PETSC ERROR in coupling Poroelasticity and coseismic processes

PyLith
1

Dear Aagaard,

I encountered some issues while using PyLith 4.2.1 to simulate coseismic poroelastic changes and would like to seek your advice.

In my model, I defined a fault with dimensions of 100 km × 20 km within a 300 km × 300 km × 40 km domain and prescribed a 2 m dislocation on the fault. I used SolnDispPresTracStrainLagrange to investigate the coupled effects of coseismic deformation and pore-pressure diffusion.

However, during the second stage of the simulation, the calculation failed and produced an error. I have attached the model setup and error information for reference.

Best regards.

image
image1809×409 25.8 KB

model.zip (111.2 KB)

2

We have found this error when the initial guess is not accurate. The workaround is to adjust the divergence tolerance for the linear solver:

[pylithapp.petsc]
# Increase divergence tolerance. Initial guess at second time step is not accurate.
ksp_divtol = 1.0e+6

This is discussed in Step2 of the magma-2d example.

3

Thanks for your reply. I adjusted the divergence tolerance, but the same errors still occur. The settings I used for this simulation are as follows:

[pylithapp.petsc]
ts_type = beuler
pc_type = fieldsplit
pc_fieldsplit_type = multiplicative
pc_fieldsplit_0_fields = 2
pc_fieldsplit_1_fields = 1
pc_fieldsplit_2_fields = 0
pc_fieldsplit_3_fields = 3

fieldsplit_trace_strain_pc_type = bjacobi
fieldsplit_pressure_pc_type = bjacobi

fieldsplit_displacement_pc_type = ml
fieldsplit_displacement_ksp_type = gmres
fieldsplit_displacement_mg_fine_pc_type = vpbjacobi

fieldsplit_lagrange_multiplier_fault_pc_type = jacobi

dm_reorder_section = True
dm_reorder_section_type = cohesive

snes_rtol = 1.0e-6
snes_atol = 1.0e-8
ksp_rtol = 1.0e-8
ksp_atol = 1.0e-10

snes_converged_reason = true
ksp_converged_reason = true

ksp_divtol = 1.0e+5

Under the same configuration, the simulation runs successfully for an elastic material model [step01 in model.zip]; however, it fails to converge when using a poroelastic formulation.

4

A couple options:

  1. Update to the new release (v5.0.0; we are doing a soft rollout for a workshop next week). We have improved the default solver settings and nondimensionalization.
  2. Increase ksp_gmres_restart to avoid reduce how often GMRES orthogonalizes vectors against the previous ones using a Gram-Schmidt process.
[pylithapp.petsc]
# Try increasing to half the number of KSP iterations or even a bit larger
ksp_gmres_restart = 300