The open-source stack behind modern simulation

The codes that actually compute an answer. Each is mature, actively developed, and used for work that gets published or certified.

Computational fluid dynamics · GPL-3.0

The most widely used open-source CFD code, in development since the 1990s. Finite volume, unstructured meshes, strong parallel scaling, and very broad physics: incompressible and compressible flow, multiphase and free surface, combustion, conjugate heat transfer, rotating machinery and more.

It is a family of solvers rather than one binary — simpleFoam for steady incompressible flow, pimpleFoam for transient, rhoPimpleFoam and sonicFoam for compressible, interFoam for free surface, chtMultiRegionFoam for conjugate heat transfer — and choosing the right one for the physics is a real part of using it. That trade-off is covered in OpenFOAM and the case for a GUI.

Structural mechanics, thermal, structural acoustics · GPL-3.0

Developed by EDF, the French electricity utility, and used in the safety assessment of their nuclear fleet. That heritage shows in the coverage: nonlinear material behaviour, fracture mechanics, creep, damage, contact, and a documented validation base running to thousands of test cases.

It is the deepest open-source structural code available, and it is demanding — the command language and much of the documentation are French in origin, and it expects a user who knows what they are asking for. It also handles structural acoustics, where the structure rather than the air is the hard part.

Low-frequency electromagnetics, thermal, acoustics, elasticity · LGPL / GPL

A multiphysics finite element package from CSC in Finland, and the mature open-source answer for low-frequency electromagnetics: magnetostatics, eddy currents, induction and Joule heating, motors and transformers.

Its real strength is breadth inside one code — electromagnetics, heat, elasticity and acoustics all present, with coupled-field problems solved internally rather than through an external coupling layer. Driven by a readable .sif solver input file. One integration therefore covers a lot of ground, which is why it carries several SHD Sim module families at once.

CFD with adjoint shape optimisation · LGPL-2.1

Originating at Stanford, SU2 is a compressible-flow oriented CFD suite whose distinctive capability is a first-class adjoint implementation — both discrete and continuous. That makes gradient-based shape optimisation practical at a cost that does not scale with the number of design variables.

OpenFOAM has adjointOptimisationFoam as an alternative, less maturely. Where OpenFOAM is the broader general-purpose tool, SU2 is the one to reach for when aerodynamic shape design is the actual objective.

Explicit dynamics — crash, impact, drop test · AGPL-3.0

Altair open-sourced its Radioss explicit solver in 2022, putting a genuinely industrially validated crash code in the open for the first time. Automotive crash, drop test, blast and impact, with the strain-rate dependent material models and contact algorithms that class of work requires. See explicit dynamics for what that actually involves.

Note the licence: AGPL-3.0 carries obligations that differ from GPL for network-delivered services, which is worth checking against how you intend to deploy it.

Radio-frequency electromagnetics · GPL-3.0 / Apache-2.0

The open-source names for high-frequency work — antennas, waveguides, filters, EMC. openEMS is a finite-difference time-domain solver; Palace, from AWS, is a finite-element solver aimed at electromagnetic device and quantum hardware design.

They are listed here for completeness and honesty rather than as a commitment. RF is a different discretisation, a different set of outputs — S-parameters and radiation patterns rather than torque — and a different audience from the low-frequency electromagnetics Elmer covers. It is on our list only if RF turns out to be genuinely wanted, and we would rather say that than imply a capability we have not committed to building well.

Where most of the human effort in simulation actually goes, and the part that decides whether a solver ever gets a fair chance.

CAD geometry kernel — STEP, IGES, BREP · LGPL-2.1 with exception

The geometry kernel that makes real CAD usable. Polygon formats such as STL and OBJ describe a surface as triangles and lose the underlying shape; STEP, IGES and BREP carry the actual analytic surfaces, which is what you need to identify faces reliably, apply boundary conditions per patch, and mesh curvature properly. OCCT is what reads them, and it is the reason importing a supplier’s STEP file is a normal operation rather than a conversion exercise.

Automatic hex-dominant meshing for CFD · GPL-3.0 (part of OpenFOAM)

OpenFOAM’s automatic mesher: it starts from a background block mesh, refines it towards your surfaces, snaps to them, and inserts boundary layers. It is capable and it is controlled by a dictionary with a great many interacting settings — refinement levels, feature edges, surface and volume refinement regions, and layer parameters that can silently fail to insert if the cells underneath are not suitable. Most CFD results that are wrong are wrong because of decisions taken here.

Finite element meshing for structural and EM · LGPL / GPL

Structural and electromagnetic analysis needs a different kind of mesh from CFD: tetrahedra and hexahedra with mid-side nodes for second-order elements, which snappyHexMesh cannot produce. Netgen and SMESH cover that FE meshing path, and ElmerGrid handles mesh preparation and partitioning for Elmer. This is a real prerequisite for the structural and electromagnetic modules rather than a detail — a solver is no use without a mesh it can accept.

Not solvers, but the plumbing between them — and, as multiphysics work usually discovers, where a coupled setup goes wrong before either solver does.

Conservative field mapping between non-matching meshes · LGPL-3.0

Two physics rarely share a mesh — a CFD mesh is refined for boundary layers, a structural mesh around stress concentrations — so every exchange needs a field interpolated from one surface discretisation onto another. MEDCoupling does that conservatively, meaning the total force or heat handed over matches what was computed. It is all a one-way chain needs, and one-way chains answer more industrial questions than people expect.

Two-way coupling between independent solvers · LGPL-3.0

A coupling library that orchestrates true two-way exchange: it handles data transfer, interpolation and the coupling iteration between codes that know nothing about each other, with adapters for OpenFOAM, code_aster and others. This is what genuine fluid–structure interaction needs, and OpenFOAM + code_aster is the realistic starting point because those are the two solvers the other modules already run.

The layer that turns one solver run into an answer about a design space. Notably, the most valuable part of this needs no external dependency at all.

Parameter sweeps, design of experiments, the run matrix · Ours

Running the same case many times with different inputs and collecting the results is orchestration, not numerics — it needs no external tool, and it delivers most of the value of the whole studies category. Sweeps, DoE sampling and the run matrix are built in rather than delegated, which is also why they can arrive before the optimisation libraries below.

Derivative-free optimisation, in-process · LGPL / GPL-3.0

Optimisation algorithm libraries that need only the objective value, not a gradient — which suits simulation, where the response can be noisy, discontinuous or expensive. NLopt provides a broad set of local and global algorithms; pagmo adds parallel island models for population-based methods such as genetic algorithms. Both run in-process around the solver loop.

Uncertainty quantification and calibration · LGPL-2.1

From Sandia National Laboratories, and the serious open-source tool for the questions beyond “what is the best design”: how much does the output move given scatter in the inputs, which parameters actually drive it, and how do you calibrate a model against measured data. This is the uncertainty side of studies, and it matters because an optimum that collapses within manufacturing tolerance is not an optimum.