window

view

loading…

The experiment

Start with A by itself: a disc of radius 6 km dipping 30°, its top cut off by the free surface so that it breaks the ground along a straight 8.9 km trace, with the published rough surface (rms facet slope 0.25, correlation length 4 km) and 3770 elements of side h = 0.25 km. It carries a uniform friction of 0.6 on an effective normal stress of 150 MPa and is loaded by a tectonic shear that is optimal for a 30° thrust, raised to λ = 1.10 λ0 = 129 MPa, just past the system’s first failure. Solved alone, A slips over 82 % of its area with a peak of 0.82 m (mollified kernel); the maps show where it is locked and by how much (the Coulomb excess, blue), and the normal and shear stress its own slip induces on its own rough surface.

Now add B, a disc of radius 4 km with the same dip whose trace lies 6 km in front of A’s, so that it sits 3 km above A’s plane in the hanging wall where A’s slip relaxes the load; and C, a 4 km disc dipping 50° whose lower edge is the line where its plane meets A’s, 4.5 km deep, so that it is a splay growing out of A. A is bit for bit the same mesh, the load is the same, and the three faults are solved together. What B and C do to A is then a subtraction, the system − alone view, and what A does to them is read off B and C, which exist only in the second experiment. Per kernel:

kernelA active, alone → system A’s moment, system / alone Coulomb load on A from B and C A max slip, alone → system
mollified P00.82 → 0.82×0.976−0.11 MPa0.82 → 0.82 m
classical TDE0.83 → 0.82×0.976−0.10 MPa0.81 → 0.81 m
continuous P20.84 → 0.83×0.975−0.11 MPa0.88 → 0.88 m

The neighbours take 2.4 % of A’s moment and a tenth of a megapascal of Coulomb load from it, and leave its active area and its peak slip where they were: B sits in A’s hanging wall where A’s own slip has already relaxed the load, and C is a misoriented splay that barely moves. All three kernels say so to the second decimal, which is the first thing the two experiments establish — the interaction is a property of the geometry and the load, not of the kernel. Read the other way round, A does more to its neighbours than they do to it:

faultactive aloneactive in the system moment, system / aloneCoulomb load from the othersmax slip
A0.820.82×0.976−0.11 MPa0.82 m
B0.830.83×0.967−0.12 MPa1.20 m
C0.110.11×0.883−1.25 MPa0.21 m

B, 3 km away in A’s hanging wall, loses 3 % of its moment to A’s shadow. C, rooted on A, loses 12 %: A’s slip unloads the splay above it by 1.25 MPa, which on a fault 32 MPa below strength keeps it locked everywhere but its deep end. The classical and the continuous kernels give the same fractions, ratios and shadows to the second decimal (the manifest carries all three).

The system

Every fault strikes along y and dips toward +x; its trace is where the plane meets z = 0, and the disc’s centre sits 4, 3 and 2.9 km down-dip of it, so each circle is cut by the surface and the trace is a straight chord whose vertices lie exactly on z = 0. C is cut twice: by the surface, and by the line where its plane meets A’s, 7.8 km from A’s trace, 4.5 km deep, 5.5 km long, at 20° to A. That chord is C’s lower boundary and an interior line of A’s mesh at once: its 23 vertices are computed once and written into both meshes, so the two tilings share them bit for bit.

faultradiusdiptraceelementsrole
A6 km30°8.9 km3770 the main thrust, present in both experiments
B4 km30°5.3 km1730 a parallel ramp 3 km into A’s hanging wall, where A’s slip shadows it
C4 km50°5.5 km1558 a splay rooted on A at 4.5 km depth, misoriented for this load

7058 elements in the system. The outline of each fault is sampled on its circle and its chords at a spacing of h, and the interior is a constrained Delaunay triangulation (Shewchuk’s Triangle) with no point added on any segment, so every rim vertex lies on its curve and the median edge is 1.00 h on all three faults. The roughness is displaced along each fault’s normal and tapered to zero within 1.5 km of a rim or a trace and within 3 km of the junction, so a rim vertex stays on its curve, a trace vertex on the surface and a junction vertex shared. A and B keep the published tilt (rms 14.0° and 14.9°); C, 5.8 km wide between its two chords, is tapered over most of its width and tilts 8.9°. Off the junction, C comes within 53 m of A, two eps; B’s plane is 3 km from A’s. The mollified kernels run at eps = 0.1 h = 25 m, and the continuous one at 0.2 h (why, below).

The solve

The condition is Coulomb. At the load λ the solve finds the slip δ on every element such that each is either locked and below strength, or slipping and exactly at it: w = μ(σeff,0 + λσnload) − (λ|τload| + PTδ) ≥ 0, with δ · w = 0, a linear complementarity problem. The slip-induced normal stress is read out but not fed back into the strength: with the splay rooted on A that coupling costs the operator its positive definiteness (the smallest eigenvalue of sym(−Peff) is −1.7 for the mollified kernels and −2.7 for the classical one), so the solve is the uncoupled problem for every kernel, and whatever the Coulomb excess means, it means the same thing in every map.

The load is set by a rule, not by eye: the smallest multiple of the system’s first failure at which A slips over at least 70 % of its area and B and C are each partly locked and partly slipping. Walked on the mollified operator of the three-fault system and then used unchanged for every kernel and for A alone:

λ / λ0A activeB activeC active
1.050.500.480.06
1.100.820.830.11
1.200.980.930.17
1.351.000.980.22
1.501.001.000.26

At that load the remote shear on A is 105 MPa against a strength of 106; on C it is 123 against 155, which is what misoriented means for a 50° plane under a load that is optimal at 30°. Solving A alone uses the same matrix, restricted to A’s rows and columns, and the same load, so nothing but the presence of B and C differs between the two experiments.

What the maps show

The page opens on the Coulomb excess S = τs − μσn′ of the three-fault system at a window of ±79 MPa, the 99th percentile of |S| over all elements. On a slipping element S is exactly zero — it is −w, the solver’s own margin — and those are 66 % of the area, painted white. On a locked element it is the deficit below strength, painted blue: a median of 26 MPa, and up to 92 on the steepest facets of C, which under this load are the farthest from what it would take to move them. Switch the metric to the normal-stress change to see what the coupling would have fed back: slip clamps and unclamps the rough surfaces by −10 to +12 MPa on A, −12 to +6 on B and −6 to +4 on C MPa at a window of ±6.2; the shear change along the slip direction runs −9 to +15 MPa on A and −9 to +13 on B MPa at a window of ±9.3. On a slipping element that shear is pinned by the friction law, so there the map is the remote deficit made up by slip; on a locked one it is what the neighbours transferred.

Switch the experiment to system − alone and the map is what B and C do to A, at a window of its own: at the maps’ ±79 MPa it would be invisible, and that is the first finding. It is a stress shadow, small and one-signed. The Coulomb excess on A falls by more than 0.5 MPa over 8 % of its area and by more than 1 MPa over 2 %, 2.6 MPa at the most, in a patch 1 to 2 km deep up-dip of B’s trace; nowhere does it rise by more than 0.2. The normal stress on A is clamped by up to 1 MPa beneath B and beneath C’s root and unclamped by about as much elsewhere. The slip changes by millimetres in the median and 3 cm at the most, and 19 of A’s 3770 elements lock that slipped when A was alone (under the classical kernel 23 lock and one unlocks). That is the 2.4 % of moment in the table, drawn where it happens.

Three kernels, one answer

The mollified kernel spreads each element’s slip over a width eps on either side of its plane, which is what makes the stress on the element itself finite and interpretable; the classical element (cutde) is the exact dislocation with its edge singularities; the third kernel is the mollified one with continuous quadratic slip, one slip vector at every vertex and every edge midpoint shared by the elements that meet there (14,459 nodes), the condition enforced at the nodes and the rim of each fault held at zero, so that between the centroids the slip is a smooth surface rather than a staircase of line dislocations. On a frictionless penny crack that removes the residual between the centroids — a quarter of the stress drop in the facet cores for either P0 kernel, 0.2 % for the continuous one — and brings the crack’s moment to 1 % of Eshelby’s where one slip per element gives 7 %.

With the free surface folded into the kernel the three agree over the whole fault, trace rows included. The two P0 kernels, one slip per element, differ in their active sets on 3 of the 7058 elements (34 before the fold); on the locked elements their Coulomb excess differs by 0.03 MPa in the median and 0.6 at the most, their normal-stress change by 0.04 MPa in the median and 0.3 on the trace rows (6 to 8 there before), and their shear change by less than 0.01. The mollified moment is 2 % above the classical on each fault, which is the mollification itself at eps = 0.1 h; B’s peak slip is 1.20 m against 1.12. The continuous kernel runs at twice that eps: its moment is 6 to 9 % above the classical, its trace-row normal-stress change within 1 to 2 MPa of the classical one, and its active set differs on 97 elements. At its centroids the Coulomb excess of a slipping element is a readout between the nodes where the condition was enforced, not the solver’s margin: 0.24 % of strength in the median, 1.5 % at the 90th percentile, 0.5 % on the trace rows, and 17 % on its worst element.

The free surface, folded

Until this revision the two mollified kernels disagreed with the classical one on exactly one row of elements: the ones that touch the surface. A mollified element spreads its slip over eps on either side of its plane, and for an element reaching z = 0 part of that smeared slip lies above ground. The kernel’s image construction, Mindlin’s, is right to second order in eps for a buried source, but continued above the surface it places its image inside the body, and within eps of the trace that is an O(1/eps) error. With one slip per element, sampled a third of an element below the surface, the error was eps’s and the trace row slipped 40 % too little; with slip continuous to the trace the nodes on the surface read it directly and slipped two to three times too much, moving the moment of the whole fault by 10 to 20 %.

The fix is in the kernel and is the classical image idea applied to the smeared source: the part of the slip above the surface is folded back into the body as its mirror image, which is the same as adding a mirror-image fault with reflected slip. For a vertical strike-slip fault that alone is exact; for a dipping fault with dip slip the fault and its mirror meet at a kink along the trace, leave a normal traction on the free surface, and a Boussinesq correction removes it. Only the row of elements that touch the surface needs it, and only at points within one element height of the surface. Checked where the answer is known: on a vertical strike-slip fault against the antiplane solution the trace row went from 5.0× exact (continuous slip, collocation) and 3.3× (Galerkin) to 0.98; on B alone, against the classical kernel, the continuous kernel’s moment went from 0.82 of classical to 1.01 with every depth bin within 3 %. On this page it shows as the agreement of the three kernels on the trace rows: the normal-stress change on A’s trace row reads −4.5 MPa in the median under the mollified P0 kernel against −4.4 classical, and −5.4 against −5.1 on B, where before the fold it read +4.5 and +4.9 with maxima of +26 and +48; the continuous kernel reads −3.6 and −3.0 where it read +12.5 and +25.5 with a maximum of +143. The selfcheck pins the convergence on B alone: the classical-to-mollified slip ratio on the trace row is 0.94, 0.97, 0.995 at eps/h = 0.1, 0.05, 0.025 (1.60, 1.20, 1.05 before), and the readout of one fixed slip differs by 1.2 MPa at 0.1 h and then sits at a 0.2 MPa floor, where it was 7.3 MPa falling to 0.1 only at eps/h = 0.0125.

What the continuous kernel still does not do is run at eps = 0.1 h in this, the collocation, form: its operator loses positive definiteness through broad body modes on each fault at that eps (fourteen broad modes on the three faults at 0.1 h, down to −13 against a median of +240, the asymmetry of collocation at a small eps/h on a large rough fault), so it is solved at 0.2 h by an active-set Newton method, where it converges in six iterations (its symmetric part on the free nodes has a smallest eigenvalue of −1.6 against a median of about +200, a long-wavelength mode of the kind the two P0 operators carry on this system as well). The symmetric Galerkin form of the same kernel is positive definite at every eps and, with the fold, matches the classical solve on B to 1 %; it costs four times the assembly and is the form to build the next page on.

Signs and windows

Traction is σ·n with each facet’s upward normal, so its normal component is tension positive; the normal stress in the Coulomb strength is compression positive, which is the sign the friction module uses, and the two are pinned against each other rather than reasoned about. Every window is symmetric, ±W, with W the area-weighted 99th percentile of the P0 mollified map of the three-fault system rounded to two figures, and it is the same window for every kernel and both experiments; the slider moves it and the legend counts what saturates. Nothing is smoothed, decimated or interpolated: each element carries the sample taken at its centroid, 0.29 h from its edges on an equilateral element and never nearer than 1.3 eps on this mesh.

Every array ships with its byte count and sha256 in onfault_manifest.json, with the geometry of the junction, the load, the ladder it was chosen from, the window rule, both experiments’ active fractions, moments and peak slips per kernel, and every kernel’s definiteness, sweep counts and residuals. See Rough crack for the kernel head-to-head at the centroids of a buried crack, Isosurfaces for the same disagreement off the fault, and Concepts for what eps is.