Towards adjoint-based inversion for rheological parameters in nonlinear viscous mantle flow

Jennifer Worthen, Georg Stadler, Noemi Petra, Michael Gurnis, Omar Ghattas

Research output: Contribution to journalArticle

Abstract

We address the problem of inferring mantle rheological parameter fields from surface velocity observations and instantaneous nonlinear mantle flow models. We formulate this inverse problem as an infinite-dimensional nonlinear least squares optimization problem governed by nonlinear Stokes equations. We provide expressions for the gradient of the cost functional of this optimization problem with respect to two spatially-varying rheological parameter fields: the viscosity prefactor and the exponent of the second invariant of the strain rate tensor. Adjoint (linearized) Stokes equations, which are characterized by a 4th order anisotropic viscosity tensor, facilitates efficient computation of the gradient. A quasi-Newton method for the solution of this optimization problem is presented, which requires the repeated solution of both nonlinear forward Stokes and linearized adjoint Stokes equations. For the solution of the nonlinear Stokes equations, we find that Newton's method is significantly more efficient than a Picard fixed point method. Spectral analysis of the inverse operator given by the Hessian of the optimization problem reveals that the numerical eigenvalues collapse rapidly to zero, suggesting a high degree of ill-posedness of the inverse problem. To overcome this ill-posedness, we employ Tikhonov regularization (favoring smooth parameter fields) or total variation (TV) regularization (favoring piecewise-smooth parameter fields). Solution of two- and three-dimensional finite element-based model inverse problems show that a constant parameter in the constitutive law can be recovered well from surface velocity observations. Inverting for a spatially-varying parameter field leads to its reasonable recovery, in particular close to the surface. When inferring two spatially varying parameter fields, only an effective viscosity field and the total viscous dissipation are recoverable. Finally, a model of a subducting plate shows that a localized weak zone at the plate boundary can be partially recovered, especially with TV regularization.

Original languageEnglish (US)
Pages (from-to)23-34
Number of pages12
JournalPhysics of the Earth and Planetary Interiors
Volume234
DOIs
StatePublished - 2014

Fingerprint

Earth mantle
inversions
mantle
inverse problem
Newton methods
optimization
viscosity
tensors
gradients
eigenvalue
parameter
inversion
plate boundary
strain rate
spectral analysis
spectrum analysis
dissipation
eigenvalues
recovery
exponents

Keywords

  • Adjoint method
  • Inverse problem
  • Mantle flow
  • Mantle viscosity
  • Newton's method
  • Nonlinear Stokes equations
  • Tikhonov and total variation regularization

ASJC Scopus subject areas

  • Geophysics
  • Space and Planetary Science
  • Physics and Astronomy (miscellaneous)
  • Astronomy and Astrophysics

Cite this

Towards adjoint-based inversion for rheological parameters in nonlinear viscous mantle flow. / Worthen, Jennifer; Stadler, Georg; Petra, Noemi; Gurnis, Michael; Ghattas, Omar.

In: Physics of the Earth and Planetary Interiors, Vol. 234, 2014, p. 23-34.

Research output: Contribution to journalArticle

Worthen, Jennifer ; Stadler, Georg ; Petra, Noemi ; Gurnis, Michael ; Ghattas, Omar. / Towards adjoint-based inversion for rheological parameters in nonlinear viscous mantle flow. In: Physics of the Earth and Planetary Interiors. 2014 ; Vol. 234. pp. 23-34.
@article{7b880d8e1a1f420f8c636a027e583525,
title = "Towards adjoint-based inversion for rheological parameters in nonlinear viscous mantle flow",
abstract = "We address the problem of inferring mantle rheological parameter fields from surface velocity observations and instantaneous nonlinear mantle flow models. We formulate this inverse problem as an infinite-dimensional nonlinear least squares optimization problem governed by nonlinear Stokes equations. We provide expressions for the gradient of the cost functional of this optimization problem with respect to two spatially-varying rheological parameter fields: the viscosity prefactor and the exponent of the second invariant of the strain rate tensor. Adjoint (linearized) Stokes equations, which are characterized by a 4th order anisotropic viscosity tensor, facilitates efficient computation of the gradient. A quasi-Newton method for the solution of this optimization problem is presented, which requires the repeated solution of both nonlinear forward Stokes and linearized adjoint Stokes equations. For the solution of the nonlinear Stokes equations, we find that Newton's method is significantly more efficient than a Picard fixed point method. Spectral analysis of the inverse operator given by the Hessian of the optimization problem reveals that the numerical eigenvalues collapse rapidly to zero, suggesting a high degree of ill-posedness of the inverse problem. To overcome this ill-posedness, we employ Tikhonov regularization (favoring smooth parameter fields) or total variation (TV) regularization (favoring piecewise-smooth parameter fields). Solution of two- and three-dimensional finite element-based model inverse problems show that a constant parameter in the constitutive law can be recovered well from surface velocity observations. Inverting for a spatially-varying parameter field leads to its reasonable recovery, in particular close to the surface. When inferring two spatially varying parameter fields, only an effective viscosity field and the total viscous dissipation are recoverable. Finally, a model of a subducting plate shows that a localized weak zone at the plate boundary can be partially recovered, especially with TV regularization.",
keywords = "Adjoint method, Inverse problem, Mantle flow, Mantle viscosity, Newton's method, Nonlinear Stokes equations, Tikhonov and total variation regularization",
author = "Jennifer Worthen and Georg Stadler and Noemi Petra and Michael Gurnis and Omar Ghattas",
year = "2014",
doi = "10.1016/j.pepi.2014.06.006",
language = "English (US)",
volume = "234",
pages = "23--34",
journal = "Physics of the Earth and Planetary Interiors",
issn = "0031-9201",
publisher = "Elsevier",

}

TY - JOUR

T1 - Towards adjoint-based inversion for rheological parameters in nonlinear viscous mantle flow

AU - Worthen, Jennifer

AU - Stadler, Georg

AU - Petra, Noemi

AU - Gurnis, Michael

AU - Ghattas, Omar

PY - 2014

Y1 - 2014

N2 - We address the problem of inferring mantle rheological parameter fields from surface velocity observations and instantaneous nonlinear mantle flow models. We formulate this inverse problem as an infinite-dimensional nonlinear least squares optimization problem governed by nonlinear Stokes equations. We provide expressions for the gradient of the cost functional of this optimization problem with respect to two spatially-varying rheological parameter fields: the viscosity prefactor and the exponent of the second invariant of the strain rate tensor. Adjoint (linearized) Stokes equations, which are characterized by a 4th order anisotropic viscosity tensor, facilitates efficient computation of the gradient. A quasi-Newton method for the solution of this optimization problem is presented, which requires the repeated solution of both nonlinear forward Stokes and linearized adjoint Stokes equations. For the solution of the nonlinear Stokes equations, we find that Newton's method is significantly more efficient than a Picard fixed point method. Spectral analysis of the inverse operator given by the Hessian of the optimization problem reveals that the numerical eigenvalues collapse rapidly to zero, suggesting a high degree of ill-posedness of the inverse problem. To overcome this ill-posedness, we employ Tikhonov regularization (favoring smooth parameter fields) or total variation (TV) regularization (favoring piecewise-smooth parameter fields). Solution of two- and three-dimensional finite element-based model inverse problems show that a constant parameter in the constitutive law can be recovered well from surface velocity observations. Inverting for a spatially-varying parameter field leads to its reasonable recovery, in particular close to the surface. When inferring two spatially varying parameter fields, only an effective viscosity field and the total viscous dissipation are recoverable. Finally, a model of a subducting plate shows that a localized weak zone at the plate boundary can be partially recovered, especially with TV regularization.

AB - We address the problem of inferring mantle rheological parameter fields from surface velocity observations and instantaneous nonlinear mantle flow models. We formulate this inverse problem as an infinite-dimensional nonlinear least squares optimization problem governed by nonlinear Stokes equations. We provide expressions for the gradient of the cost functional of this optimization problem with respect to two spatially-varying rheological parameter fields: the viscosity prefactor and the exponent of the second invariant of the strain rate tensor. Adjoint (linearized) Stokes equations, which are characterized by a 4th order anisotropic viscosity tensor, facilitates efficient computation of the gradient. A quasi-Newton method for the solution of this optimization problem is presented, which requires the repeated solution of both nonlinear forward Stokes and linearized adjoint Stokes equations. For the solution of the nonlinear Stokes equations, we find that Newton's method is significantly more efficient than a Picard fixed point method. Spectral analysis of the inverse operator given by the Hessian of the optimization problem reveals that the numerical eigenvalues collapse rapidly to zero, suggesting a high degree of ill-posedness of the inverse problem. To overcome this ill-posedness, we employ Tikhonov regularization (favoring smooth parameter fields) or total variation (TV) regularization (favoring piecewise-smooth parameter fields). Solution of two- and three-dimensional finite element-based model inverse problems show that a constant parameter in the constitutive law can be recovered well from surface velocity observations. Inverting for a spatially-varying parameter field leads to its reasonable recovery, in particular close to the surface. When inferring two spatially varying parameter fields, only an effective viscosity field and the total viscous dissipation are recoverable. Finally, a model of a subducting plate shows that a localized weak zone at the plate boundary can be partially recovered, especially with TV regularization.

KW - Adjoint method

KW - Inverse problem

KW - Mantle flow

KW - Mantle viscosity

KW - Newton's method

KW - Nonlinear Stokes equations

KW - Tikhonov and total variation regularization

UR - http://www.scopus.com/inward/record.url?scp=84904613702&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84904613702&partnerID=8YFLogxK

U2 - 10.1016/j.pepi.2014.06.006

DO - 10.1016/j.pepi.2014.06.006

M3 - Article

VL - 234

SP - 23

EP - 34

JO - Physics of the Earth and Planetary Interiors

JF - Physics of the Earth and Planetary Interiors

SN - 0031-9201

ER -