Adjoint inverse reconstruction¶
Introduction¶
In this tutorial, we will demonstrate how to perform an inversion to recover the initial temperature field of an idealised mantle convection simulation using G-ADOPT. This tutorial is published as the first synthetic experiment in Ghelichkhan et al. (2024). The full inversion showcased in the publication involves a total number of 80 timesteps. For the tutorial here we start with only 5 timesteps to go through the basics.
The tutorial involves a twin experiment, where we assess the performance of the inversion scheme by inverting the initial state of a synthetic reference simulation, known as the "Reference Twin". To create this reference twin, we run a forward mantle convection simulation and record all relevant fields (velocity and temperature) at each time step.
We have pre-run this simulation by running the forward case, and stored model output as a checkpoint file on our servers. These fields serve as benchmarks for evaluating our inverse problem's performance. To download the reference benchmark checkpoint file if it doesn't already exist, execute the following command:
![ ! -f adjoint-demo-checkpoint-state.h5 ] && wget https://data.gadopt.org/demos/adjoint-demo-checkpoint-state.h5
In this file, fields from the reference simulation are stored under the names "Temperature" and "Velocity". After importing g-adopt and the associated inverse module (gadopt.inverse - discussed further below), we can retrieve timestepping information from the pre-computed forward run as follows
from gadopt import *
from gadopt.inverse import *
# Open the checkpoint file and subsequently load the mesh:
checkpoint_filename = "adjoint-demo-checkpoint-state.h5"
checkpoint_file = CheckpointFile(checkpoint_filename, mode="r")
mesh = checkpoint_file.load_mesh("firedrake_default_extruded")
mesh.cartesian = True
# Specify boundary markers, noting that for extruded meshes the upper and lower boundaries are tagged as
# "top" and "bottom" respectively.
bottom_id, top_id, left_id, right_id = "bottom", "top", 1, 2
# Retrieve the timestepping information for the Velocity and Temperature functions from checkpoint file:
temperature_timestepping_info = checkpoint_file.get_timestepping_history(mesh, "Temperature")
velocity_timestepping_info = checkpoint_file.get_timestepping_history(mesh, "Velocity")
We can check the information for each:
print("Timestepping info for Temperature", temperature_timestepping_info)
print("Timestepping info for Velocity", velocity_timestepping_info)
Timestepping info for Temperature {'index': array([ 0., 79.])}
Timestepping info for Velocity {'index': array([ 0., 1., 2., 3., 4., 5., 6., 7., 8., 9., 10., 11., 12.,
13., 14., 15., 16., 17., 18., 19., 20., 21., 22., 23., 24., 25.,
26., 27., 28., 29., 30., 31., 32., 33., 34., 35., 36., 37., 38.,
39., 40., 41., 42., 43., 44., 45., 46., 47., 48., 49., 50., 51.,
52., 53., 54., 55., 56., 57., 58., 59., 60., 61., 62., 63., 64.,
65., 66., 67., 68., 69., 70., 71., 72., 73., 74., 75., 76., 77.,
78., 79.]), 'delta_t': array([4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06,
4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06, 4.e-06])}
The timestepping information reveals that there are 80 time-steps (from 0 to 79) in the reference simulation, with the temperature field stored only at the initial (index=0) and final (index=79) timesteps, while the velocity field is stored at all timesteps. We can visualise the benchmark fields using Firedrake's built-in VTK functionality. For example, initial and final temperature fields can be loaded:
# Load the final state, analagous to the present-day "observed" state:
Tobs = checkpoint_file.load_function(mesh, "Temperature", idx=int(temperature_timestepping_info["index"][-1]))
Tobs.rename("Observed Temperature")
# Load the reference initial state - i.e. the state that we wish to recover:
Tic_ref = checkpoint_file.load_function(mesh, "Temperature", idx=int(temperature_timestepping_info["index"][0]))
Tic_ref.rename("Reference Initial Temperature")
checkpoint_file.close()
These fields can be visualised using standard VTK software, such as Paraview or pyvista.
import pyvista as pv
VTKFile("./visualisation_vtk.pvd").write(Tobs, Tic_ref)
dataset = pv.read('./visualisation_vtk.pvd')
# Create a plotter object
plotter = pv.Plotter()
# Add the dataset to the plotter
plotter.add_mesh(dataset, scalars='Observed Temperature', cmap='coolwarm')
# Adjust the camera position
plotter.camera_position = [(0.5, 0.5, 2.5), (0.5, 0.5, 0), (0, 1, 0)]
# Show the plot
plotter.show(jupyter_backend="static")