Strong vs. Weak Coupling
Concept: how and when information is exchanged
Overview
I've been asked whether weak coupling is enough for a fluid–structure analysis or whether it needs strong coupling. I don't know how to decide.
Separate two questions first. One is direction: one-way (fluid to structure only) or two-way (the structure's deformation feeds back to the fluid). The other is timing: weak coupling solves each field once per step in turn and uses the other's result one step late; strong coupling iterates within the step until both agree, or solves them together. Which you need depends on how strong the interaction is, and in fluid–structure problems there's a clear indicator for that.
Classifying coupling
| Type | Approach | Suited to |
|---|---|---|
| One-way | Solve one field and pass results to the other | Small deformation that doesn't affect the flow (thermal stress, wind-load strength checks) |
| Two-way, weak | Alternate once per step (explicit) | Weak interaction (e.g. wing flutter in air) |
| Two-way, strong (partitioned) | Iterate within each step until converged | Strong interaction (liquid with light structures) |
| Two-way, strong (monolithic) | Solve everything as one system | Very strong interaction or poorly converging iterations |
Added mass as an indicator
When a structure accelerates it must also move surrounding fluid, so it behaves as if heavier. This is the added mass $m_a$, and its ratio $r$ to the structural mass $m_s$ is a good measure of interaction strength. For a thin shell it scales roughly as below ($\rho_f$, $\rho_s$ fluid and structure densities, $R$ a characteristic size, $t$ the thickness).
| Example | Approximate ratio | Coupling needed |
|---|---|---|
| Aircraft wing in air (aluminium, 3 mm skin, 0.5 m scale) | 0.07 | Weak coupling works |
| Artery wall in blood (0.5 mm wall, 4 mm radius) | 7.7 | Strong coupling required |
A lesson from blood-flow simulation
Aircraft flutter analyses long used weak coupling — alternating between fluid and structure — with success. But when the same method was applied in the 2000s to blood flow and arterial wall deformation, a strange behaviour appeared: it diverged sooner the smaller the time step. The cause was that blood and vessel wall have almost the same density and the wall is thin, so the added mass exceeds the structural mass. Explaining this added-mass instability drove major advances in partitioned strong-coupling methods and relaxation schemes.
Worked example: a 1-DOF model
When weak (explicit) coupling diverges
In weak coupling the structure receives a fluid force based on the previous step's acceleration, so the acceleration error is multiplied by $-r$ every step and diverges for $r > 1$ regardless of time step. For a unit mass on a unit spring, 200 steps were run and the change in acceleration amplitude recorded:
| Added-mass ratio r | Δt = 0.2 | Δt = 0.05 | Δt = 0.01 |
|---|---|---|---|
| 0.9 | ×0.28 (stable) | ×0.54 (stable) | ×0.12 (stable) |
| 1.1 | about ×6.7 million | about ×64,000 | about ×2,300 |
| 2.0 | ×10⁵⁸ | ×10⁵⁶ | ×10⁵⁵ |
Even with a smaller time step, once r exceeds 1 it still diverges.
That's what makes added-mass instability so troublesome. Ordinary numerical instability stops with a smaller step, but this comes from the coupling scheme itself, so no step size cures it. The table compares a fixed 200 steps, but because the error grows per step, a smaller step packs more steps into the same physical time and actually diverges sooner in time. In practice, estimate r from density ratio and thickness as in the table above, and if it's near or above 1, start with strong coupling.
Strong-coupling iterations and relaxation
Strong coupling updates the acceleration iteratively within a step. Plain substitution still diverges for r > 1, so the update is scaled by a factor $\omega$ (relaxation). Iterations to reach a relative error of $10^{-8}$ in the 1-DOF model:
| Added-mass ratio r | No relaxation | Fixed ω = 0.5 | Aitken |
|---|---|---|---|
| 0.2 | 12 | 21 | 2 |
| 1.0 | no convergence | 1 | 1 |
| 2.0 | no convergence | 29 | 2 |
| 10 | no convergence | no convergence | 2 |
A fixed factor slows down or diverges when it is off the ideal value ($1/(1+r)$ here). Aitken's method sets the factor automatically from the last two residuals and converges quickly even for large r. Because this is a linear 1-DOF problem it converges in two iterations; real problems typically need a few to a dozen or more.
Choosing in practice
Decision steps
- Ask whether structural deformation affects the flow (or temperature or fields). If not, one-way coupling is enough.
- If two-way, estimate interaction strength with indicators such as the added-mass ratio.
- If weak (small r), start with weak coupling and confirm results don't change with time step.
- If strong, or weak coupling oscillates or diverges, switch to strong coupling with sub-iterations (e.g. Aitken relaxation). If it still struggles, consider a monolithic solver.
Beyond fluid–structure interaction
- Thermal–structural: two-way if deformation changes contact or gaps and hence heat transfer; one-way if expansion is small and contact unchanged.
- Electromagnetic–thermal: induction heating, where conductivity and permeability change strongly with temperature, needs two-way coupling.
- Different timescales: when one field is fast (electromagnetics) and the other slow (heat), the fast field is often period-averaged and passed to the slow one.
“Coupling made no difference” is good news
Re-solving with two-way strong coupling and getting almost the same answer as one-way coupling is not wasted time — it's evidence that one-way coupling was sufficient. The efficient order in coupled analysis is to solve simply first, then check whether a more rigorous method changes the result. If it doesn't, you have grounds to use the simple method next time; if it does, you have a clue to which effect matters.
Common mistakes
Mistakes and fixes
| Mistake | Effect | Fix |
|---|---|---|
| Shrinking the time step when it diverges | Added-mass instability isn't cured | Estimate the added-mass ratio and switch to strong coupling |
| Keeping a fixed relaxation factor | Slow or divergent iterations | Use adaptive relaxation such as Aitken |
| Loose sub-iteration tolerance | Energy grows; instability | Check both interface displacement and force |
| Not checking data transfer between meshes | Total force not conserved | Compare total force before and after mapping |
| Starting with the most rigorous method | Excessive run time | Start simple and step up |
I'd like to experiment with iteration and relaxation.
Try the Gauss–Seidel simulator and the SOR relaxation simulator to see how the relaxation factor changes convergence, and the Newton–Raphson simulator for nonlinear iteration. Coupling schemes are compared in partitioned versus monolithic and staggered coupling, and data transfer between meshes in mapping between dissimilar meshes.
Related Topics
Experience the theory firsthand with the interactive simulator for this field
All Simulators