Sensitivity Calculations for Poisson's Equation via the Adjoint Field Method

Alireza Aghasi, Eric L. MillerIEEE GRSL 2012 · DOI

TL;DR

This paper presents a unified and clear derivation of adjoint field methods for calculating sensitivity information in inverse problems governed by Poisson's equation. It explicitly addresses the role of general boundary conditions and clarifies the relationship between adjoint calculations for gradient descent and Gauss-Newton inversion methods, making the approach more accessible and adaptable.

What Problem Does This Paper Address?

Many geophysical imaging techniques, such as electrical resistance tomography (ERT), electrical impedance tomography (EIT), and induced polarization (IP), rely on solving inverse problems. A crucial step in these inversions is calculating "sensitivity information," which quantifies how changes in the subsurface properties (like conductivity) affect the measured data. Traditional methods for obtaining this information can be computationally intensive and complex, especially when dealing with various types of boundary conditions (e.g., Neumann, Dirichlet, mixed). Existing adjoint field methods, while powerful, often lack a unified derivation that clearly explains the role of boundary conditions and the distinct requirements for different inversion algorithms (like gradient descent versus Gauss-Newton). This paper aims to bridge this gap by providing a transparent, step-by-step derivation of adjoint field sensitivities for Poisson's equation, making it easier for researchers to apply these techniques to their specific problems.

Methods

The core of this paper is the derivation of sensitivity calculations using the adjoint field method. The authors start with the forward model described by Poisson's equation with general boundary conditions. They then consider a small perturbation in the conductivity and analyze its effect on the potential field and the measured data.

The key steps involve:

  • Defining the Adjoint System: An adjoint Poisson equation is introduced, driven by "adjoint sources" that are essentially Dirac delta functions located at the measurement points. The solution to this adjoint equation is the "adjoint field."
  • Relating Perturbations to Adjoint Fields: Using Green's theorem and algebraic manipulation, the authors show that the change in the measured data due to a conductivity perturbation can be expressed as an integral involving the adjoint field and the conductivity perturbation.
  • Deriving Sensitivity Kernels: This integral form is then related to a desired sensitivity kernel, which directly maps a conductivity perturbation to a change in the measurement. The paper explicitly derives these kernels for the Poisson equation with general boundary conditions.
  • Differentiating Inversion Approaches: A significant contribution is the clear distinction and relationship between the adjoint calculations needed for:
  • Gradient Descent: Requires a single "composite" adjoint field, derived from all measurement residuals.
  • Gauss-Newton: Requires multiple adjoint fields, one for each measurement, to construct the Jacobian matrix. The paper shows how the gradient descent sensitivity can be obtained from the Gauss-Newton sensitivities.

The derivation emphasizes using standard vector calculus identities and transparent manipulation of partial differential equations, aiming for clarity and adaptability.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Unified Derivation for General Boundary Conditions: The paper provides a comprehensive treatment of how various boundary conditions (Neumann, Dirichlet, mixed) are incorporated into the adjoint field derivation, which is often a point of confusion or simplification in other works.
  • Clear Link Between Gradient Descent and Gauss-Newton: A key insight is the explicit demonstration of how the adjoint fields calculated for a Gauss-Newton method can be combined to obtain the sensitivity required for gradient descent methods. This highlights the efficiency of the adjoint approach, as fewer adjoint solves might be needed for gradient-based methods.
  • Pedagogical Approach: The authors prioritize a step-by-step derivation using accessible mathematical tools, aiming to make the concepts easily adaptable to other physics-based inverse problems beyond Poisson's equation.
  • Self-Adjoint Nature of Poisson's Equation: The paper leverages the fact that Poisson's equation is self-adjoint, simplifying the adjoint system to have the same form as the forward problem, differing only in the source term.

Key Results

The paper successfully derives explicit expressions for the sensitivity kernels (Frechet derivatives) for Poisson's equation with general boundary conditions. It demonstrates that:

  • For gradient descent methods, the sensitivity can be computed using one forward solve and one "composite" adjoint solve.
  • For Gauss-Newton methods, the sensitivity requires one forward solve and N adjoint solves (where N is the number of measurements) to construct the Jacobian.
  • The paper provides the mathematical link to show how the N adjoint solutions for Gauss-Newton can be used to form the single composite adjoint solution for gradient descent.

Limitations and Future Work

  • Real-Valued Conductivity: The derivation is primarily presented for real-valued conductivity (ERT problems), although the authors note that the approach can be adapted for complex-valued conductivities (EIT) or other physical quantities.
  • Single Source: The derivation considers a single source for simplicity, but the authors mention that the approach generalizes to multiple sources.
  • Nonhomogeneous Boundary Conditions: While the paper discusses the impact of nonhomogeneous boundary conditions, it notes that the adjoint system derived is not strictly the adjoint of the nonhomogeneous forward system, with further details provided in the Appendix.
  • Future Work: The authors suggest that the clear derivation and explicit treatment of boundary conditions make the approach easily adaptable to other systems of interest, implying potential applications to different physical models.

Summary updated

BibTeX
@article{aghasi2012sensitivity,
  author = {Alireza Aghasi and Eric L. Miller},
  title = {{Sensitivity Calculations for Poisson's Equation via the Adjoint Field Method}},
  journal = {IEEE Geosci. Remote. Sens. Lett.},
  year = {2012},
  doi = {10.1109/lgrs.2011.2164052},
}

Loading related content...

Figure