view

loading…

What a level set is a test of

On the rough-crack page the classical triangular dislocation element comes out well at the one place a collocation scheme ever asks: an element’s own centroid, which sits a fixed 0.29 h from its own edges and so a fixed distance from the singularity of its bounding loop. It is finite there for that geometric reason, which is why it assembles a good interaction matrix and recovers Eshelby marginally better than the mollified kernel does. What it cannot do is hand you the field between the elements, or anywhere else on the fault — the On the fault page is that measurement. Two adjacent elements that slip by different amounts leave a residual dislocation on the edge they share, and its stress diverges as 1/d — on 1518 interior edges, none of which is a physical feature of the crack.

A level set is where that becomes visible rather than tabulated, because it is the one readout that asks the kernel for a value at every point at once. Where the two kernels agree, the two surfaces are the same surface. Where they do not, one of them wraps the mesh.

Two scales, because one grid cannot hold both

The classical field reaches a level L at a distance r ≈ A/L from a shared edge, with A ≈ 2×10−4 GPa km measured on this mesh. At 0.05 GPa that is 4 m. Resolving 4 m over a 12 km crack is of order 109 grid points, or several hundred hours of kernel evaluation — so this page does not pretend to show the singular lines at full scale. It shows two honest things instead:

sceneboxspacingpoints what it is for
envelope14 × 14 × 9 km 0.2 km = 4 eps231,886 the shape of the field around the whole crack, and how far apart the two kernels put the same level
detail1.6 × 1.6 × 0.6 km 0.02 km = 0.4 eps203,391 one patch at 20× the resolution, fine enough for a shared edge’s own structure to be several cells across

The detail slab is centred on the interior edge carrying the largest slip mismatch on the whole crack, and it is oriented on the local mean plane of the facets it contains rather than on the crack’s global mean plane: across a 2.4 km footprint this surface wanders 0.47 km about the global plane and 0.18 km about the local one, so the global alignment would have sampled mostly empty half space and clipped the crack at both ends.

What the three scalars are

Stress is a tensor, and “a stress isosurface” means nothing until the scalar is named. All three here come from the same elastic stress tensor, with the mollified double layer’s eigenstress subtracted, in GPa:

Levels are set by a rule rather than by eye: the 90th and 99th percentiles of the mollified scalar inside the box, and both kernels are then contoured at the same number — which is the only way the comparison means anything.

What the comparison measured

Both kernels were evaluated at all 231,886 points of the envelope grid, and contoured at the same levels — percentiles of the mollified field, so the classical surface is never handed a level chosen to flatter it. Two numbers come out of each pair. The distance is from every vertex of one surface to the other surface, point to surface rather than point to vertex, so a denser mesh cannot read as a closer one. The area is the discriminator the distance cannot be: two level sets at the same level enclose the same region or they do not.

levelmedian apartworst apart mollified areaclassical areaexcess
von Mises p901.2 m (0.006 cell) 758 m (3.8 cells) 251.2 km2251.7 km2 +0.2 %
von Mises p999.3 m (0.046 cell) 498 m (2.5 cells) 121.4 km2130.2 km2 +7.3 %
ΔCFS +L lobe31.6 m (0.158 cell) 1231 m (6.2 cells) 20.6 km230.5 km2 +48 %

So the answer is not one word. Over nearly all of their area the two surfaces are the same surface: a median departure of 1.2 m at p90 is six thousandths of a grid cell, and the two fields’ minima over the whole box agree to five digits (1.1168 against 1.1166×10−5 GPa) — the far field is not in question. The departures are local, and they are large where they occur: several cells, at isolated places, which is what a defect on a shared edge looks like from 0.2 km away.

The area says which kernel is the one moving, and the trend says why. Move the level from p90 into the tail and the classical surface inflates — 0.2 %, then 7.3 %, then 48 % — because the tail of the classical field is the edge singularities. Taken over the whole box, the classical von Mises maximum is 0.217 GPa against the mollified 0.0134, sixteen times larger, on a grid whose cells are 4 eps wide and which therefore comes nowhere near resolving the lines carrying it. That factor is a lower bound set by the grid, not a property of the field.

One statistic was measured and does not discriminate: the number of separate pieces. A tail level set of a rough crack is dozens of lobes rather than one shell, and at von Mises p99 the mollified surface carries more of them than the classical one — 38 against 34. Fragmentation here is set by the roughness of the crack and by where it slips, not by the kernel, so the piece count is published as a description of what you are looking at and not as evidence.

And what it measured at twenty times the resolution

The detail slab samples at 0.02 km — 0.4 eps, ten times finer than the envelope — and there the two kernels are not the same object at all:

levelmedian apartworst apart mollified areaclassical areaexcess
von Mises p9018.1 m (0.90 cell) 394 m (19.7 cells) 3.70 km25.26 km2 +42 %
von Mises p9958.6 m (2.93 cells) 448 m (22.4 cells) 0.93 km23.09 km2 +232 %
ΔCFS +L lobe161 m (8.1 cells) 965 m (48.2 cells) 0.13 km21.30 km2 +918 %

The field itself says the same thing more bluntly. On these same points, from this same slip, von Mises runs to 0.0142 GPa under the mollified kernel and 4.11 GPa under the classical one — a factor of 290. And the mollified maximum barely moved from the envelope’s 0.0134: a bounded field does not care what grid you ask it on. The classical maximum went from 0.217 to 4.11 GPa for the same refinement, because what it is reporting is not a feature of the crack but how close the grid came to a line.

The tubes, predicted before they were looked at

A residual dislocation on a shared edge gives a 1/d field, so a level L should be reached at r = A/L and the classical level set should wrap that edge in a tube of radius r. That is a falsifiable statement, and it was written down before the surfaces were extracted: 3.4 to 7.6 cells at these levels, which is why the slab is sampled this finely. Measured afterwards off the extracted geometry, as the median standoff along the middle of the edge the slab is centred on:

levelpredicted rmeasured ratiomollified, same edge
von Mises p90 6.40 MPa 112.7 m78.1 m0.6915.3 m
von Mises p99 9.30 MPa 77.5 m51.9 m0.6732.9 m
σ1−σ3 p90 7.28 MPa 99.1 m79.2 m0.8015.3 m
σ1−σ3 p99 10.59 MPa 68.1 m52.2 m0.7733.0 m

The tubes are there, and they are about two thirds to four fifths of the predicted radius. The prefactor A is the weak half of that prediction — it was measured at one edge whose slip mismatch is 3.6 times smaller and scaled linearly — so the law is tested separately from it. r ∝ 1/L holds to 3.4 % on von Mises and 4.0 % on differential stress: the ratio of the two measured radii is 0.664 and 0.660 against the 0.687 that the two levels alone predict, with no fitted quantity anywhere in it.

The mollified surface is the control, and it behaves in the opposite sense. Its standoff from that same edge grows with the level, 15.3 m to 32.9 m, where a tube must shrink as 1/L. It is not wrapping the edge; it does not know the edge is there, and its position is set by the smooth field around it.

The tube rule is not applicable to ΔCFS, and the measurement says so rather than being quietly omitted: both lobes come within 13.1 m of the edge, under one cell, against a predicted 152 m. The 1/d coefficient of a signed scalar changes sign with azimuth around the line, so there is no radius at which the field reaches +L all the way round. Both lobes run into the line instead of standing off it — which is why they measure the same 13.1 m.

How much to trust the surfaces

Each surface is checked against the field it claims to come from: every vertex of the undecimated contour is compared with the trilinearly interpolated grid, and the worst departure from the requested level is reported in the manifest. Across all 24 published surfaces that worst departure is 0.11 %. The check is not a tautology, and the selfcheck proves it on a closed form by breaking it on purpose: against an analytic ellipsoid the contour sits on its level to 3×10−5 %, while a y↔z transposed grid reads 44 % off, an x↔z transposed one 100 %, and a 10 % error in the grid spacing 56 %.

Nothing here was decimated or smoothed. The budget is 60,000 triangles and the largest surface on the page is 38,434, so every published mesh is the contour as it came out, and every decimation deviation in the manifest is exactly zero. That is deliberate: a smoothing pass is precisely what would remove a structure a few cells across and manufacture the agreement this page is testing.

One grid node in the classical detail field — 1 of 203,391 — came out non-finite, where the point landed close enough to an element edge for the kernel’s logarithms to overflow. The field really is unbounded there, so rather than invent a value, that node is given the largest-magnitude value among its six neighbours: because the field grows toward the edge, that puts it on the side of the level it is genuinely on. Every surface’s manifest entry carries how many nodes this happened to. The mollified field has none, at either scale.

The meshes are binary glTF in kilometres, z up and depth negative, with no transform node — the same files this page loads will open in Blender, which declares y up and will want one rotation. Every surface carries its scene, kernel, scalar, level, triangle count and sha256 in iso_manifest.json. See Rough crack for the measurement this picture is of, and Concepts for what eps is.