Multiscale modeling of natural hazards

MPM, LBM-DEM and ML


Krishna Kumar, krishnak@utexas.edu
Rei Hosseini, Qiuyu (Amber) Wang, Joseph Vantassel, Chihun Sung and Thiago Araujo
University of Texas at Austin



Prof. Boominathan's Felicitation

IITM, 25th June 2021

Thank you Sir! You are the best!!!

Geoelements Extremescale Computational Geomechanics

  • Material Point Method
  • Lattice-Boltzmann + Discrete Element Method
  • Finite Element Method - Thermo-Hydro Mechanical Coupling
  • Lattice Element Method
  • Machine Learning
View the Geoelements website for more information and software tools

Oso landslide (2014)

8 million cubic meters of glacial deposits and water-filled debris material transported to a distance of 1 km (Haugerud., 2014).

Oso landslide (2014)

Oso landslide: Geology

Glaciolacustrine deposits on lower portion of Bear Lake Rhythmites with failure surface.
Deformation till with flame structures in fine-grained glacio-lacustrine deposit EB7 (depth 65 ft).


Courtesy of Dr Gunnar Schlieder

Mesh-based vs Mesh-free techniques

Material Point Method

MPM model setup

MPM simulation of Oso landslide

Yerro et al., 2017

Oso landslide simulation

What is Ray Tracing?

Two-phase MPM rendering

MPM Oso landslide rendering

Submarine run-out

Credit: Amanda Murphy (2016)

Mechanism of submarine runout

Lattice Boltzmann - MRT

Real Fluid vs LBM Idealisation
LBM D2Q9 Model

\[f_{i}(x + dx, t +\Delta t) - f_{i}(x, t) = -S_{\alpha i}( f_{i}(x, t) - f_{i} ^ {eq}(x, t))\]
  • $S_{\alpha i}$ is the collisional matrix.
  • Probability density of finding a particle : $f(x,\varepsilon, t) $, where, x is position, $\varepsilon$ is velocity, and t is time.
Streaming
Collision

LBM - DEM a = 0.8 & 10,000 particles



  • LBM Nodes = 50 Million : DEM grains = 10000 discs
  • Run-time = 4 hours
  • Speedup = 125x on a Pascal P100

Collapse on slopes: loose v dense

runout evolution ('a' of 0.8)

Loose v dense: Runout distance

Loose
Dense

Loose v dense: Initiation phase

initial runout evolution ('a' of 0.8)

Loose v dense: Initiation phase

Loose
Dense

Pore-pressure distribution along the failure plane during initiation.

Loose v dense: Runout phase

Attack angle ('a' of 0.8) $t = 3 \tau_c $

Loose v dense: Runout phase

Loose
Dense
Water entrainment front (~15d length) at a slope of 5*

Two-phase MPM Submarine landslide

Multiphase LBM

Multiphase LBM: Effect of varying the contact angle

Fully hydrophilic surface
(θ = 0°)
Fully hydrophobic surface
(θ = 180°)
Neutral surface
(θ = 90°)

Multiphase LBM: Hysteresis in Hamburg sand

Multiphase LBM: Capillary structures

Neutral surface
(θ = 90°)

Multiphase LBM: Hysteresis

Multiphase LBM: Origin of Hysteresis

Two-phase MPM Rainfall induced landslide

Multichannel Analysis of Surface Waves (MASW)

Machine Learning with Convolution Neural Network

Machine Learning with CNN

Machine Learning with Convolution Neural Network

CNN Full Waveform Inversion: Architecture

Vantassel et al., (2021)

CNN Full Waveform Inversion: Training

CNN Full Waveform Inversion

CNN prediction of subsoil profile

Thank you!



Krishna Kumar

krishnak@utexas.edu