Gaussian Processes for Probabilistic Estimates of Earthquake Ground Shaking: A 1-D Proof-of-Concept

Sam A. Scivier, Tarje Nissen-Meyer, Paula Koelemeijer, Atılım Güneş Baydin · 2024

Read Gaussian Processes for Probabilistic Estimates of Earthquake Ground Shaking: A 1-D Proof-of-Concept by Sam A. Scivier, Tarje Nissen-Meyer, Paula Koelemeijer, Atılım Güneş Baydin (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-03299.txt.

Introduction

Sam Scivier and co-authors fold disagreement between Earth models into earthquake shaking forecasts. Their 2024 paper notes that many maps of wave speed fit the same region, and picking one at random bends the forecast. Current hazard work ignores this source of doubt.

As a proof of concept, they fit a Gaussian process to two made-up one-dimensional speed profiles at once. The fit's spread grew exactly where the two profiles disagreed. Shaking was then simulated through many sampled profiles. The range of ground motion came out far wider than either input model alone would suggest. The team argues hazard studies should carry this doubt openly in maps that guide building and planning choices.

Catalog record

CREATORSam A. Scivier, Tarje Nissen-Meyer, Paula Koelemeijer, Atılım Güneş Baydin
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-03299

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Gaussian Processes for Probabilistic Estimates of Earthquake Ground Shaking: A 1-D Proof-of-Concept Sam A. Scivier, Tarje Nissen-Meyer, Paula Koelemeijer, Atılım Güneş Baydin Abstract Estimates of seismic wave speeds in the Earth (seismic velocity models) are key input parameters to earthquake simulations for ground motion prediction. Owing to the non-uniqueness of the seismic inverse problem, typically many velocity models exist for any given region. The arbitrary choice of which velocity model to use in earthquake simulations impacts ground motion predictions. However, current hazard analysis methods do not account for this source of uncertainty. We present a proof-of-concept ground motion prediction workflow for incorporating uncertainties arising from inconsistencies between existing seismic velocity models. Our analysis is based on the probabilistic fusion of overlapping seismic velocity models using scalable Gaussian process (GP) regression. Specifically, we fit a GP to two synthetic 1-D velocity profiles simultaneously, and show that the predictive uncertainty accounts for the differences between the models. We subsequently draw velocity model samples from the predictive distribution and estimate peak ground displacement using acoustic wave propagation through the velocity models. The resulting distribution of possible ground motion amplitudes is much wider than would be predicted by simulating shaking using only the two input velocity models. This proof-of-concept illustrates the importance of probabilistic methods for physics-based seismic hazard analysis. 1 Introduction Seismic velocity models — estimates of the Earth’s seismic wave speeds — underpin earthquake ground motion prediction in seismic hazard analysis, as they are key inputs to wave equation solvers. They continue to be produced at different resolutions and scales, stemming from different methods (e.g., tomography [ 1 ] , reflection surveys [ 2 ] ). The seismic inverse problem is ill-posed as there are not enough data to constrain a unique true Earth model [ 3 ] . As such, many overlapping velocity models exist for a given region. Consequently, the choice of which velocity model to use in ground motion prediction is often arbitrary. Nevertheless, it has a significant impact on the results as different models have different structures, length scales, and amplitudes. The key output of seismic hazard analysis is an estimate of peak ground motion, to assess potential infrastructure damage and inform earthquake engineering. Most commonly, empirical ground motion models (GMMs) [ 4 , 5 ] are used to predict the median and uncertainty of a ground motion parameter (e.g., peak ground displacement — PGD) for earthquake scenarios [ 6 ] . They make rapid predictions, but drastically simplify the underlying physical processes. Importantly, they approximate the effect of seismic velocities on ground motion, typically only using the average shear wave velocity in the uppermost 30 m [ 7 ] . GMMs are thus limited in accuracy and reliability. An alternative is to simulate earthquake scenarios in 3-D by solving the wave equation and extracting PGD estimates, requiring 3-D seismic velocity information as input. However, there are two issues: (i) simulating many earthquakes is computationally costly, and (ii) choices of input parameters are subjective, including the input velocity model. To address the first issue, recent advances in machine learning have begun accelerating wave propagation methods [ 8 , 9 , 10 , 11 ] . However, current physics-based hazard analysis workflows do not consider inconsistencies between velocity models. This omits a key source of uncertainty, given that predicted ground motion can be drastically impacted by velocity structure. One possible solution is to fuse different velocity models, and use the output in earthquake simulations. Unfortunately, existing methods for velocity model fusion e.g., 12 , 13 , 14 typically do not produce probabilistic outputs, limiting their ability to account for differences between velocity models. In this study, we propose a workflow to account for inconsistencies between seismic velocity models in ground motion prediction. Our method is based on the probabilistic fusion of velocity models using Gaussian processes (GPs), and estimates uncertainties owing to differences between them. We then produce probabilistic ground motion predictions with respect to these uncertainties by drawing velocity model samples from the GP predictive distribution, simulating acoustic wave propagation using each sample, and extracting the PGD predictions. We illustrate that such a probabilistic method is necessary to capture the spread of possible ground motion scenarios. Our key contributions are as follows: (i) We present a workflow for probabilistic earthquake ground motion prediction that accounts for inconsistencies between seismic velocity models. (ii) We demonstrate the capability of scalable GPs for the probabilistic fusion of different estimates of the same physical parameter, through a synthetic example using 1-D seismic velocity models. The code for this work is written in Python and is available at Scivier et al. 2024 . Figure 1: Comparison of SVGPR and PPGPR for the probabilistic fusion of seismic velocity models. (a) shows the input synthetic 1-D seismic velocity profiles with depth. (b) and (c) show the fusion results of SVGPR and PPGPR, and (d) shows the 200 function samples drawn from the PPGPR predictive distribution used in Section 4 . The shading in (b) and (c) show the SVGP posterior predictive distribution, q_{\text{SVGP}}\left(\mathbf{y_{*}}\right) , and the PPGPR latent predictive distribution, q_{\text{PPGPR}}\left(\mathbf{f_{*}}\right) , respectively, in terms of distance from the predictive means in standard deviations. 2 Gaussian processes and data fusion GPs [ 16 ] are a class of non-parametric models for defining a distribution over function spaces. They are widely used for regression, providing robust uncertainty quantification and predictive performance. Unfortunately, exact GP regression is limited in scalability owing to a computational cost of \mathcal{O}(n^{3}) , where n is the number of data points. Despite the small datasets used in this study, seismic datasets of realistic size can have n\sim 10^{6}\textup{--}10^{7} . To overcome this, approximate GP methods have been developed [ 17 ] . One popular method for scalable GP inference is the sparse variational Gaussian process (SVGP) [ 18 , 19 ] , which applies variational inference to fit GPs. SVGPs introduce a set of inducing variables to approximate the full dataset using a smaller set of points m\ll n . Thus, the computational cost is reduced to \mathcal{O}(nm^{2}+m^{3}) , making SVGPs practical to apply to large 2-D and 3-D datasets. In this study, we use scalable GP regression for fusing seismic velocity models by fitting a GP to multiple datasets simultaneously. A key advantage of GPs is their modelling of covariance structure, which enables samples matching the spatial patterns of the input data to be drawn from the predictive distribution. We aim to model inconsistencies between velocity estimates as uncertainty in the GP predictive distribution. Thus, we need to understand the form of the predictive variance in the SVGP model. Below we provide a brief summary of the key SVGP results to highlight the relevant context for our work. The interested reader is referred to Titsias 2009 , Matthews et al. 2016 , Jankowiak et al. 2020 , and Murphy 2023 for full derivations and explanations. The inputs are the training set (coordinates of the input velocity models), \mathbf{X} , the inducing point locations, \mathbf{Z} , and the points at which we wish to predict, \mathbf{X_{*}} . Then \mathbf{f_{X}},\mathbf{f_{Z}},\mathbf{f_{*}} are the (unknown) velocity values that we predict at these locations; and \mathbf{y} are the observed data (input velocity values). SVGP-based methods approximate

Read the complete text