Evidentally...

the microtubule structure is energetically favorable for tubulin dimers near the ends and unfavorable for those in the bulk. Thus, all models of dynamic instability to date invoke a so-called "cap" mechanism.
The idea is that dimers at the microtubule ends act like a cap that prevents the otherwise unstable dimers in the bulk from escaping into solution. Stochastic loss of this cap is supposed to occur when a chemical reaction in the body of the microtubule overtakes a growing end. Bulk-type dimers are then free to escape and the microtubule disassembles. Disassembly stops, by definition, when a cap somehow reforms.
At first, it was conjectured that the chemical reaction competing with microtubule growth was GTP hydrolysis (the GTP-cap model). However, experiments found that GTP hydrolysis is practically co-incident with tubulin binding, so that the GTP-cap does not vary with growth rate, as the rate of catastrophe is known to do.
Next, it was suggested that release of the phosphate ion product of hydrolysis might be the transforming step (a Pi -cap model), but large concentrations of phosphate ions in solution did not have the predicted stabilizing effect.
What remains is a conformational-cap model, which says that GTP hydrolysis and/or phosphate release induce a time-delayed conformational change in the tubulin dimers. However, without greater insight into the nature of the conformational change, the model has little predictive power and is almost tautological.

That is, suppose the conformational change of tubulin involves a portion of b-tubulin unfolding into a random coil. The random coil domain would be a volume of radius r centered on the surface of the b-tubulin sphere (which itself excludes a portion of the volume).
In the microtubule, a row of b-tubulins wraps around the cylinder (radius Rµ) at an angle f. If random coils unfold into the microtubule interior, and if r > b (1 - b cos f /Rµ), their domains will overlap and they will sterically hinder one another.
In other words, a polymer brush will form inside the microtubule.

...unfolding

Unfolding is simple from a physical point of view in that it is completely characterized by a scalar property of the polymer, its density, and does not constrain the detailed geometry of the folded or unfolded states.
It is likewise simple from a biological point of view, where the lack of geometrical constraints implies only a weak dependence on sequence, making it an easily evolved and mutationally robust feature.
Finally, it is simple from a chemical point of view, since the relative stability of the folded and unfolded states can be reversed by a single, simple chemical reaction such as (de)phosphorylation or (de)methylation.

Then...

The polymer brush generates a pressure radially outward. Stress on the longitudinal bonds is not as great, because a-tubulins (which do not hydrolyze GTP) do not release random coils. If r > b (÷5 - 1) ª 1.2 b, random coil domains on b-tubulins will overlap with neighboring a-tubulins instead.
If the tubulin-tubulin lateral bonds fail, the geomtery of this polymer brush suggests that a microtubule will disassemble into individual protofilaments which curl back and away from the microtubule axis.
This is consistent with the morphology of shortening microtubules, often observed by electron microscopy. Protofilaments do separate from one another and curl back, away from the microtubule axis, forming "blossoms" at the microtubule ends.
Thus, the first prediction is that the radius of protofilament curvature Rpf should relate to the size of the random coil domain as
r = b ((5 + 4b/Rpf)1/2- 1).
Electron micrographs indicate Rpf = 19 nm. Taking b= 2 nm gives r = 2.7 nm.