OK, it's now working fairly well.
Localized Mesh Refinement is pretty much guaranteed to prevent convergence, whilst SPA works some of the time. Quick Check is obviously the most likely to work.
I've manually set my loadsteps (9) to follow a ^0.5 curve, giving finer steps as the displacement increases.
The final limitation on convergence seems to be the point at which the pull-out force starts to drop - I've ended up first letting it fail, noting the timestep at which is fails and adjusting the displacement value to just below that value, although with SPA this has been requiring several attempts (per design iteration!).
I've also done some investigation on run-time, trying SPA and Quick Check with varying general mesh size (note that my rads, ~0.5 mm, were set to have 5 nodes around 90° in all cases):
- 5mm mesh, SPA: 124 sec
- 3mm mesh, SPA: 199 sec
- 2mm mesh, SPA: 217 sec
- 1.5mm mesh, SPA: 155 sec
- 1mm mesh, SPA: >1200 sec (aborted)
- 3mm mesh, QC: 141 sec
- 1.5mm mesh, QC: 77 sec
- 1.2mm mesh, QC: 151 sec
- 1mm mesh, QC: 72 sec
- 0.5mm mesh, QC: 250 sec!
So, finer mesh can help convergence and actually give a faster run in some cases - but it's not a continuous function, as shown by the 1.2 mesh size QC, which I tried after the 1.0 and 1.5 runs.
Results (peak pull-out force) were quite consistent regardless of mesh size, with the SPA predicting about a 5% lower value.
Thanks again for the pointers!