EMopt

A toolkit for shape (and topology) optimization of 2D and 3D electromagnetic structures.

EMopt offers a suite of tools for simulating and optimizing electromagnetic structures. It includes 2D and 3D finite difference frequency domain solvers, 1D and 2D mode solvers, a flexible and easily extensible adjoint method implementation, and a simple wrapper around scipy.minimize. Out of the box, it provides just about everything needed to apply cutting-edge inverse design techniques to your electromagnetic devices.

A key emphasis of EMopt's is shape optimization. Using boundary smoothing techniques, EMopt allows you to compute sensitivities (i.e. gradient of a figure of merit with respect to design variables which define an electromagnetic device's shape) with very high accuracy. This allows you to easily take adavantage of powerful minimization techniques in order to optimize your electromagnetic device.

Documentation

Details on how to install and use EMopt can be found on readthedocs. Check this link periodically as the documentation is constantly being improved and examples added.

Installation

Requirements: Ubuntu 22.04 / WSL2 Linux, sudo access, ~5 GB disk, ~30 min first run.

1. System dependencies + PETSc/SLEPc (run once)

bash setup-system-deps.sh

This installs g++, gfortran, OpenMPI, Eigen, Boost, and Poppler via apt, then builds PETSc (with complex arithmetic) and SLEPc from source into ~/.emopt.

2. Python environment (uv)

bash setup-python.sh

Creates .venv/ (Python 3.13), installs mpi4py, petsc4py, slepc4py, CPU-only PyTorch, pandas, pyyaml, pdftotext, and emopt itself, then builds the native C++ extensions.

Useful overrides:

bash setup-python.sh --no-torch
bash setup-python.sh --torch-gpu

Validated working combination on this branch:

• Python 3.13
• PETSc 3.21.5
• SLEPc 3.21.2
• petsc4py==3.21.5
• slepc4py==3.21.2
• setuptools<70
• Cython==3.0.10

3. Activate

source .venv/bin/activate

To recreate the Python environment later (without rebuilding PETSc):

source ~/.emopt_deps && bash setup-python.sh

Using EMopt From Another Repo

If you want to use emopt from a separate uv-managed project, treat this repo as the system-dependency provider and install emopt into the other project's virtual environment.

1. On a fresh machine, prepare EMopt system dependencies once

git clone <your-emopt-fork-url>
cd emopt
bash setup-system-deps.sh

This installs PETSc and SLEPc into ~/.emopt and writes the environment file ~/.emopt_deps.

2. In your other repo, create a uv environment

cd /path/to/your-other-repo
uv venv --python 3.13
source .venv/bin/activate
source ~/.emopt_deps

3. Install the PETSc Python bindings into that environment

uv pip install "setuptools<70" wheel "Cython==3.0.10"
uv pip install numpy scipy matplotlib h5py pyyaml pandas requests future pdftotext
uv pip install mpi4py --no-build-isolation
uv pip install "petsc4py==3.21.5" --no-build-isolation --no-deps
uv pip install "slepc4py==3.21.2" --no-build-isolation --no-deps

4. Install emopt from your cloned checkout

uv pip install -e /path/to/emopt --no-build-isolation

Or, from the active environment in the other repo, run the helper shipped in this repo:

/path/to/emopt/scripts/install-into-active-venv.sh

If you change native code in emopt/src, rebuild from the emopt checkout:

cd /path/to/emopt
source ~/.emopt_deps
make

5. Verify from the other repo

python - <<'PY'
import emopt
from petsc4py import PETSc
print("emopt ok")
print("PETSc scalar type:", PETSc.ScalarType)
PY

Notes:

• source ~/.emopt_deps must be active whenever you build or reinstall petsc4py, slepc4py, or emopt.
• The other repo does not need to vendor PETSc or SLEPc itself; it can reuse the shared ~/.emopt installation created by setup-system-deps.sh.
• Experimental emopt.experimental workflows additionally require PyTorch.

Free-Form Topology and AutoDiff-Enhanced Feature-Mapping Approaches

New optional experimental modules for topology optimization and automatic differentiation enhanced feature-mapping approaches are implemented in emopt/experimental, with corresponding examples in examples/experimental. The AutoDiff methods can result in large improvements in optimization speed for designs with variables that parameterize global geometric features. Please see our preprint below and examples for correct usage. Note: Requires PyTorch installation. These features are still in development.

Authors

Andrew Michaels

Sean Hooten (Topology and AutoDiff methods)

License

EMOpt is currently released under the BSD-3 license (see LICENSE.md for details)

References

The methods employed by EMopt are described in:

Andrew Michaels and Eli Yablonovitch, "Leveraging continuous material averaging for inverse electromagnetic design," Opt. Express 26, 31717-31737 (2018)

An example of applying these methods to real design problems can be found in:

Andrew Michaels and Eli Yablonovitch, "Inverse design of near unity efficiency perfectly vertical grating couplers," Opt. Express 26, 4766-4779 (2018)

Shape optimization feature-mapping methods accelerated by automatic differentiation:

S. Hooten, P. Sun, L. Gantz, M. Fiorentino, R. Beausoleil, T. Van Vaerenbergh, "Automatic Differentiation Accelerated Shape Optimization Approaches to Photonic Inverse Design on Rectilinear Simulation Grids." arXiv [cs.CE], 2311.05646 (2023). Link here.