Hysteresis, creep and drift: how far a pressure map can be trusted over time

A pressure map is the most persuasive artefact in a test report: a coloured field, a peak value, a neat standard deviation. It is also a measurement taken by a polymer sheet that is being squeezed, and polymers do not return to the same state when you let them go. Three different mechanisms make the same joint read differently on Tuesday than it did on Monday, and confusing them is how teams end up arguing about a result that was never in the part to begin with.

The three are not interchangeable. Hysteresis is a path effect — the reading depends on whether you are loading or unloading. Creep is a time effect at constant load. Drift is a slow change across cycles, campaigns or seasons. A calibration certificate usually folds all three into one accuracy figure, which is precisely why nobody can tell them apart when the numbers move.

Hysteresis: the sheet remembers where it has been

Load the interface to 1 MPa, hold it, then unload in steps and read at each step. The unloading curve does not retrace the loading curve — for the same pressure, the unloading reading sits low. The gap between the two curves at mid-range is the hysteresis, and it is a property of the sensing layer itself, not of your fixture.

In practice this matters in one specific way: a single reading taken while pressure is falling is not comparable with a reading taken while pressure is rising. If your process closes a joint and then relaxes it slightly before the reading is taken, you are on the unloading branch and you will systematically under-read. The fix is procedural, not technical — always read on the same branch, and say which one in the record.

Hysteresis grows as you approach the ends of the range. A sheet specified for 400–6000 kPa used at 300 kPa is operating near its lower limit, where the relative effect is largest. This is one reason a low-pressure interface should be measured with a low-range build rather than a standard sheet turned down: it is not only about resolution, it is about where on the curve you are sitting.

Creep: the reading is still moving when you stop

Hold a constant load on a piezoresistive sheet and watch the displayed pressure for half an hour. On most builds it decays — quickly at first, then slowly, settling into a rate that is small but not zero. Nothing about the interface changed. The contact area redistributed inside the sensing layer, and the reading followed it down.

This is the mechanism behind a very common disagreement: two people read the same joint, one at thirty seconds and one at five minutes, and report numbers that differ by more than the tolerance they are trying to hold. Neither is wrong. They measured at different times on a curve that was still falling.

The practical rule is to fix a read time and treat it as part of the method. For a press-closure check, the reading is taken at the moment the press reaches its set point, not after the operator has finished walking around the fixture. For a held-load test — a cure under clamp pressure, an assembly left clamped overnight — the interesting quantity is not one number but the curve, which is exactly what a logging array gives you and a single-use indicating film cannot. That is also the reason bond-line pressure before cure is measured across the whole cure window rather than sampled once.

Drift: the same sheet is not the same sheet next month

The slowest of the three is also the one that quietly invalidates comparisons. A sheet that has been through a few hundred cycles will not read identically to a new one at the same pressure, and the change is not linear in cycle count. Temperature cycling, humidity and mechanical handling each contribute. Over a campaign of a few months, the difference between "the process improved" and "the sheet aged" can be larger than the improvement you are trying to detect.

This is why per-sheet calibration with a stored record is worth more than a tighter specification on paper. A reference sheet measured at the start and end of a campaign separates sensor ageing from process change; without it, a six-month comparison is a comparison of two different instruments that happen to share a part number.

Drift is also easy to confuse with a more visible problem. Whether the array can resolve what you are comparing at all is a separate question of bandwidth — 20 Hz covers a press closure and does not cover a vibration test — and the two get mixed up whenever a slow change is blamed on a fast effect or the reverse.

None of the three is a defect. They are the behaviour of a compliant sensing layer, and every method that puts something compliant into a contact has them.

What this means for a specification

A single accuracy figure such as ±5 % FS is a summary, and like most summaries it hides the structure. It tells you the magnitude of the worst combined error, not whether that error will be stable across a comparison you care about. Two systems with the same headline number can behave very differently when you ask "did this joint get better than last quarter's?" — because one of them has hysteresis and creep and no reference record, and the other has all three plus the record that lets you subtract them.

When you are specifying a measurement, the questions worth asking are therefore not only about the headline accuracy. Ask three more:

• Is the reading taken on a defined branch of the load path, and is that branch recorded?
• Is creep characterised at the hold times this process actually uses — minutes, or overnight?
• Is there a reference sheet and a stored calibration history, so two campaigns can be compared as numbers?

A supplier who can answer those three is selling a measurement. One who can only quote an accuracy figure is selling a sensor. The difference matters exactly when you try to prove that a change worked.

Where this page stops

These are quasi-static effects. None of this describes what happens in an impact or a vibration test, where the relevant problems are bandwidth, mass loading and cable fatigue, not hysteresis.

Nor is the analysis a substitute for recalibration. Understanding the mechanisms tells you which comparisons are safe and which are not; it does not let you skip the calibration schedule. If anything, the opposite: once you can see creep on a graph, you stop trusting an uncalibrated sheet for a trend, which is the correct instinct.

And the numbers behind these effects are specific to the sensing layer — its material, its microstructure, its range. Any figure quoted here as a general rule would be an invented one. The honest position is that the mechanisms are universal and the magnitudes are not, which is why the test procedure above is more useful than a percentage.

Frequently asked questions

Is the first reading after clamping the worst one?

Usually, yes — for creep. The rate of decay is highest immediately after the load is applied, so the first seconds move the most. That is also why "read it as soon as it closes" and "read it after it settles" give different answers, and why the read time has to be fixed by the procedure rather than left to the operator.

Can software compensate for creep and hysteresis?

Partly. Creep can be modelled from a calibration curve for a known hold time, and that is standard practice in load cells. For a sheet that is also being compressed unevenly across its area, the model is weaker, because different points are at different stages of their own creep. Compensation reduces the error; it does not remove the need to state the read time.

Does a new sheet need the same treatment as an old one?

A new sheet has less drift but the same hysteresis and creep behaviour, because those are properties of the layer rather than of its age. The reference measurement matters most across campaigns, not across individual joints.

What hold time should a cure-pressure check use?

The one that matches the process. If the adhesive is under pressure for two hours, a reading taken at thirty seconds describes a moment that the part never experiences for the rest of its life. Log the whole hold and quote the value at the process-relevant time.