Introduction

This is the code base for our paper Amortized Finite Element Analysis for Fast PDE-Constrained Optimization on ICML 2020.

Requirements

• python 3.7.3
• numpy
• scipy
• matplotlib
• fenics
• dolfin-adjoint
• mshr
• pytorch

We recommend using conda

conda install python==3.6.3
conda install numpy scipy matplotlib
conda install -c conda-forge fenics
conda install -c conda-forge dolfin-adjoint
conda install -c conda-forge mshr
conda install pytorch torchvision cudatoolkit=$YOUR_TOOLKIT_VERSION -c pytorch

Usage

To run a runner module, do:

python -m src.ml.trainer_dolfin

or similar. The package currently contains three examples (linear, dolfin and robot) and module names should be indicative.

Descriptions

The following descriptions use dolfin example for instructions.

To compute traditional finite element analysis (FEA) solutions, do:

python -m src.pde.poisson_dolfin

To draw i.i.d. data for control parameters, do:

python -m src.ml.generator

To train a neural network model using amortized finite element analysis (AmorFEA), do:

python -m src.ml.trainer_dolfin

To perform PDE-constrained optimization with AmorFEA enabled surrogate model, do:

python -m src.opt.optimizer_dolfin

To visualize results, do:

python -m src.vis.demo_dolfin

Dev

To define hyperparameters / arguments / defaults for a module, create a file src/arguments/MODULENAME_args.py, with a method add_args(parser) which adds argparse arguments to the parser.

To access arguments for all modules:

from src import arguments
args = arguments.args

Modules use relative imports (e.g. to import arguments from a module in src/, from . import arguments, to do it from a module in src/pde/, from .. import arguments, ...).

Generally follow PEP8.