Nobody warns you that the hardest number in a machine build is temperature. Stress you can compute from a drawing, vibration you can fix with a gusset, but heat arrives quietly and then demands a gap, a torque value, or a shim that nobody budgeted for. I found this out the expensive way on a welded steel frame for a drying tunnel, where the top platen ran at 120 degrees Celsius and the frame underneath stayed at shop temperature. At first review the alignment between the upper and lower rails looked fine on the bench. Two weeks into commissioning the customer measured a 0.4 millimeter vertical mismatch that was pushing the product out of tolerance. The drawbar was gone before anyone looked at the physics.
Thermal Structural Coupled Analysis Is Two Solves With One Story
The phrase sounds fancier than it is. In practice a coupled thermal-structural analysis is a temperature solve and a structural solve that share a model. You first get the steady or transient temperature field from a heat-transfer analysis, then you feed that temperature field into a structural analysis as thermal loads, and the solver expands each element according to its coefficient of thermal expansion. The subtlety is that the two solves may or may not be bidirectional: a stiffness change that barely moves temperature can be ignored and you run them one-way, but a problem where deformation closes an air gap and changes the cooling flow needs a real iteration, and that is where beginners burn their weekend.
| Workflow stage | What I actually set up | The question it answers |
|---|---|---|
| Thermal model | Heat source, convection, radiation if it matters | Where is the part hot and how hot |
| Temperature mapping | Element or node interpolation between meshes | Do both models speak the same heat |
| Structural model | Material, supports, temperature as load | How far and in which direction it moves |
| Strain and stress | Expansion, restraint, bolt preload | Will it bend, bind, or crack |
My default for machine frames is a one-way coupling with the temperature field passed from the thermal solve, because the solid steel parts I work with do not change their geometry enough to disturb the thermal solution. The moment that assumption fails, and it fails when you have a thin sheet metal duct that bulges under pressure and changes its own convection area, I switch to a bidirectional loop and let the geometry update between iterations. That classification right there, one-way versus two-way, is the single most useful decision in any thermal-structural job, and I would rather get it right than chase a fancier solver feature.
The 0.4 mm Gap: What The First Run Missed
Let me be honest about how I missed it, because the story is more useful than the math. I ran the thermal analysis on the platen alone, saw a uniform 120 degrees, and ran the structural analysis with the frame fully grounded, and the report showed a few tenths of a millimeter of expansion that I shrugged at. The failure was my boundary condition, not the physics: I had released the platen edges but treated the frame legs as fixed, which is exactly wrong. The frame is the soft, long lever in the chain. When the top platen grows it pushes the unheated side rails, and the frame, being welded and massive, resists and bows the whole assembly by far more than the platen’s own expansion predicts. A 0.4 mm mismatch came from the frame trying to stay rigid while two rails wanted to grow at different rates.
In a coupled thermal-structural model, the biggest error is almost never the solver. It is the boundary condition that pretends the cold member stays exactly where the drawing says it should.
The fix was a thermal relief slot and a looser tolerance on one rail, plus a shim pack for commissioning, and the analysis went from a decoration to the thing that placed the shims. Since then I run every heated assembly the same way: hot member, cold frame, real supports, and a displacement review at all the critical inspection points, not just the loudest stress.
Bolt Relaxation Is Where Coupled Analysis Earns Its Keep
Here is a number that sticks with me: a M16 class 8.8 bolt torqued to its book value at 20 degrees can lose between ten and twenty percent of its clamp load when the joint sees a fifty degree rise, without a single thread moving. The bolt material and the clamped parts expand at different rates, the interface gets a tiny gap in the corner of the flange, and the preload quietly bleeds into strain that never comes back. In a pure structural analysis nobody asks this question, because the model has room temperature baked into every coefficient. Throw the temperature field in and suddenly the joint becomes the honest part of the model.
| Steel grade | CTE, x10^-6 per K | Conductivity, W/m.K | What I use it for |
|---|---|---|---|
| EN 10025 S235JR (mild) | 12.0 | 50 | Frames, brackets, low-heat |
| Grade 316 stainless | 16.0 | 15 | Process contact, steam |
| 6061-T6 aluminium | 23.6 | 167 | Heatsinks, stiff light plates |
| Cast iron GG25 | 10.5 | 50 | Machine beds, damping |
Notice what this table does to a design. Aluminium grows nearly twice as far as mild steel per degree, and it conducts more than three times as well, so a motor plate that reads fine at assembly can grab a bearing ring by the time the motor reaches running temperature. If you bolt aluminium to steel at 20 degrees you have already decided the preload behavior for every degree above that. I have stopped quoting torque values without also quoting a temperature, and I write the same reminder into every drawing that carries a critical joint: retorque spec, test temperature, and the delta that matters.
Step By Step, The Way I Actually Run It
My sequence on any coupled job is boring on purpose. First, model the heat source honestly, which for an electric motor means the copper losses in the stator winding more often than the shaft, and for a heated platen means the cartridge heaters plus the convection to the part on top. Second, run the thermal solve once and save the temperature field as a named result, because you will re-use it a dozen times. Third, swap in the structural model with the same mesh seams, apply the mapped temperatures, and add the boundary conditions that represent how the assembly is really mounted. Fourth, review displacements at inspection points, not just stress, and fifth, loop back to the thermal side only if geometry changed enough to matter. I put the sequence in that order because skipping the saved temperature field is the fastest way to re-solve the whole problem for every small change you make afterward.
Do not spend the afternoon chasing a convergence tolerance on a steel weldment at 120 degrees. Spend it asking whether the cold rail is allowed to move, because that is the number the inspector will measure at the end.
One last practical note on meshing. For the structural side I let the solver use its default element sizes because expansion strains are smooth and forgiving, but I keep a fine mesh at the bolt seats and the plate edges where real stress concentrates. For the thermal side the mesh is driven by the gradient, so I refine where heat flows into a thin flange, and let the thick body stay coarse. A good thermal mesh and a good structural mesh are rarely the same mesh, and pretending they are is the second fastest way to a slow, wrong answer.
When One-Way Stops Being Enough
You will run into the two-way case sooner than you think. The classic is a press tool or a mold where the heated insert warps enough to change the contact with the knockout pins, or a thin-walled plenum whose cross section depends on its own pressure. In those situations I run a very small number of staggered iterations, update the heat-transfer area from the new shape, and watch the results settle. I have never needed more than four or five iterations on mechanical hardware, and if the answer is still swinging after that, the model is wrong, not the coupling scheme. That is the honest summary after nearly a decade of thermal-structural work: coupling is about deciding who moves first, and the frame always moves first, and remembering that has saved me more 0.4 mm gaps than any solver option ever will.
Verify With a Thermocouple, Not With Conviction
The most humbling part of any thermal-structural job is the measurement, because the numbers on the report almost never match the numbers on the floor at first. I now spend budget on this in every heated design: a handful of type K thermocouples glued to the critical joints, one thermal imaging pass after warm-up, and a dial indicator or a laser target at the points the analysis said would move most. The 0.4 mm case came back with a 0.36 mm reading on the dial gauge after I fixed the boundary condition, and that kind of agreement, a few tenths of a millimeter out of a few hundred, is what makes the whole discipline worth trusting at all. If your measured gap is off by a factor of two, assume the model told you where to look, not how far to look.
There is also a time-constant trap that trips up everyone on the first heat-up cycle. A welded steel frame skims slowly, so a thirty minute soak can still leave the far rail twenty degrees cooler than the platen at the moment you read the gauge. I take readings at steady state, which on steel hardware usually means waiting until the temperature drift drops below a degree per five minutes, and I log the gradient across the frame, not just the hottest point. The model is a snapshot of a state, and if you compare it to a measurement taken while the frame is still climbing, you are comparing two different problems.
Honestly, I have come to think of thermal-structural coupled analysis as the quiet workhorse of machine design. It does not make the dramatic pictures that a crash simulation makes, and it does not get the meeting time a vibration analysis gets, but it is the one that explains why the drawing that fit perfectly on Monday binds on Wednesday afternoon. Once you learn to read the temperature field as a load, and once you accept that the cold member is the one that moves, the method stops being an exotic solver feature and becomes the ordinary way a frame earns a place in a workshop. That is the change I hope this writes into your workflow: start with the hot member, free the cold frame, and check the gap, because the gap is where customers find your mistakes.
One more habit worth stealing before I close: I always reprint the material card before a hot run, not after. Steel changes its yield strength with temperature in a way that surprises people who only ever looked at the room temperature table, and aluminium creeps if you hold it hot long enough, and coefficient of thermal expansion is not the constant the catalogue implies. I keep a column in my standard model set for the 100 degree and 200 degree variants of every material I use on a heated machine, so the temperature field never feeds numbers that were written for a colder life. It is a small discipline, but it is the difference between a coupled analysis that decorates the report and one that predicts the shim pack.
