Major issue in the ordering of Predict and InitializeSolutionStep (and of ConvergenceCriterion initialization)

[SADPR]

Description

We are experiencing convergence issues when using a residual-based criterion in a nonlinear dynamic simulation with the following configuration. The solver fails due to converge (including singular matrix), even under moderate prescribed displacements. Switching to a displacement-based criterion allows convergence for some steps, but the simulation still fails eventually (possibly due to inverted elements).

Setup

• Software versions:

  • KratosMultiphysics: master branch (latest)
  • GiD GUI: latest distributed version

• Problem configuration:

  • Domain: 3D beam of length 43.2 m, approximate width: ~2 m, thickness: ~0.3–0.5 m
  • Element: Solid total Lagrangian
  • Constitutive law: Hyperelastic NeoHookean (HyperElastic3DLaw), also with KirchhoffSaintVenant3DLaw
  • Material: Steel
    • Density: 7850 kg/m³
    • Young's Modulus: 206.9 GPa
    • Poisson Ratio: 0.29
  • Analysis type: Dynamic nonlinear
  • Time integration: Implicit Newmark
  • Boundary conditions:
    • Left end: Fully fixed (Dirichlet = 0)
    • Right end: Prescribed displacement in the X direction: $$u_x(t) = t$$ with t being the current time in seconds.

Observed behavior

• Using residual-based convergence criterion:

  • The simulation fails to converge at the first step.

• Using displacement-based criterion:

  • Simulation progresses for several time steps.
  • Eventually fails — possibly due to element inversion or mesh distortion.

[SADPR]

Issue.zip

[SADPR]

Displacement Criterion:

[AlejandroCornejo]

Hi!

I've downloaded the files and managed to reproduce your problem. Let me cc @loumalouomega just in case he has some extra comments on this.

Increasing the echo_level to 1, we can see how the initial Obtained ratio = 4.96964e+11, so even though it manages to reduce it, it gets stucked in Obtained ratio = 70.5188. This is an issue that I identified in the past and, due to lack of time, I ended up increasing the abs norm tolerance, since this value is relatively low (Absolute norm = 3.9567e-06).

However, I feel that the calculation of the initial residual (the one with respect we compute the first ratio) may be not properly computed somehow. Maybe, since it is a nonlinear simulation, this is the best we can do, what do you think @loumalouomega ?

An additional tip, if you use the bossak scheme instead of newmark, besides the first step (similar behaviour), the remaining steps converge properly.

[loumalouomega]

Hello, I am out of office until Wednesday. My first guess was going to test with the UL element, to check if the issue may come with a error in the reference position calculation. Your checks help too.

[matekelemen]

Just here to mention that I think deciding when and how to compute the initial residual of convergence criteria is something we should take a second look at.

Right now, all coupled analyses use incorrect residual ratios (#12394). ConvergenceCriteria computes its initial residual in InitializeSolutionStep, which only gets called once per time step. However, strong coupling runs SolveSolutionStep multiple times within the same time step, effectively launching nonlinear iterations with a completely unrelated initial residual.

FYI @juancamarotti

[loumalouomega]

I checked my idea, with the UL element it works without any issue, but it is strange it does not work and converges for TL... I will do some further testing.

[loumalouomega]

Just here to mention that I think deciding when and how to compute the initial residual of convergence criteria is something we should take a second look at.

Right now, all coupled analyses use incorrect residual ratios (#12394). ConvergenceCriteria computes its initial residual in InitializeSolutionStep, which only gets called once per time step. However, strong coupling runs SolveSolutionStep multiple times within the same time step, effectively launching nonlinear iterations with a completely unrelated initial residual.

FYI @juancamarotti

I agree, but this is a different issue. We must update the NR for this and the discussion will be painful IMHO, I recommend to open a new discussion just for this.

[loumalouomega]

One more comment, using the skyline linear solver (the worst one) it crashes due to singular matrix.

[loumalouomega]

The local RHS for UL and TL are the same in first 2 NL iterations, strange that the residual criterion diverges...