P92 nonuniform structural screen¶
Purpose and scope¶
Exercise the existing physical-time user-material consumer in bending and stress redistribution, beyond homogeneous patches. The material is the research-only isotropic tensor extension of the fixed 650 C parameters in Zhang et al., Metals 10 (2020) 307. The supplied Poisson ratio is 0.3. This is not the separate 2019 creep-fracture law, a new calibration, or a component-life prediction.
The solver, accepted state and checkpoint belong to the ordinary model.step
workflow. The repository tool supplies geometry, loading and measurements;
it does not implement a second global solver.
Test definition¶
A 4 x 1 x 1 mm solid is clamped on x=0. Its x=4 face has prescribed y displacement, with the other components free. Displacement rises linearly to 0.02 mm in 4 s and is held for 60 s. The temperature is fixed at 650 C. This is a small-strain test, not a finite-rotation benchmark.
Hexahedral meshes have (4r, r, r) cells, for r=1,2,4,8. The base physical
time grid has eight ramp increments and twelve hold increments. Time
refinement doubles both counts without changing the waveform or duration.
Each solve is serial and limits numerical library threads to one.
Before interpreting the fine-grid result, set the numerical screening goals: the final mesh doubling changes ramp-end and hold-end reactions by less than 5%; time-step halving at a fixed mesh changes them by less than 1%. These are bounded numerical screening goals, not estimates of exact error or engineering allowables. Report the actual differences even if a goal fails. Do not tune material parameters to make them pass.
Measure reactions at the constrained support and loaded component separately. Their vector sum tests force equilibrium. Summing the unconstrained internal force over every node alone is insufficient: internal assembly may cancel globally even when the nonlinear equations have not converged. Also retain the accepted Newton residual. The clamp has boundary-condition transitions; maximum sampled stress is diagnostic, not a converged hotspot quantity.
The selected material-point history uses cell 0 in each mesh; it is useful for same-layout restart, not a common physical point for cross-mesh comparison. Restore at the accepted end of the ramp into a fresh Step and require identical selected histories and endpoint observables.
Reproduction¶
With the repository research tools and AgentFEM available:
python -m tools.p92_nonuniform_structure --resolution 1 --checkpoint-dir bending-restart --output bending-r1.json
python -m tools.p92_nonuniform_structure --resolution 4 --refinement 2 --output bending-r4-t2.json
Use new paths: evidence is not overwritten. The 2048-cell mesh is an explicit manual screen, not part of every CI test. CI uses the four-cell composition and restart test. Run numerical screens sequentially and retain wall times without calling them a general performance comparison.
Reference practice¶
Abaqus creep integration verification separates constitutive integration checks against exact solutions from coupled loading tests, and distinguishes stable integration from time accuracy. This screen follows that separation: existing independent scalar/local checks are retained; mesh refinement here does not replace experimental validation.
Recorded numerical evidence¶
The six raw reports are retained in
evidence/creep/2026-11-07-p92-nonuniform.json, including tool identity,
constraint reactions, accepted residuals and wall times.
| Cells | Increments | Ramp-end reaction (N) | Hold-end reaction (N) |
|---|---|---|---|
| 4 | 20 | 15.837950 | 15.817883 |
| 32 | 20 | 11.787592 | 11.575751 |
| 256 | 20 | 10.596423 | 10.236817 |
| 2048 | 20 | 10.262701 | 9.854165 |
| 256 | 40 | 10.605809 | 10.234115 |
| 2048 | 40 | 10.273501 | 9.851034 |
Relative differences use the finer result as denominator. The last mesh doubling changes total reactions by 3.25% and 3.88%; time refinement at 2048 cells changes them by 0.105% and 0.032%. These pass the stated bounded total-reaction screens. The four-cell interrupted run reproduces selected histories and endpoint observables exactly; other runs did not test restart.
The relaxation drop itself changes by about 12% between the last two meshes at 20 increments (fine-drop denominator), so its spatial convergence remains open. The maximum sampled clamp stress also continues to grow. Neither quantity is certified by the total-reaction screen. Stop this mesh ladder here rather than treating progressively larger solves as experimental validation. Engineering multiaxial material validation remains a separate gate.