shipengcheng1230 / quaycle.jl Goto Github PK

• Triangle
  https://www.sciencedirect.com/science/article/pii/S0098300407000593
  Better alternative:
  https://academic.oup.com/gji/article/201/2/1119/572006#86405621

• Polygon & Transversely isotropic
  https://pubs.geoscienceworld.org/ssa/bssa/article-abstract/104/6/2698/331796/elastic-deformation-due-to-polygonal-dislocations?redirectedFrom=fulltext

Would cause problem if embeded in functional call with argument named as names.

Need to rewrite dθ_dt, dμ_dvdθ and their in-place version

There is a problem when assembling the stiffness tensor when adding a buffer zone. It violates the basic FFT procedure if not extending the fault length accordingly.

Need to expand this:

https://github.com/shipengcheng1230/JuEQ.jl/blob/0406b206a712a20744818441f6175718273af98c/src/gf_okada.jl#L85-L88

Might need additional attributes in property struct to indicate velocity direction and slip type.

Need benchmark against use of Strided.jl.

Reference:

Erickson, B.A., Dunham, E.M., 2014. An efficient numerical method for earthquake cycles in heterogeneous media: Alternating subbasin and surface-rupturing events on faults crossing a sedimentary basin. Journal of Geophysical Research: Solid Earth 119, 3290–3316. https://doi.org/10.1002/2013JB010614

One key issue is how to assign arbitrary Dirichlet boundary. A potential solution is seeking a mapping from vertex to dof.

Some breaking changes in HDF5 need to be fixed.

• New modeling API to accommodate various input type

• Distributed BLAS using COSMA

• CUDA integration

• New solution saving interface

• Remove Green's function API

• No mixed stress/strain components in 3D

Use set_dims!.

A number of features for analyzing results:

• saving procedure (functional callback)
• max velocity for various input types
• cumulative slip (integrate slip into odes?)
• moment magnitude
• event detection
• stress drop (energy-based (Noda et al., 2013))

A number of features for usage:

• FFTW flag & threads options
• Devectorize computation for better handling buffer zones
• Further minimize allocation (mostly from out-place FFT)

https://github.com/shipengcheng1230/JuEQ.jl/blob/4de55917919aa2b16236d549871c9b121a97de29/src/tools/gmshtools.jl#L186-L187

By using Luxor.jl.

https://github.com/shipengcheng1230/JuEQ.jl/blob/f377778afbc54dd6b4251d5b703617761f087808/src/assemble.jl#L107

This may cause a problem if only diffusional creep is considered.

• The clamped θ will cause infinite dμ_dθ, hence infinite dv_dt:

https://github.com/shipengcheng1230/JuEQ.jl/blob/f377778afbc54dd6b4251d5b703617761f087808/src/assemble.jl#L56-L70

Given BP1:
on 1.1

julia> @time sol = solve(prob, TsitPap8(), reltol=1e-6, abstol=1e-6);
  1.411736 seconds (1.55 M allocations: 78.117 MiB, 6.79% gc time)

on 1.2-rc2

julia> @time sol = solve(prob, TsitPap8(), reltol=1e-6, abstol=1e-6);
  5.128751 seconds (7.36 M allocations: 675.122 MiB, 6.69% gc time)

Strain green's function isn't right:

https://github.com/shipengcheng1230/JuEQ.jl/blob/7fddd4acd39f48cfdc9aee77e652a2b950be30d2/src/gf_okada.jl#L208-L216

Potential references are here.

This is related to how slips are incremented. Middle-difference does not resolve this as expected.

Various utils function dealing with velocity, stress drop, catalog.

Integrate with plotting lib for basic quick visualization.

Refer: shipengcheng1230/GmshTools.jl#1

• https://academic.oup.com/gji/article-abstract/208/2/877/2454292?redirectedFrom=fulltext

It is possible to organize all target solution into one vector by a proper handler. For now, ArrayPartition does not work with all the ODE algorithm.

Implementation based on

Liu, Yajing, and James R. Rice. 2005. “Aseismic Slip Transients Emerge Spontaneously in Three-Dimensional Rate and State Modeling of Subduction Earthquake Sequences.” Journal of Geophysical Research: Solid Earth 110 (B8). https://doi.org/10.1029/2004JB003424.

Need to decouple mesh, assemble, greens_function etc.