1D Shallow water

Numerical code modelling the 1D shallow water equations to build a toy problem of robust estimation of the bottom friction. Quadratic cost function is implemented between a reference simulation (twin experiments setting) and the simulation performed with given parameters. The gradient with respect to the bottom friction of this cost function is obtained via the adjoint model. In the module \HR_config\, wrapper.py contains readily available cost function(s)

The code is a bit messy, and could use a good refactoring

Installation

This project is not on the python package index (PyPI), hence has to be cloned from this git repository (root access may be needed to install):

git clone [email protected]:vtrappler/SWE1D.git
pip install -r requirements.txt

Examples of utilisation

The file example_forward_tlm.ipynb may be more up to date with the code. More generally, I think the HR_config folder may not be useful anymore

May be obsolete below

The file example.ipynb is a notebook containing a few examples on how to use the modules

Animation of the evolution of the sea surface heights

import HR_config.wrapper as swe
import code_swe.animation_SWE

[xr, h, u, t] = swe.swe_KAP(swe.Kref * 2, swe.amplitude - 1.0, swe.period + 1.0)
code_swe.animation_SWE.animate_SWE(xr, [swe.href, h], swe.b, swe.D, ylim = [0,10])

Evaluation of the cost function (where swe.href is the simulation reference), and verification of the gradient

import HR_config.wrapper as swe

swe.J_KAP(swe.Kref, swe.amplitude, swe.period)
cost0, gradient0 = swe.J_KAP([0], swe.amplitude, swe.period)
epsilon = 1e-8
cost_eps = swe.J_KAP_nograd([epsilon], swe.amplitude, swe.period)
gradient_finite_diff = (cost_eps - cost0) / epsilon
print gradient_finite_diff, gradient0

Evaluation of the cost function parallelized with gradient

J_KAP_array is readily implemented, taking advantage of the Multiprocessing module of python in order to parallelize the computations. The function takes as input an array of tuples, each one in the following format: (Coeff_K, Amplitude, Period), where Coeff_K is an array that will produce the piecewise constant interpolation on all the grid points, and Amplitude and Period are scalar that parametrize the left boundary condition.

import HR_config.wrapper as swe

response, gradient = swe.J_KAP_array([([0.1, 0.2, 0.5], 5.0, 15.0),   # Example array to evaluate
                                      ([0.1, 0.2, 0.5], 5.1, 15.1),   # Dim K = 3
                                      ([0.1, 0.1, 0.1], 5.0, 15.2),
                                      ([0.1, 0.1, 0.4], 5.1, 15.0),
                                      ([0.2, 0.2, 0.5], 5.0, 15.1),
                                      ([0.2, 0.2, 0.5], 5.1, 15.2),
                                      ([0.6, 0.1, 0.7], 5.0, 15.0),
                                      ([0.2, 0.2, 0.5], 5.1, 15.1),
                                      ([0.2, 0.2, 0.5], 5.0, 15.2),
                                      ([0.1, 0.7, 0.5], 5.1, 15.0),
                                      ([0.2, 0.2, 0.2], 5.0, 15.1)],
                                      idx_to_observe = None,
                                      hreference = swe.href,
                                      parallel=True, ncores=4,
                                      adj_gradient=True)

Technical details

Numerical scheme

Numerical solution computed via finite volume. Adjoint code has been derived for reflexive boundary on the right, and Lax-Friedrich's flux inbetween the volumes.

Boundary condition on the left

The boundary condition on the left is parametrized as following:

$$h(0,t) = \texttt{mean.h} + \texttt{amplitude} \cdot \sin\left(t \frac{2\pi}{\texttt{period}} + \texttt{phase}\right)$$