Uncertainty Quantification of Landslide Generated Waves Using Gaussian Process Emulation and Variance-Based Sensitivity Analysis

Branwen Snelling, S.j. Neethling, Kevin Horsburgh, G. S. Collins, Matthew D. PiggottWater 2020 · DOI

TL;DR

This paper tackles the challenge of predicting the hazard posed by landslide-generated waves (LGWs) by quantifying the uncertainty in their simulation. They use Gaussian Process Emulation and variance-based sensitivity analysis to pinpoint which landslide characteristics most significantly influence wave behavior, enabling more efficient and focused hazard assessments.

What Problem Does This Paper Address?

Landslide-generated waves (LGWs) are a significant natural hazard, but predicting their severity is difficult due to the inherent uncertainties in the complex physical processes involved. Simulating these events, often using computationally intensive methods like Smooth Particle Hydrodynamics (SPH), requires numerous input parameters related to the landslide's properties (e.g., volume, rheology, initial conditions) and the surrounding environment. Running enough simulations to fully explore the impact of all these uncertainties on the resulting wave characteristics (like height or inundation potential) is computationally prohibitive. This paper aims to address this by identifying which input parameters contribute most to the uncertainty in LGW predictions, allowing researchers to prioritize efforts in reducing uncertainty for the most critical factors.

Methods

The core of this study lies in combining two powerful techniques:

  • Smooth Particle Hydrodynamics (SPH): This is the "truth" model used to simulate LGWs. SPH is a meshless Lagrangian method that excels at modeling fluid dynamics with complex behaviors, including flexible boundary conditions and various rheologies (like the Herschel-Bulkley model used here for the landslide).
  • Gaussian Process Emulation (GPE): Because running thousands of SPH simulations is infeasible, the authors build a computationally cheaper "emulator" using GPE. This emulator learns the relationship between input parameters and output wave characteristics from a limited set of SPH simulations.
  • Variance-Based Sensitivity Analysis (VBSA): Once the emulator is trained and validated, VBSA is employed. This technique, specifically using Sobol' indices, quantifies how much each input parameter (and their interactions) contributes to the overall variance (uncertainty) in the simulated wave output.

The process involves:

  1. Generating a training dataset of SPH simulations with carefully chosen input parameter ranges.
  2. Training a GPE on a subset of this data.
  3. Validating the GPE against the remaining SPH simulation data.
  4. Using the validated GPE to extensively sample the parameter space and calculate Sobol' indices.

Datasets

The study utilizes a simulated dataset generated by their SPH model. They created a training dataset of 120 SPH simulations, sampling input parameters using a Sobol' sequence to ensure good coverage of the parameter space. This dataset was then used to train and validate the Gaussian Process Emulator. No external real-world datasets were used for training or direct validation of the SPH simulations themselves, though parameter ranges were informed by literature on real-world landslide events.

Key Insights and Ideas

  • Volume and Submergence Dominate Wave Height Uncertainty: The most surprising and impactful finding is that the landslide's volume and its initial submergence depth are the primary drivers of uncertainty in maximum wave height. This suggests that accurately characterizing these two parameters is crucial for reliable LGW hazard assessment.
  • Rheology Matters Less Than You Might Think (for Wave Height): Counterintuitively, the detailed rheological parameters of the landslide (viscosity, yield stress, shear-thinning exponent) contribute relatively little to the uncertainty in maximum wave height. This is likely due to the broad range of boundary conditions explored, which can mask the subtle effects of rheology.
  • Run-up Introduces New Key Players: When considering wave run-up on a shore as the hazard metric, the slope angle of the inundated shore becomes a highly influential parameter, second only to landslide volume. This highlights the importance of characterizing the target location as much as the source for accurate hazard assessment.
  • Interactions Can Be Substantial: While individual rheological parameters might have low main Sobol' indices, their interactions with other parameters (like volume and slope angle) can be significant, particularly in unlocking the potential for small slides to generate large waves under specific conditions (e.g., low yield stress).

Key Results

  • Wave Height: Landslide volume (V) and initial submergence depth (D) were identified as the most influential parameters, contributing the most to uncertainty in maximum wave height. Slope angle (θ) and boundary layer viscosity (ψ) also showed moderate contributions.
  • Displaced Volume: The results for positive displaced water volume mirrored those for wave height, with V and D being the dominant contributors to uncertainty. Slope angle (θ) showed a slightly larger influence compared to wave height.
  • Run-up: For approximated run-up, landslide volume (V) remained the most influential parameter. However, the inundated shore slope angle (β) emerged as the second most influential parameter, underscoring the importance of the target environment.
  • Convergence: The variance-based decomposition analysis achieved satisfactory convergence for Sobol' indices by using 8000 emulator samples.

Limitations and Future Work

  • Idealized Geometry: The study uses an idealized geometry for the landslide and SPH domain, which may not fully capture the complexities of real-world scenarios.
  • Run-up Approximation: The use of an analytical approximation for run-up, while practical, may not fully represent the intricate physics of wave inundation.
  • Boundary Condition Uncertainty: The boundary layer viscosity parameter, while influential, lacks precise physical constraints, making its accurate determination challenging.
  • Future Work:
    • Physically constraining the boundary layer viscosity would reduce uncertainty.
  • Incorporating parameters related to wave propagation and inundation phases (e.g., propagation distance, detailed shore bathymetry) using more sophisticated hydrodynamic models would provide deeper insights.
  • Extending the analysis to 3D LGW models could reveal previously unexplored parameter influences and interactions.
  • Further investigation into the interactions between yield stress and boundary conditions is warranted.

Summary updated

Abstract

Simulations of landslide generated waves (LGWs) are prone to high levels of uncertainty. Here we present a probabilistic sensitivity analysis of an LGW model. The LGW model was realised through a smooth particle hydrodynamics (SPH) simulator, which is capable of modelling fluids with complex rheologies and includes flexible boundary conditions. This LGW model has parameters defining the landslide, including its rheology, that contribute to uncertainty in the simulated wave characteristics. Given the computational expense of this simulator, we made use of the extensive uncertainty quantification functionality of the Dakota toolkit to train a Gaussian process emulator (GPE) using a dataset derived from SPH simulations. Using the emulator we conducted a variance-based decomposition to quantify how much each input parameter to the SPH simulation contributed to the uncertainty in the simulated wave characteristics. Our results indicate that the landslide’s volume and initial submergence depth contribute the most to uncertainty in the wave characteristics, while the landslide rheological parameters have a much smaller influence. When estimated run-up is used as the indicator for LGW hazard, the slope angle of the shore being inundated is shown to be an additional influential parameter. This study facilitates probabilistic hazard analysis of LGWs, because it reveals which source characteristics contribute most to uncertainty in terms of how hazardous a wave will be, thereby allowing computational resources to be focused on better understanding that uncertainty.
BibTeX
@inproceedings{snelling2020uncertainty,
  author = {Branwen Snelling and S.J. Neethling and Kevin Horsburgh and G. S. Collins and Matthew D. Piggott},
  title = {{Uncertainty Quantification of Landslide Generated Waves Using Gaussian Process Emulation and Variance-Based Sensitivity Analysis}},
  booktitle = {Water},
  year = {2020},
  doi = {10.3390/w12020416},
}
Figure