semba-dgtd

Maxwell curl-equation solver using discontinuous Galerkin methods (OpenSEMBA / UGR).

Repository layout

Path	Contents
src/	Driver, evolution operators, DG components, MFEM extensions
external/mfem-geg/	Required MFEM fork (submodule)
testData/maxwellInputs/	Example simulation cases (JSON + mesh per folder)
testData/rcsInputs/	Offline opensemba_rcs JSON (frequency/angle sweeps)
test/	Unit and integration tests (GoogleTest)
docs/	Input format, tools, and feature usage notes
pythonBindings/	Optional Python bindings

Compiling

Requirements: CMake ≥ 3.25.2, vcpkg ($VCPKG_ROOT), Ninja, GCC.

vcpkg installs (see vcpkg.json): eigen3, gtest, fftw3, nlohmann-json.

MPI builds

Requires METIS 5 and HYPRE built from source; set METIS_DIR and HYPRE_DIR before configuring.

METIS:

wget https://github.com/mfem/tpls/raw/gh-pages/metis-5.1.0.tar.gz
tar -zxvf metis-5.1.0.tar.gz
cd metis-5.1.0
make BUILDDIR=lib config
make BUILDDIR=lib
cp lib/libmetis/libmetis.a lib/
export METIS_DIR=$PWD

HYPRE:

wget https://github.com/hypre-space/hypre/archive/refs/tags/v2.31.0.tar.gz
tar -zxvf v2.31.0.tar.gz
cd hypre-2.31.0/src
./configure --disable-fortran
make -j $(nproc)
export HYPRE_DIR=$PWD/hypre

CUDA builds

Requires HYPRE built with CUDA (CMake, not autoconf ./configure).

Prerequisites: For RTX 50-series (sm_120), build MFEM/dgtd with CUDA ≥ 13.x (CUDA 12.8’s cicc can OOM on MFEM kernels; see mfem#5363). Set -DCMAKE_CUDA_ARCHITECTURES to your GPU (e.g. 120, 89, 86) and match CMakePresets.json.

HYPRE with CUDA — stay on HYPRE 2.31 (matches stock mfem-geg; no HYPRE-3 API patches). Build HYPRE with CUDA 12.8 if needed (2.31 does not build against CUDA 13 CCCL); link that install into a dgtd/MFEM build that uses CUDA 13.3 for sm_120:

# Example: HYPRE 2.31 + CUDA 12.8, arch matched to the GPU
cmake -S hypre-2.31.0/src -B hypre-cuda-build \
      -DHYPRE_WITH_CUDA=ON \
      -DHYPRE_CUDA_SM=120 \
      -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.8/bin/nvcc \
      -DCMAKE_INSTALL_PREFIX=$HOME/workspace/hypre-cuda-install
cmake --build hypre-cuda-build -j $(nproc)
cmake --install hypre-cuda-build
export HYPRE_CUDA_DIR=$HOME/workspace/hypre-cuda-install

Then (with METIS_DIR set; CUDA 13.3 for the dgtd/MFEM compile via the preset):

export PATH=/usr/local/cuda-13.3/bin:$PATH
cmake --preset gnu-release-cuda
cmake --build --preset build-gnu-release-cuda --parallel

On sm_120, the top-level CMake compiles a small set of heavy MFEM CUDA TUs at -O0 to keep cicc compile times reasonable.

CMake presets

Preset	Description
gnu-debug-mpi	Debug, MPI, OpenMP
gnu-release-mpi	Release, MPI, OpenMP
gnu-debug-cuda	Debug, MPI, OpenMP, CUDA (gcc-12)
gnu-release-cuda	Release, MPI, OpenMP, CUDA (gcc-12)

cmake --preset gnu-release-mpi
cmake --build --preset build-gnu-release-mpi --parallel

MFEM

Built automatically from external/mfem-geg. For an external install: -DSEMBA_DGTD_ENABLE_MFEM_AS_SUBDIRECTORY=OFF.

Warning: Use only OpenSEMBA/mfem-geg. Upstream MFEM will not build this project.

Running a case

Cases live under testData/maxwellInputs/<case_name>/ with matching <case_name>.json and mesh file.

Full JSON reference: docs/json-input-format.md.

Example:

./build/gnu-release-mpi/bin/opensemba_dgtd testData/maxwellInputs/1D_PEC/1D_PEC.json

(Confirm binary name/path for your preset.)

Exports appear under exports/ by kind, then run mode and case name:

• Simulation probes / stats / RCS dumps: exports/SimulationData/<runmode>/<casename>/
• ParaView exporter: exports/ParaView/<runmode>/
• Sparse operators: exports/Operators/<casename>/

RCS post-processing (opensemba_rcs)

Offline far-field / RCS from an existing rcssurface export. The Maxwell case must already have been run with a probes.rcssurface probe so that

exports/SimulationData/<runmode>/<casename>/RCSSurface/<probe_name>/rank*/surface_data.bin

exists. The post-processor also reads the case JSON at testData/maxwellInputs/<casename>/<casename>.json (plane-wave data and probe names).

RCS input files live under testData/rcsInputs/ (often *.rcs.json). Full parameter reference: docs/json-input-format.md.

Field	Required	Description
runmode	yes	Export folder segment (e.g. cuda-1, single-core, mpi-8)
casename	yes	Must match testData/maxwellInputs/<casename>/ and the export tree
frequencies	yes	{ start, end, steps } in Hz (steps = linspace count)
angles	yes	theta / phi each { start, end, steps } in radians
max_time	no	Keep snapshots with time ≤ this value
every_n_steps	no	Subsample while reading (default 1; not equivalent to full history)
ram_gate	no	Hard RAM budget in GiB; omit → ~50% of MemAvailable. Over budget → streaming DFT

Example:

mpiexec -n 1 ./build/gnu-release-mpi/bin/opensemba_rcs \
  -i testData/rcsInputs/3D_Nasa_Almond_G2_25cm_5GHz.rcs.json

Results are written under the same probe directory as farfield/ and rcs/.

Further documentation

Topic	Link
All docs	docs/README.md
JSON input (solver + offline RCS)	docs/json-input-format.md
MOR → ParaView	docs/mor2paraview.md
Box PML (mesh, JSON, absorption table)	docs/designing-a-box-pml.md

Funding

• Spanish Ministry of Science and Innovation (MICIN/AEI) (Grant PID2022-137495OB-C31).
• European Union, HECATE project (HE-HORIZON-JU-Clean-Aviation-2022-01).
• European Union, FEDER 2020 (B-TIC-700-UGR20).