COMSOL Multiphysics Approach
Theory: solve everything at once, or alternate?
Overview
When solving coupled problems in a general tool like COMSOL, which solver setting should I choose?
Broadly there are two: fully coupled (monolithic) and segregated. Fully coupled solving gathers every unknown — potential, temperature, displacement — into one system and solves it with Newton's method including how each affects the others. Convergence is fast and robust, but the matrix grows, raising memory and per-iteration time. Segregated solving solves each physics in turn and repeats: each pass is light, but strong coupling increases iterations and can prevent convergence. The choice depends on coupling strength and problem size. Here we compare both on a strongly coupled case — a device that conducts better as it warms.
Test problem: electro-thermal thermistor
Voltage U is applied across a 10 mm bar with both ends held at 300 K; σ₀ = 1 S/m, B = 3000 K, thermal conductivity k = 2 W/mK, 41 nodes. Conductivity rises (resistance falls) with temperature.
Fully coupled Newton method
The off-diagonal blocks are temperature's effect on electrics and electrics' effect on temperature. The segregated method omits them and alternates single-field solves.
Newton and Raphson
Finding a root by repeatedly following tangents appears in Newton's manuscript of around 1669, but it was published only in 1711. In between, in 1690, Raphson published it in a more practical form — hence the name Newton–Raphson method. A method over 300 years old sits at the heart of solving coupled problems with millions of unknowns today.
Worked examples
Example 1: voltage and iterations to converge (tolerance 10⁻⁸ K)
| Voltage | Peak temperature | Segregated | Fully coupled |
|---|---|---|---|
| 10 V | 307.0 K | 11 | 4 |
| 15 V | 318.7 K | 18 | 5 |
| 20 V | 347.5 K | 38 | 6 |
| 22 V | 377.4 K | 69 | 8 |
| 23 V | 415.6 K | 166 | 9 |
| 23.5 V | — | diverged | not converged |
Example 2: where solutions exist (continuation)
Raising voltage in 0.1 V steps and starting Newton from the previous solution, solutions exist up to 23.2 V (peak 443.3 K) and vanish at 23.3 V. Beyond that there is no steady state and temperature keeps rising.
So at 23.5 V neither converged not because of bad solver settings, but because there's no solution.
Exactly. When a coupled calculation won't converge, many people first change relaxation factors or shrink time steps, but if physically no steady solution exists, no setting will converge it — and a forced answer would be more dangerous. To tell, tighten conditions gradually while tracking the solution (continuation) and see whether peak temperature shoots up and then ends. Here peak temperature rose 38 K between 22 and 23 V and ended about 0.3 V later. With that shape, the design should keep ample margin below it. A sudden rise in segregated iterations is also a warning that you are near the limit. The fully coupled method's advantage is tracking solutions right up to the edge with few iterations, so the limit is located correctly.
Modelling workflow
- Solve first at mild conditions (low voltage or load) and confirm convergence.
- Choose fully coupled for small, strongly coupled problems and segregated for large, weakly coupled ones.
- Ramp conditions gradually, starting each solve from the previous solution.
- If iterations or peak values jump, suspect the edge of solution existence.
- Where steady solves fail, run transient analysis to see what actually happens.
“An analysis that wouldn't converge however much relaxation”
In an electro-thermal analysis of a power-supply component, only the 1.3× rated load case refused to converge. The engineer repeatedly lowered segregated relaxation factors and raised iteration limits, without change. Switching to fully coupled and ramping load while tracking the solution showed peak temperature shooting up near 1.25× rating with no solution beyond: the component's material had resistance falling with temperature and sat at a thermal-runaway threshold. The design was changed to strengthen heat dissipation.
Common mistakes
Mistakes and fixes
| Mistake | Effect | Fix |
|---|---|---|
| Solving at final conditions directly | Poor initial guess, no convergence | Ramp conditions |
| Blaming settings only | Runaway missed | Check solution existence |
| Segregated for strong coupling | Many iterations, divergence | Use fully coupled |
| Fully coupled on huge models | Out of memory | Segregated with preconditioning |
| Loosening tolerance to 'converge' | Wrong answer | Check residuals and physics |
I'd like to learn related topics.
Related pages include convergence of coupled analysis, staggered coupling, co-simulation coupling, reduced-order models for coupled problems and Joule heating.
Related Topics
Experience the theory firsthand with the interactive simulator for this field
All Simulators