Adjoint Data Assimilation into a 3D Unstructured Mesh Coastal Finite Element Model

F. Fang, Christopher C. Pain, Matthew D. Piggott, Gerard Gorman, Graham M. Copeland, C. Mendes de Oliveira, A.j.h. Goddard, Igor Gejadze2004 · DOI

TL;DR

This paper introduces a novel adaptive mesh adjoint model for simulating coastal flows, enabling more accurate representations of complex water dynamics. By allowing the forward and adjoint models to use different, dynamically adapting meshes, the method achieves good inversion results for boundary conditions, even with a slight inconsistency between the discrete schemes.

What Problem Does This Paper Address?

Accurate simulation of coastal flows is crucial for understanding and predicting phenomena like tides, currents, and pollutant dispersion. However, these simulations are often hampered by uncertainties in model parameters, such as initial and boundary conditions, and the complexities of coastal topography. Traditional data assimilation techniques, which aim to improve model predictions by incorporating observational data, can be computationally expensive and may struggle with the high resolution often required in specific areas of interest (e.g., near detectors or complex features). This paper tackles the challenge of improving coastal flow simulations by developing an adaptive mesh adjoint model that can dynamically adjust mesh resolution and anisotropy to efficiently capture both the flow dynamics and the observational data.

Methods

The core of this work is the development of an adaptive mesh adjoint model. Here's a breakdown of the key methods:

  • Adjoint Model Derivation: The authors derive a continuous adjoint model from the continuous forward model (based on Navier-Stokes equations). This continuous adjoint model is then discretized to obtain the discrete adjoint model.
  • Mesh Adaptivity: A significant innovation is the use of mesh adaptivity in both the transient forward and adjoint models. This means the mesh (the grid used for calculations) can dynamically change its resolution and shape over time and space to better represent local flow features and observational data locations.
  • Inconsistent Discrete Schemes: A key aspect is that the discrete adjoint model is not necessarily consistent with the discrete forward model. This allows for greater flexibility, enabling the forward and adjoint models to use different, optimized meshes. For instance, the forward model might need high resolution near a headland, while the adjoint model might prioritize resolution near observation points.
  • Nonlinear Conjugate Gradient Algorithm: The optimization process, which involves adjusting control variables (like boundary conditions) to minimize the difference between model predictions and observations, uses a nonlinear conjugate gradient algorithm.
  • Interface Module: A crucial component is an interface module that handles communication between the forward and adjoint models. This module manages the control variables, interpolates information between potentially different meshes, and ensures that the adjoint solution can be used to update the forward model's parameters.

Datasets

Not discussed in the paper. The study uses synthetic pseudo-observations generated by running the forward model with known "exact" boundary conditions.

Key Insights and Ideas

  • Flexibility Through Inconsistency: The paper demonstrates that sacrificing strict consistency between the discrete forward and adjoint models can be advantageous. This allows for independent mesh adaptivity, leading to more efficient and accurate simulations by tailoring the mesh to the specific needs of each model component.
  • Adaptive Meshes for Coastal Flows: Mesh adaptivity is shown to be highly effective in capturing complex flow features, such as local eddies near a headland, which are missed by fixed-mesh simulations.
  • Efficient Inversion: The adaptive mesh adjoint approach significantly improves the efficiency of inverting for uncertain parameters, like boundary conditions, by focusing computational effort where it's most needed.
  • Practical Implementation: The development of an interface module is critical for making adaptive mesh adjoint methods practical, enabling the transfer of information between dynamically changing meshes.

Key Results

  • The adaptive mesh adjoint model successfully inverted for inflow boundary conditions for a flow past a headland.
  • After three nonlinear conjugate gradient iterations, the maximum misfit between inverted and expected boundary conditions was reduced from approximately 1.0 m/s to 0.2 m/s.
  • The adapted meshes for both forward and adjoint models conformed well to local flow features, with the adjoint mesh showing refinement near simulated observation points.
  • Local eddies near the headland were only simulated when mesh adaptivity was employed in the forward model.

Limitations and Future Work

  • Inconsistency: While beneficial for flexibility, the inconsistency between discrete schemes means the adjoint model is not a perfect time-reversed representation of the forward model. The authors acknowledge this as a "route cause of the inconsistency."
  • Smoothing: The use of linear interpolation between meshes introduces some smoothing, which is deemed negligible in this study because adaptivity is not invoked at every time step. However, the authors suggest that more sophisticated higher-order shape-preserving interpolation algorithms could further minimize this.
  • Future Work: The paper highlights the need for further investigation into the full advantages and efficiency gains of this approach compared to consistent adjoint fixed-mesh methods. Exploring more sophisticated interpolation techniques is also suggested.

Summary updated

Abstract

An adaptive mesh adjoint model has been developed to invert for uncertain parameters in the simulation of coastal flows. These may include, for example, open boundary conditions and initial conditions. Our objective in this paper is to obtain a better representation of water flows around and over coastal topography. In this paper a first attempt has been made at using mesh adaptivity in both transient forward and adjoint models. Although the resulting method has inconsistent discrete schemes for the forward and adjoint models, the inversion results shown here are nevertheless good. The derivation of the continuous adjoint model is presented. A description of the adaptive mesh adjoint technique is also given. As a study case, the method is applied to the inversion of boundary conditions for flow past a headland.
BibTeX
@misc{fang2004adjoint,
  author = {F. Fang and Christopher C. Pain and Matthew D. Piggott and Gerard Gorman and Graham M. Copeland and C. Mendes de Oliveira and A.J.H. Goddard and Igor Gejadze},
  title = {{Adjoint Data Assimilation into a 3D Unstructured Mesh Coastal Finite Element Model}},
  year = {2004},
  doi = {10.1061/40734(145)20},
}
Figure