loading…
What the difference is
A collocation scheme only ever asks a kernel for traction at an element’s own centroid, and that is the one place where the classical finite part exists — finite for a geometric reason, the centroid being a fixed 0.29 h from the element’s own edges. Approached along its normal that value is flat: stepping in from 30 eps to 0.003 eps the von Mises stress changes by less than a per cent, with a log-log slope of +0.002 — flat, and matching the mollified kernel to 1–3 %. That is why a classical triangular dislocation element assembles a perfectly good interaction matrix on this mesh, and on a separate head-to-head over four roughness levels it recovers the Eshelby solution marginally better than the mollified kernel does.
A stress field is a different question, because it asks for stress everywhere rather than at centroids. Two adjacent elements that slip by different amounts leave a residual dislocation on the edge they share, and its stress diverges. Approaching the midpoint of one interior shared edge of this mesh — two faces 4.6° apart, slipping 0.256 and 0.233 m, so a 9.9 % mismatch:
| d/eps | mollified (GPa) | classical (GPa) | ratio |
|---|---|---|---|
| 0.001 | 6.01e-3 | 3.95e+0 | 658 |
| 0.01 | 6.01e-3 | 3.95e-1 | 65.8 |
| 0.1 | 6.10e-3 | 4.02e-2 | 6.59 |
| 1 | 6.87e-3 | 7.43e-3 | 1.08 |
| 3 | 5.47e-3 | 5.50e-3 | 1.004 |
The classical log-log slope over the inner two decades is −0.9963: a clean 1/d. The mollified slope is +0.0032. At 0.05 m from the edge the classical von Mises stress has reached 3.95 GPa — 13 % of the shear modulus, from a quarter-metre of slip.
The cause is the slip mismatch, not the geometry, and that is settled by a control rather than argued: set those two elements’ slip equal and change nothing else, and the singularity disappears outright — slope +0.0003, ratio 1.017. Make the slip uniform over the whole crack and the two kernels agree to 1.000 at every distance. The divergent line is not physics. It is the discretisation’s own residue, and there is one on every interior edge where the slip varies: 1518 interior edges, 770 km of spurious singular line against 49 km of physical rim, a factor of 15.7, or 6.8 km of singular line per km² of crack.
Mollification removes all of it, because eps is a declared physical length rather than a property of the mesh: the jump is smeared over a width eps, so a shared edge carries a smooth profile instead of a line. The physical crack front — the rim, where the stress really is singular — is the only one left, and it is regularised to the same declared width.
What this volume does and does not show
The grid is 0.4 km, which is 8 eps. It therefore undersamples the singular lines by design, and the ratios it reports are a lower bound: the largest classical/mollified ratio anywhere on this grid is 6.29, against the 658 above, because no grid point lands closer than about half an eps to an edge. Resolving a 1/d line in a volume needs spacing well under eps in all three directions, which is a million times this cost. What the volume is for is the extent of the disagreement, and that it reads honestly: a thin shell on the crack, and agreement to one part in ten thousand everywhere else.
The crack surface is the mesh itself: 1044 facets on 571 shared vertices, welded from the same triangles the field was computed from and refused unless it reproduces them. The wireframe on it is its own edge set, and the 1518 interior edges in that picture are the lines this page is about — there is nothing physical at any of them. Translucent, so it never hides a slice; turn it off to read the slices alone.
Colour is log10 of the field and always was — the payload stores log10 in each byte, one level per 1/253 of the span — so the ticks are whole powers of ten. What the decades slider sets is how much of that ramp is shown, counted down from the array’s own maximum, with both ends saturating. It opens on all 3.0 decades, because a 1/r3 field in a 20 km box is mostly far field and narrowing the window buys contrast at the crack by flattening the rest: 2.5 decades puts 28 % of the voxels on one colour, 2.0 decades 70 %. The difference state is the magnitude of a signed difference, floored four decades below its own maximum — the state arrays quantise at about 1.6 % relative, so anything below that floor is under the payload’s own resolution rather than a measurement, and it is drawn as no-data rather than as a colour. See Examples for the quantisation bound this payload is checked against, and Concepts for what eps is.