opt f(x)

Note (updated May 2024): This repository is not currently maintained. An updated version is continued to pytop.

Motivation and significance

optfx is a library for the large scale field optimization platform based on FEniCS ecosystem. This module enables reusable and straightforward UFL coding for inverse problem of multi-physics.

This software is the third-party module of fenics project.

This software is based on Lagacy FEniCS (FEniCS2019.1.0). The new version, FEniCSx, is not supported.

This README and documentation is now under construction.

Getting started

import numpy as np
import optfx as opt

opt.parameters["form_compiler"]["optimize"] = True
opt.parameters["form_compiler"]["cpp_optimize"] = True
opt.parameters['form_compiler']['quadrature_degree'] = 5

comm = opt.MPI.comm_world
m = 0.30   # Target rate of the material amount
p = 5      # Penalty parameter
eps = 1e-3 # Material lower bound
R = 0.01   # Helmholtz filter radius
n = 256    # Resolution

def k(a):
    return eps + (1 - eps) * a ** p

mesh = opt.UnitSquareMesh(comm, n, n)
X = opt.FunctionSpace(mesh, 'CG', 1)
f = opt.interpolate(opt.Constant(1e-2), X)

class Initial_density(opt.UserExpression):
    def eval(self, value, x):
        value[0] = m

class Left(opt.SubDomain):
    def inside(self, x, on_boundary):
        gamma = 1/250 + 1e-5
        return x[0] == 0.0 and 0.5 - gamma < x[1] < 0.5 + gamma and on_boundary

U = opt.FunctionSpace(mesh, 'CG', 1)
t = opt.TrialFunction(U)
dt = opt.TestFunction(U)
bc = opt.DirichletBC(U, opt.Constant(0.0), Left())

class PoissonProblem(opt.Module):
    def problem(self, controls):
        rho = controls[0]
        rho = opt.helmholtzFilter(rho, X, R)
        a = opt.inner(opt.grad(t), k(rho)*opt.grad(dt))*opt.dx
        L = opt.inner(f, dt)*opt.dx
        Th = opt.Function(U, name='Temperture')
        opt.solve(a==L, Th, bc)
        J = opt.assemble(opt.inner(opt.grad(Th), k(rho)*opt.grad(Th))*opt.dx)
        rho_bulk = opt.project(opt.Constant(1.0), X)
        rho_0 = opt.assemble(rho_bulk*opt.dx)
        rho_total = opt.assemble(controls[0]*opt.dx)
        rel = rho_total/rho_0
        self.volumeFraction = rel
        return J
    def constraint_volume(self):
        return self.volumeFraction - m

x0 = opt.Function(X)
x0.interpolate(Initial_density())
N = opt.Function(X).vector().size()
min_bounds = np.zeros(N)
max_bounds = np.ones(N)

setting = {'set_lower_bounds': min_bounds,
           'set_upper_bounds': max_bounds,
           'set_maxeval': 1000
          }
params = {'verbosity': 1}

problem = PoissonProblem()
solution = opt.optimize(problem, [x0], {'constraint_volume': [0]}, setting, params)

Installation

Docker

We recomennd to use our docker image. First you should install docker.io or relavant systems. Our docker image was stored in the dockerhub.

If you use the docker.io and docker-compose, the following command pull and run the above image.

docker-compose up

install on your local host

First, make sure to install the following dependencies:

FEniCS2019.1.0 + pyadjoint
nlopt

Second, install using pip.

pip install git+https://github.com/Naruki-Ichihara/fenics_optimize.git@main

Contributors

Naruki-Ichihara

API and documentation

API: API Documentation: Docs

References

Cite

To cite this repository:

@software{fenics_optimize2023github,
  author = {Naruki Ichihara},
  title = {{optfx}: Large scale field optimization platform based on fenics in python},
  url = {https://github.com/Naruki-Ichihara/fenics_optimize},
  version = {0.3.0},
  year = {2023},
}

Naruki-Ichihara / fenics_optimize

readme