FES — Finite Element Solver

FES is a 2D/3D finite element solver for computational electromagnetics based on the E-field formulation with H(curl) conforming elements. It supports S-parameter extraction, eigenmode analysis, electrostatics, and nonlinear Kerr materials.

The solver has three language backends sharing the same model files:

Backend	Dir	Description
C++	cpp/	Production 3D solver — curl-curl, MUMPS direct + GMRES iterative, DD, waveports
Python	py/	2D solver + DNN-GP surrogate modelling — scalar Helmholtz, HB, ML surrogates
MATLAB	m/	Reference / legacy implementation — full-featured, research-oriented

---

Quick Start

C++ backend (primary)

./setup # install dependencies (compiled from source)
make build # cmake configure + build (Release)
make test# run all model tests (load & mesh only)

The binary is at cpp/build/fes:

# 3D waveguide — mesh with TetGen, then solvecd data && ../cpp/build/fes WR90 +poly AafeeQ +f 1e10 +p 2
# 2D TMz filter — mesh with Triangle, then solvecd data && ../cpp/build/fes BilatFilter +poly q34a +f 150e9 +p 2
# Waveport eigenmodescd data && ../cpp/build/fes WR90 +poly AafeeQ +f 1e10 +p 2 +formula em_e_tl_eig
# Electrostatic (requires voltage assignment)cd data && ../cpp/build/fes CapSense +poly +f 0 +volt Elec 1

Python backend

./setup --py # or: cd py && ./configure
make py-test # or: py/.venv/bin/python -m pytest py/tests/ -v

# Run waveguide simulationcd py && .venv/bin/python -c "from fes.projects import run_waveguide; run_waveguide()"# Train DNN-GP surrogate modelcd py && .venv/bin/python -c "from fes.projects import bilateral_filter_dnngp; bilateral_filter_dnngp()"

MATLAB backend

./setup --m # or: cd m && ./configure
make m-build # cd m && make all (builds IOrMesh mesher)# make m-test # runs MATLAB/Octave tests

% In MATLAB/Octave:
addpath(genpath('m'));
ProjectWaveGuide;

---

Formulations (C++ CLI)

Formulations are auto-detected from the #Formula tag in each .poly file. CLI flags (+formula em_e_fd, +em_e_fd, etc.) override the file tag.

CLI flag	Description
+formula em_e_fd or +em_e_fd	3D frequency-domain EM (curl-curl, default)
+formula em_ez_fd or +em_ez_fd	2D TMz (scalar Helmholtz on Triangle mesh)
+formula em_e_tl_eig or +em_e_tl_eig	2D cross-section eigenmode or 3D waveport eigenmodes
+formula em_e_qs or +em_e_qs	Electrostatic quasistatic (f = 0, use +volt bnd V)

CLI Options (C++)

Flag	Default	Description
	Model I/O
+poly [CMD]	—	Import .poly file (auto-detects 2D/3D; optional TetGen quality switches)
	Formulation
+formula NAME	—	Select formulation by snake_case name (em_e_fd, em_ez_fd, etc.)
+em_e_fd	—	3D frequency-domain EM (curl-curl)
+em_e_fd_dd N	—	3D EM with domain decomposition
+em_e_fd_nl	—	3D EM with nonlinear Kerr
+em_e_tl_eig	—	Waveport eigenmodes / 2D cross-section eigenmode
+em_ez_fd	—	2D TMz (scalar Helmholtz)
+em_e_qs	—	Electrostatic quasistatic
+em_h_qs	—	Magnetic quasistatic
	Frequency
+f FREQ	—	Main frequency [Hz] (required; 0 = electrostatic)
	Mesh & discretization
+p N	1	Polynomial order (1–4)
+h N	0	Homogeneous mesh refinement level
+href CMD	—	Quality mesh refinement with TetGen
	Output
+field	—	Export VTK field data
+rad Nθ Nφ	—	Export far-field radiation pattern
+sparam	on	Write S-parameters (-sparam to disable)
	Ports & excitation
+tfe	on	Transfinite element formulation on waveports (-tfe to disable)
+volt bnd V	—	Apply voltage V to PEC boundary (electrostatic)
+pow P	1	Port power scaling [W]
+einc LABEL = {Ex,Ey,Ez,kx,ky,kz}	—	Incident plane wave (disables S-params)
	Frequency sweep
+fr lf hf n	—	Discrete frequency sweep over n points
	Solver
+direct	default	MUMPS direct sparse solver
+gmres tol [restart]	—	GMRES iterative solver
+sgl	—	Single precision (default: double)
+dbl	—	Double precision (explicit)
+dbg	—	Debug output
	Domain decomposition
+dd N	—	Partition mesh into N subdomains (Schur complement)
+dds N	—	DD with Schur complement (explicit)
+ddn N	—	DD with Neumann preconditioner
+gs	default	Gauss-Seidel DD preconditioner
+jc	—	Jacobi DD preconditioner
	Nonlinear
+nl H mat kerr relax	—	Kerr nonlinear, H harmonics
	Misc
++	—	Increase process priority
-verbose	—	Suppress console output

Architecture

Pipeline

import → mesh → assemble → solve → export

This pipeline is implemented independently in each backend:

C++ (cpp/): Compiled binary, TetGen/Triangle meshing, H(curl) elements, sparse direct/iterative solvers.

Python (py/): Triangle meshing via io_poly, scipy sparse assembly, numpy/scipy solve, pyVista field rendering.

MATLAB (m/): IOrMesh/Triangle meshing, native MATLAB sparse matrices, direct/DD/HB solvers.

Source layout

├── cpp/ # C++ FEM solver (production 3D)
│ ├── include/ # 25 headers: assembler, solver, mesh, options, ...
│ ├── src/ # 27 implementation files + main.cpp
│ └── CMakeLists.txt # C++14, links dep/lib/*.a
├── py/ # Python FEM + ML surrogates
│ ├── fes/ # FEM core: fem/, mesh/, post/, projects/
│ ├── tests/ # pytest suite
│ └── setup.py # pip-installable package
├── m/ # MATLAB reference implementation
│ ├── fes/ # Package root (mirrors py/fes/)
│ │ ├── core/ # Assembly routines (40+ files)
│ │ ├── mesh/ # Mesh I/O, geometry writers
│ │ ├── post/ # VTK field export
│ │ └── projects/ # Simulation project drivers
│ ├── tests/ # Standalone / debug / DD / NL scripts
│ └── Config.m # Path setup (addpath(genpath('.')))
├── data/ # Shared model files — .poly geometry + .h1.mat mesh caches (all backends)
└── dep/ # C++ library dependencies (dep/src, dep/build, dep/lib)

Key features

• H(curl) conforming elements — hierarchical vector basis functions (orders 1–3) for the curl-curl E-field formulation
• Transfinite elements (TFE) — exact port mode expansion for accurate S-parameters
• Domain decomposition — additive Schwarz or Schur complement preconditioners
• Nonlinear materials — Kerr effect, iterative fixed-point relaxation
• 2D TMz solver — P2 elements on Triangle triangulations, auto-detected from #Formula
• 2D electrostatic — P1 triangle assembly for quasistatic analysis
• 2D cross-section eigenmode — waveguide TE/TM mode computation on 2D meshes
• DNN-GP surrogate modelling — Deep Kernel Learning surrogates (fes, Wilson et al. 2016)
• Auto-formulation — #Formula tag in .poly selects assembly type automatically
• OOP architecture — polymorphic assembly (assembler base), strategy-pattern solvers (solver base)
• Sparse matrices — arma::SpMat<complex<double>> (C++), scipy.sparse.csr (Python)

.poly file format

Standard TetGen PLC sections (nodes, facets, holes, regions) plus custom trailing sections:

# NODES: num_nodes dim num_attributes num_markers
...
# SEGMENTS: num_segments num_markers
...
# REGIONS: num_regions
...
#Formula EM_EZ_FD ← auto-selects formulation
#Regions N
<name> <label> <epsr> <mur> <sigma> <matname>
#Boundaries M
<name> <label> <type> [numModes]

Boundary types: PerfectE, PerfectH, Radiation, WavePort.

#Formula values: EM_E_FD (3D), EM_EZ_FD (2D TMz), EM_E_TL_EIG (eigenmode), EM_E_QS (electrostatic).

File formats

Format	Extension	Description
Poly	.poly	TetGen/Triangle PLC with #Formula/#Solids/#Boundaries sections
Touchstone	.sNp	S-parameter output

Dependencies (C++)

Library	Role
OpenBLAS	Dense BLAS/LAPACK
ARPACK-NG	Sparse eigenvalue solver
MUMPS	Direct sparse multifrontal solver
METIS	Graph partitioning and mesh reordering
Armadillo	Dense and sparse linear algebra
Triangle	2D Delaunay triangulation
TetGen	3D tetrahedral mesh generation

All C++ dependencies built via ./setup into dep/.

Python dependencies installed via pip under py/ (see py/setup.py).

License

MIT — see LICENSE.