state of the model. We show that the model predicts seismic waveforms on global and regional scales. Active participation in the project is encouraged.

How to cite: Noe, S., van Herwaarden, D.-P., Thrastarson, S., and Fichtner, A.: The Collaborative Seismic Earth Model: Generation 2, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-8, https://doi.org/10.5194/egusphere-gc11-solidearth-8, 2023.

P9
|
GC11-solidearth-31
Alejandra Guerrero, Arnau Folch, and Leonardo Mingari

DT-GEO is a project proposed to deal with natural or anthropogenically induced geohazards (earthquakes, volcanoes, landslides and tsunamis) by deploying a Digital Twin of the planet. The prototype will provide a way to visualize, manipulate and understand the response to hypothetical or on-going events by integrating data acquisition  and models. 

Due to the complexity of the development, the project has been divided into different work packages and components. The volcanic phenomena package includes 4 Digital Twin Components (DTCs): volcanic unrest, volcanic ash clouds and ground accumulations, lava flows, and volcanic gas dispersal. The volcanic ash and dispersal deposition component implements a workflow for atmospheric dispersal and ground deposition forecast systems. The workflow is composed of four general units. The first one is the Numerical Weather Prediction (NWP) acquisition (provided by external institutions) refers to both:  automatic obtention of the forecast (up to few days ahead) or the reanalysis (preprocess data from the past) in global or regional scales at different resolutions. Then, the Triggering and Eruption Source Parameters (ESP) is based on predefined communications channels and prioritized by an accuracy rank. The FALL3D model setup and run ensemble simulations, resulting from perturbing ESP values within a range. Finally, the postprocess refers to the compilation of the simulations into hazard maps.

How to cite: Guerrero, A., Folch, A., and Mingari, L.: Volcanic ash dispersal and deposition workflow on HPC, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-31, https://doi.org/10.5194/egusphere-gc11-solidearth-31, 2023.

P10
|
GC11-solidearth-50
Marisol Monterrubio-Velasco, David Modesto, Scott Callaghan, and Josep de la Puente

Large earthquakes are among the most destructive natural phenomena. After a large-magnitude event occurs, a crucial task for hazard assessment is to rapidly and accurately estimate the ground shaking intensities in the affected region. To satisfy real-time constraints, ground shaking is traditionally evaluated with empirical relations called Ground Motion Prediction Equations (GMPE) which can be combined with local amplification factors and early data recordings, when available. Given their nature, GMPEs can be inaccurate to model rarely observed earthquakes, such as large earthquakes. Furthermore, even for very populated databases, GMPEs are characterized by large variances, as earthquakes of similar magnitude and location may have very different outcomes related to complex fault phenomena and wave physics. 

The ML Estimator for Ground Shaking maps (MLESmap) workflow is proposed as a novel procedure that exploits the predictive power of ML algorithms to estimate ground acceleration values a few seconds after a large earthquake occurs. The inferred model can produce peak (spectral) ground motion maps for quasi-real-time applications. Due to its fast assessment, it can further be used to explore uncertainties quickly and reliably. MLESmap is based upon large databases of physics-based seismic scenarios to feed the algorithms.

Our approach (i.e. simulate, train, deploy) can help produce the next generation of ground shake maps, capturing physical information from wave propagation (directivity, topography, site effects) at the velocity of simple empirical GMPE. In this work, we will present the MLESmap workflow, its precision, and a use case.

How to cite: Monterrubio-Velasco, M., Modesto, D., Callaghan, S., and de la Puente, J.: Machine Learning based Estimator for ground Shaking maps, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-50, https://doi.org/10.5194/egusphere-gc11-solidearth-50, 2023.

P11
|
GC11-solidearth-41
|
ECS
Farnaz Bayat, Milad Kowsari, Otilio Rojas, Marisol Monterrubio-Velasco, Josep de la Puente, and Benedikt Halldorsson

The strongest earthquakes in Southwest Iceland take place on a large number of North-South near-vertical dextral strike-slip faults located side-by-side along the entire zone. The capital region along with multiple small towns are in close proximity or on top of this fault system, along with all infrastructure and lifelines of our modern society. As a result, seismic hazard is the highest in this region and performing a probabilistic seismic hazard assessment (PSHA) as the most used procedure to reduce the ruinous effects of large earthquakes is vital. A reliable PSHA requires a reliable ground motion models (GMMs) that can appropriately describe the ground shaking at any given location. However, past PSHA efforts in Iceland did not account for the complex near-fault effects in the form of long-period, high-amplitude velocity pulses that are the most damaging feature of ground motions in the near-fault region. Recently, a new 3D finite-fault system model of the entire bookshelf zone has been proposed for Southwest Iceland. The model has been balanced against the rate of the tectonic plate motions and its seismic activity has been shown to be variable along the entire zone. Given the unknown fault locations, the model allows both for deterministic and random fault locations, and each fault is completely specified in terms of its maximum magnitude, its dimensions and its long-term slip and moment rate. In collaboration with ChEESE project, a realization of a 3000-year finite-fault earthquake catalogue based on the 3D finite-fault system model has been implemented in the CyberShake platform and the ground motion of each earthquake have been simulated for a dense grid of 594 stations. The simulation has been carried out on high-performance computing systems of the Barcelona Supercomputing Centre in Spain. The variation of hypocentral locations and slip distribution on each finite-fault has produced 18 million event-station pairs of synthetic two-horizontal-component low-frequency ground motion time histories that have just become available, those that are simulated less than 40 km from the faults contain near-fault high-amplitude velocity pulses at larger magnitudes, where actual data is nonexistent in Iceland (i.e., above 6.5). Therefore, the purpose of this study is to use a new and vast near-fault dataset of synthetic ground motions to develop a near-fault GMM using an advanced Bayesian Hierarchical Modeling (BHM) for Southwest Iceland.

 

 

How to cite: Bayat, F., Kowsari, M., Rojas, O., Monterrubio-Velasco, M., de la Puente, J., and Halldorsson, B.: A new Near-Fault Earthquake Ground Motion Model for Iceland from Bayesian Hierarchical Modeling  , Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-41, https://doi.org/10.5194/egusphere-gc11-solidearth-41, 2023.

P12
|
GC11-solidearth-43
|
ECS
Milad Kowsari, Manuel Titos, Farnaz Bayat, Carmen Benitez, Marisol Monterrubio-Velasco, Otilio Rojas, Josep de la Puente, and Benedikt Halldorsson

The South Iceland Seismic Zone (SISZ) and Reykjanes Peninsula Oblique Rift (RPOR) in Southwest Iceland together form one of the two major transform zones in the country that have the greatest capacity for the occurrence of destructive earthquakes. Therefore, in these regions, the seismic hazard is highest and performing a probabilistic seismic hazard assessment (PSHA) is vital as the foundation of earthquake resistant building design and seismic risk mitigation. It is well known both from observations as well as physics-based (PB) modeling of earthquake rupture and near-fault ground motion simulations, that the most damaging part of near-fault seismic motion is the velocity pulse, the large-amplitude and long-period pulse-like ground motions found along the fault and away from the ends of strike-slip faults. Such motions cause intense earthquake action primarily on large buildings, such as hydroelectric power plants, dams, powerlines, bridges and pipelines. However, the data is still too limited to enable the reliable calibration of a physically realistic, yet parsimonious, near-fault model that incorporate such effects into empirical ground motions models (GMMs), thereby allowing their incorporation into a formal PSHA. However, in the recent European H2020 ChEESE project, we established a new 3D finite-fault system model for the SISZ-RPOR system that now has facilitated the simulation of finite-fault earthquake catalogues. Moreover, the catalogues have been implemented into the CyberShake platform, the PB earthquake simulator that was adapted to the characteristics of the SISZ-RPOR earthquakes in the ChEESE project. The seismic ground motions of each earthquake in the catalogue have thus been simulated on a dense grid of 594 near-fault stations in Southwest Iceland. The simulation has been carried out on high-performance computing systems of the Barcelona Supercomputing Centre in Spain. Moreover, the hypocentral locations and slip distributions on each synthetic fault have been varied, resulting in approximately 1 million earthquake-station-specific pairs of synthetic low-frequency and high-amplitude near-fault ground motion time histories. In this study, we analyse this dataset using an artificial neural network to reveal its characteristics in terms of amplitudes and the characteristics of near-fault velocity pulses, capturing all key features of such effects. The results will facilitate the incorporation of the near-fault effects into new near-fault and far-field GMMs, that are a key element of conventional PSHA. This will both enable the near-fault PB-PSHA along with the comparison of PSHA from the synthetic dataset vs. the GMMs. This will usher in a new era of PB-PSHA in Iceland.

How to cite: Kowsari, M., Titos, M., Bayat, F., Benitez, C., Monterrubio-Velasco, M., Rojas, O., de la Puente, J., and Halldorsson, B.: Characterization of Earthquake Near-fault Ground Motion Parameters Using an Artificial Neural Network on Synthetic Big Data, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-43, https://doi.org/10.5194/egusphere-gc11-solidearth-43, 2023.

P13
|
GC11-solidearth-44
|
ECS
Milad Kowsari and Benedikt Halldorsson

Throughout history, damaging earthquakes have repeatedly struck in Southwest Iceland, the country’s most populated and seismically active region. There, the interplate earthquakes do not occur on sinistral strike-slip faults parallel to the plate margin, but instead on a dense array of near-vertical dextral faults striking perpendicular to the plate margin. This “bookshelf” faulting has not explicitly been accounted for in probabilistic seismic hazard assessment (PSHA). Instead, incomplete earthquake catalogues and simplistic seismic sources have been used in past PSHA that have used conventional methods. Recently however, a new and physics-based 3D fault system model of the Southwest Iceland transform zone has been proposed that effectively explains the observed Icelandic earthquake catalogues. The model moreover allows the systematic spatial variation of fault slip-rates to be modeled by discrete subzonation of the fault system and the equivalent parameters of seismic activity (Mmax, a- and b-values). Through random realizations of fault locations as postulated by the new model, we have simulated multiple finite-fault earthquake catalogues for the entire bookshelf system for earthquakes ranging from magnitude 4 to 7. This in fact allows us to apply conventional PSHA but instead of using e.g. seismic point sources distributed over a designated seismic source areas, the seismic activity of which is predicted by limited historical catalogues, the synthetic finite-fault catalogues are time-independent and embody fully the first two key elements of PSHA, the seismic source locations along with their activity rates. Using multiple empirical hybrid Bayesian ground motion models (GMMs) that recently have been proposed for Southwest Iceland we have predicted the amplitudes (peak ground accelerations and pseudo-acceleration spectral response) from each synthetic finite-fault earthquake on a grid of hypothetical stations. This enables us to carry out a Monte Carlo PSHA that is based on a physical earthquake fault system model. We present the provisional PSHA results for Southwest Iceland and compare them to other relevant efforts, the Icelandic National Annex to Eurocode 8 and the ESHM20, but most importantly to those of a parallel study that carries out a physics-based PSHA based on synthetic ground motion time histories (on the same hypothetical network) from kinematic earthquake rupture modeling (on the same finite-fault earthquake catalogues) implemented in the CyberShake framework adapted to the Southwest Iceland tectonic situation and earthquake source scaling.

How to cite: Kowsari, M. and Halldorsson, B.: Towards physics-based finite-fault Monte Carlo PSHA for Southwest Iceland based on a new fault system model, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-44, https://doi.org/10.5194/egusphere-gc11-solidearth-44, 2023.

P14
|
GC11-solidearth-45
|
ECS
Manuel Titos Luzon, Milad Kowsari, and Benedikt Halldorsson

The overwhelming success of data-driven models to solve complex predictive real-world problems has made them an effective alternative to the simulation-driven models. In addition to their computational cost, simulation-driven approaches need to be calibrated by actual data that links both to the physical theory and thereby improves both our knowledge and existing models. Similarly, data-driven methods allow deepening the knowledge by analyzing existing data. Therefore, to develop improved predictive models, one needs to pursue a balance between both data-driven and simulation-driven approaches, keeping data as a common pivot. This can be done by using Machine Learning (ML) which is a powerful tool to extract knowledge directly from the data and provide complementary information to the previously developed physics-based models. The main advantage of ML methods is their ability to process massive and complex structure data sets that are difficult to be processed by traditional data-processing methods. Therein lies also the main disadvantage of ML methods i.e., they need massive amounts of data that often are not available. In this study however we take advantage of a new physics-based model of the earthquake fault system of the Southwest Iceland transform zone and generate synthetic, but physically realistic, finite-fault earthquake catalogues. For each earthquake in the catalogues we simulate seismicground motion parameters at a large hypothetical station network  thus generating a massive parametric dataset of synthetic seismic data from earthquakes of magnitude 5 to 7. We will apply multiple types of new generation machine learning techniques such as deep neural network (DNN), deep Bayesian neural networks (DBNN) and deep Gaussian processes (DGP) to investigate the ability and efficiency of the methods in capturing the characteristics of the synthetic dataset in terms of key parameters e.g., ground motion amplitudes, ground motion attenuation versus source-to-site distance and site effects and independent parameters such as earthquake magnitude, fault extend and depth, etc. The ML methods will be trained using a procedure known as greedy layer-wise pretraining where each layer is initialized via the unsupervised pretraining, and the output of the previous layer can be used as the input for the next one. A typical advantage of these pre-trained networks compared to the other deep learning models is that the weights initialization renders the optimization process more effective, providing faster convergence by initializing the network parameters near a convergence region. This can help to avoid underfitting/overfitting problems when the training samples are highly correlated. The results will provide a new insight into the efficiency and usefulness of ML methods on synthetic seismic datasets with implications for their use on actual and more sparser datasets.

How to cite: Titos Luzon, M., Kowsari, M., and Halldorsson, B.: Feasibility of Multiple Advanced Machine Learning Techniques for Synthetic Finite-fault Earthquake Ground Motion Data, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-45, https://doi.org/10.5194/egusphere-gc11-solidearth-45, 2023.

P15
|
GC11-solidearth-48
|
ECS
Luciano Galone, Emanuele Colica, Peter Iregbeyen, Luca Piroddi, Deidun Alan, Gianluca Valentino, Adam Gauci, and Sebastiano D’Amico

Pocket beaches are small beaches bounded by natural promontories, free from direct sedimentary inputs other than those coming from the erosion of their cliffs.

Malta's pocket beaches are one of the most significant geomorphological features of the archipelago. They play an important role for a variety of ecological and economic reasons. Sediment dynamics (mainly sand) is one of the most relevant factors to be considered in those beach system. As the pocket beach system behaves as an integrated unit, periodic bathymetric monitoring is essential - and challenging - from an environmental management perspective.

The SIPOBED project (Satellite Investigation to study POcket BEach Dynamics) develops an integrated tool capable of monitoring sediment dynamics using SAR and digital photogrammetry to monitor beach topographic variations and multispectral UAV and satellite images to derive bathymetry.

Obtaining updated in situ bathymetric measurements is essential to calibrate and re-calibrate the model over time and conduct more actualized and accurate multispectral-derived bathymetry.

In this context, the collection of data by citizens, for instance, bathymetric data collected by private boats abundant in the archipelago, in conjunction with the processing power of modern computing, represents the new challenge of Maltese pocket beach monitoring.

The SIPOBED project is financed by the Malta Council for Science and Technology (MCST, https://mcst.gov.mt/) through the Space Research Fund (Building capacity in the downstream Earth Observation Sector), a programme supported by the European Space Agency.

How to cite: Galone, L., Colica, E., Iregbeyen, P., Piroddi, L., Alan, D., Valentino, G., Gauci, A., and D’Amico, S.: Monitoring the sediment dynamics of Maltese beaches. The SIPOBED project and its future challenges., Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-48, https://doi.org/10.5194/egusphere-gc11-solidearth-48, 2023.

P16
|
GC11-solidearth-72
Wei Li, Megha Chakraborty, Georg Rümpker, and Nishtha Srivastava

Seismic event detection and its magnitude estimation are the two crucial steps in real-time earthquake monitoring and early warning systems. Traditional EEW systems can be limited in their ability to accurately detect the arrival phases (P and S waves) and locate earthquakes, particularly for events with high levels of background noise. Deep learning has emerged as a promising alternative to traditional EEW algorithms, since these algorithms can automatically learn complex patterns and features in seismic data, allowing them to more accurately detect the seismic phase arrival times in the signals. In this study, we first propose a deep learning based architecture DynaPicker which uses a dynamic convolutional neural network to detect seismic body wave phases. Then, the pre-trained model is used to pick the seismic phases on the continuous seismic recording. This model is further combined with another deep-learning model CREIME to perform magnitude estimation. The experimental results on several open-source seismic datasets demonstrate that DynaPicker achieved a higher testing accuracy in seismic phase identification compared to other benchmark models. Additionally, DynaPicker’s robustness in classifying seismic phases was tested on the low-magnitude seismic data polluted by noise. DynaPicker can be adapted to handle input data of varying lengths, making it well-suited for P/S phase picking. When applied to continuous seismic data, DynaPicker can identify more seismic events accurately and produce lower arrival time picking errors than baseline methods. We also found that using the estimated P-phase arrival time of DynaPicker, the CREIME model shows reliable results in estimating the magnitude of the aftershocks of the Turkey earthquake.

How to cite: Li, W., Chakraborty, M., Rümpker, G., and Srivastava, N.: Real time Earthquake detection using Deep Learning, Galileo Conference: Solid Earth and Geohazards in the Exascale Era, Barcelona, Spain, 23–26 May 2023, GC11-solidearth-72, https://doi.org/10.5194/egusphere-gc11-solidearth-72, 2023.