What is engineering simulation?

Engineering simulation is solving the equations that govern a physical process on a computer, over the actual geometry of the thing you are designing, so you can find out how it behaves before you build it. The umbrella term is CAE — computer-aided engineering.

Every kind of simulation works the same way underneath. The physics is written as differential equations that hold everywhere in a continuous region. A computer cannot solve those directly, so the region is chopped into a large number of small pieces — the mesh — and the equations are rewritten as algebraic relationships between neighbouring pieces. That produces an enormous system of simultaneous equations, which is solved numerically. What differs between disciplines is only which equations, and what the pieces represent.

That shared structure is why the disciplines feel similar once you have learned one, and why the same three questions decide whether any of them are worth trusting: is the geometry right, is the mesh fine enough where it matters, and are the boundary conditions the ones that actually apply?

Simulation is usually split by which physics is being solved. The boundaries between the families are conventions rather than laws, but they map closely onto which solver you reach for.

Fluids — CFD. Computational fluid dynamics solves the motion of gases and liquids: pressure, velocity, drag and lift, mixing, free surfaces, combustion, cavitation. It answers questions of the form how does the fluid move, and what force or heat does that put on my part? Explained in full here.

Structures — FEA. Finite element analysis solves deformation and stress in solids: does the bracket yield, how far does the beam deflect, at what frequency does the panel resonate, will the joint survive a million cycles. Linear static analysis is the workhorse; contact, plasticity, buckling, fatigue and fracture are where it gets hard.

Heat. Thermal analysis solves temperature: conduction through solids, convection at surfaces, radiation between them. If the fluid moving the heat around matters as much as the solid conducting it, the problem becomes conjugate heat transfer and needs a flow solution too.

Electromagnetics. At low frequency this covers magnets, motors, transformers, eddy currents and induction heating — Maxwell’s equations in a regime where the wavelength is large compared with the device. Radio-frequency and antenna work is a genuinely different discretisation and a different class of tool.

Acoustics. Sound as a wave: duct and cavity modes, how a vibrating structure radiates noise, and noise generated by flow itself.

Multiphysics. Anything where two of the above have to be solved together because each changes the other — a flexible structure deflecting in a flow, a coil heating the part it is inducing current in.

Whatever the physics, a run passes through the same stages.

• Geometry. A clean, watertight description of the region being solved. For CFD that is usually the space the fluid occupies — the negative of your CAD part — which catches people out the first time.
• Meshing. Dividing that region into cells or elements. This is where most of the human effort goes and where most bad results are decided. Too coarse and the answer is wrong; uniformly fine and the run never finishes.
• Physics and materials. Which equations to solve, and the properties that appear in them — density, viscosity, conductivity, modulus.
• Boundary conditions. What happens at the edges of the region: an inlet velocity, a fixed face, an applied load, an ambient temperature. A model is only ever as representative as these.
• Solving. Iterating until the equations are satisfied to a tolerance you choose. Convergence is not a formality — an unconverged run produces numbers that look exactly like converged ones.
• Post-processing. Turning fields of numbers into the two or three quantities the decision actually rests on, plus enough evidence that somebody else can check them.

Two choices cut across every discipline and change the cost of a run by orders of magnitude.

Steady versus transient. A steady analysis asks what the answer settles down to; a transient one asks how it gets there, resolving time explicitly. If the thing you care about is inherently unsteady — vortex shedding, a shock moving, a thermal soak — a steady run cannot answer it no matter how well it converges.

Linear versus nonlinear. Linear means the response is proportional to the load and can be solved in one pass. Nonlinearity arrives with large deflections, materials past yield, contact that opens and closes, and turbulence. Nonlinear problems are solved by repeated linear steps, and they are where convergence stops being automatic.

The honest framing is verification and validation. Verification asks whether you solved the equations correctly — is the mesh fine enough, did it converge, does refining it further change the answer? Validation asks whether they were the right equations, which can only be settled against measurement or an established benchmark.

The single most useful habit is a mesh independence study: run the same case at two or three mesh densities and watch the quantity you care about. If it is still moving, the mesh is still deciding your answer. A result quoted without one is a number with unknown error bars, however convincing the contour plot.

It is also worth being clear about what simulation is for. It is unmatched at comparing designs — this bracket against that one, this inlet position against another — because systematic errors largely cancel between runs of the same setup. Absolute prediction to engineering accuracy is a much harder claim and needs validation evidence behind it.

If you know which physics you need, go straight to it: CFD for fluids, or the full map of analysis types if you want to see everything laid out and find the one that matches your question. If you are weighing up tools rather than physics, OpenFOAM and the case for a GUI covers the open-source end of the market.