triflow 0.5.2

Automatic finite difference discretization for 1D PDE with fast temporal solvers.

Triflow

Master	Dev

Installation

External requirements

This library is written for python >= 3.5.

The library is based on Theano, thus extra dependencies like fortran and C compiler are needed, see Theano install page for extra information:

http://deeplearning.net/software/theano/install.html

On v0.5.0, it is possible to choose between theano and numpy (which provide similar features). numpy will be slower but with no compilation time, which is handy for testing and prototyping.

via PyPI

pip install triflow

will install the package and

pip install triflow --upgrade

will update an old version of the library.

use sudo if needed, and the user flag if you want to install it without the root privileges:

pip install --user triflow

via github

You can install the last version of the library using pip and the github repository:

pip install git+git://github.com/locie/triflow.git

via anaconda (strongly advised for windows users)

conda install -c conda-forge -c celliern triflow

Introduction

Motivation

The aim of this library is to have a (relatively) easy way to write transient dynamic systems with 1D finite difference discretization, with fast temporal solvers.

The main two parts of the library are:

• symbolic tools defining the spatial discretization, with boundary taking into account in a separated part
• a fast temporal solver written in order to use the sparsity of the finite difference method to reduce the memory and CPU usage during the solving

Moreover, extra tools are provided and the library is written in a modular way, allowing an easy extension of these different parts (see the plug-in module of the library.)

The library fits well with an interactive usage (in a jupyter notebook). The dependency list is actually larger, but on-going work target a reduction of the stack complexity.

Model writing

All the models are written as function generating the F vector and the Jacobian matrix of the model defined as dtU = F(U).

The symbolic model is written as a simple mathematic equation. For example, a diffusion advection model can be written as:

from triflow import Model

equation_diff = "k * dxxU - c * dxU"
dependent_var = "U"
physical_parameters = ["k", "c"]

model = Model(equation_diff, dependent_var, physical_parameters)

Example

import numpy as np
import pylab as pl
from triflow import Model, Simulation

model = Model("k * dxxU - c * dxU",
              "U", ["k", "c"])

x, dx = np.linspace(0, 1, 200, retstep=True)
U = np.cos(2 * np.pi * x * 5)
fields = model.fields_template(x=x, U=U)

parameters = dict(c=.03, k=.001, dx=dx, periodic=False)

t = 0
dt = 5E-1
tmax = 2.5

pl.plot(fields.x, fields.U, label=f't: {t:g}')


def dirichlet_condition(t, fields, pars):
    fields.U[0] = 1
    fields.U[-1] = 0
    return fields, pars


simul = Simulation(model, t, fields, parameters, dt,
                   hook=dirichlet_condition, tmax=tmax)

for i, (t, fields) in enumerate(simul):
    print(f"iteration: {i}\t",
          f"t: {t:g}", end='\r')
    pl.plot(fields.x, fields.U, label=f't: {t:g}')

pl.xlim(0, 1)
legend = pl.legend(loc='best')

pl.show()

NEWS

v0.5.0:

• WARNING: some part of the API has changed:
  • Simulation signature has changed. t arg is now optional (with t=0) as default and physical_parameters is now parameters.
  • The displays have been completely rewritten, and the previous API is depreciated. Users are encouraged to modify their scripts or to stick to the ^0.4 triflow versions.
• move schemes from plugins to core
• compilers: remove tensorflow, add numpy which is way slower but has no compilation overhead.
• displays and containers are connected to the simulation via streamz
• add post-processing.
• real-time display is now based on Holoviews. Backward compatibility for display is broken and users are encouraged to modify their scripts or to stick to the ^0.4 triflow versions.
• use poetry to manage dependencies.
• use tqdm to display simulation update.

v0.4.12:

• give user choice of compiler
  • get out tensorflow compiler (not really efficient, lot of maintenance trouble)
  • give access to theano and numpy compiler
• upwind scheme support
• using xarray as fields backend, allowing easy post process and save
• update display and containers
• adding repr string to all major classes

v0.4.7:

• adding tensor flow support with full testing
• adding post-processing in bokeh fields display

ROADMAP / TODO LIST

The following items are linked to a better use of solid external libs:

• change all the display and container workflow:
  • use streamz to allow pipeline like way to add display / probing / post-process :done:
  • use holoviews as main way to do real-time plotting :done:
  • use xarray multi netcdf files to reduce IO lack of performance :done:
• better use of external solving lib:
  • merge triflow.plugins.schemes and scipy.integrate.OdeSolver API :todo:
  • use scipy.integrate.solve_ivp for triflow temporal scheme solving (making it more robust) :todo:
  • main goal is to have better two-way integration with scipy :todo:

These are linked to the triflow core

• build a robust boundary condition API :todo:
• work on dimension extension, allowing 2D resolution and more :todo:
• allow auxiliary function to make some complex model easier to write :todo:
• allow a choice on the finite difference scheme, on a global way or term by term :todo:
• test and propose other compilers (Cython, numba, pythran?) :todo:
• work on adaptive spatial and temporal mesh :todo:

These are far away but can be very interesting:

• implement continuation algorithm working with triflow (separate project?)
• try other kind of discretization scheme (separate project each?)
  • Finite volume
  • Finite element?

The final (and very ambitious goal) is to provide a robust framework allowing to link the mathematical language (and a natural way to write the model, as natural as possible) with a high performance and robust solving of the numerical system. There is a trade-off between the ease of use and the performance, for this software, all mandatory dependencies have to be pip-installable (as opposite to dedalus project or fenics project), even if some non-mandatory dependencies can be harder to install.

If we go that further, it may be interesting to split the project with the triflow language, the different spatial discretization and so on...

License

This project is licensed under the term of the MIT license