Advance V5 nonlinear benchmarks and shell solver

This commit is contained in:
shenw
2026-06-27 16:43:00 +08:00
parent a2a11179d9
commit db68340aaf
84 changed files with 38720 additions and 971 deletions
+151 -28
View File
@@ -1,41 +1,164 @@
# Mohr-Coulomb Strength Reduction Benchmark
This case is a compact slope-wedge benchmark for the V5 Mohr-Coulomb strength
reduction workflow. The structural input uses a real `*MOHR COULOMB` material
with friction angle, cohesion, dilation angle, tensile cutoff, residual
parameters and an initial reduction factor.
This directory implements the `models.geomech.strength_reduction_method.pdf`
embankment benchmark as a full reference example, not as a smoke case. The
checked-in input is generated from the PDF geometry and uses the PDF mesh size
limit of 0.75 m.
Reference model:
- Geometry: `L1 = 85 m`, `L2 = 20 m`, `H1 = 20 m`, `H2 = 10 m`.
- Slope angles: `alpha = 15, 25, 35, 45 deg`.
- Material: `E = 20 MPa`, `nu = 0.3`, `c = 20 kPa`, `phi = 25 deg`,
`rho = 1940 kg/m3`.
- Material case 1: nonassociative flow with `psi = 0 deg`, mapped by Davis
procedure B to an equivalent associated material before strength reduction.
- Material case 2: associative flow with `psi = phi = 25 deg`.
- Stop criterion: `max displacement > Lslope / 20`.
- Boundary conditions: all nodes are constrained in the out-of-plane `y`
direction and in unused rotational DOFs 4-6; the base is fixed in
translations and the two vertical sides are fixed in `x`.
Files:
- `strength_reduction_slope_static.inp`: C3D4 slope wedge with `MC_SOIL`.
- `check_strength_reduction_results.py`: CTest checker for the benchmark
report.
- `strength_reduction_slope_static.inp`: full C3D8 solid embankment model for
the PDF default `alpha = 45 deg` geometry, using native `*DLOAD, GRAV`
solid body force loading and a real `*MOHR COULOMB` material. The lower
rectangular foundation and the upper slope share interface nodes along
`z = H1`, so the checked mesh is conforming rather than two disconnected
soil blocks.
- `strength_reduction_reference_cases.csv`: digitized/reference FOS values used
for the two material cases and four slope angles.
- `generate_strength_reduction_method.py`: regenerates the checked-in input.
- `check_strength_reduction_results.py`: CTest checker for the full PDF
reference sweep.
- `run_case.bat`: double-click runner that executes FemLite and then generates
strength-reduction reports.
benchmark reports.
Generated benchmark outputs:
- `strength_reduction_history.csv`: trial factor, residual, displacement
probe, plastic share, metric and bracket status.
- `strength_reduction_summary.csv`: final bracket and converged factor.
- `plastic_zone.csv`: Mohr-Coulomb branch activation for representative slope
probes at the converged reduction factor.
- `displacement_probe.csv`: crest vertical and toe horizontal displacement
control probes.
- `strength_reduction_path.csv`: trial factor, residual, displacement criterion,
plastic-share diagnostic, metric and search status for each alpha/material
case.
- `strength_reduction_cases.csv`: computed and reference FOS for all 8 PDF
cases.
- `strength_reduction_summary.csv`: aggregate convergence and maximum FOS error.
- `plastic_zone.csv`: Mohr-Coulomb branch activation at representative slope
probes for every case.
- `displacement_probe.csv`: crest-down and toe-out displacement probes for every
case.
- `strength_reduction_nonlinear_chain.csv`: optional real FE nonlinear trial
records written when the benchmark tool is launched with `--nonlinear-chain`.
Each row contains the selected nonlinear continuation method, strength
factor, nonlinear solver error code, convergence status/failure code, cutback
count, Newton iteration count, target/actual/next continuation increment,
adaptive adjustment reason, trust radius, damping factor, residual merit
ratio, trust rejection count, residuals, maximum nodal displacement and
plastic integration-point share. Nonlinear chain/progress rows also aggregate
Mohr-Coulomb integration-point diagnostics: point count, updated point count,
accepted/requested substeps, limited return count, branch mask, maximum yield
residual and maximum plastic increment.
- `strength_reduction_nonlinear_summary.csv`: optional summary for the same
real nonlinear trial chain.
- `strength_reduction_cloud_summary.csv`: optional target-FOS nonlinear trial
summary written with `--nl-write-cloud`.
- `result_00001525.vtk`: optional target-FOS cloud VTK for the PDF
`FOS = 1.525` case. It contains nodal `displacement` vectors and element
`eqps` scalars.
The runner intentionally does not generate `result_00000001.vtk`. That file was
previously produced by a linear static gravity precheck and can show
nonphysical displacements around 200 m, while the PDF reference displacement
plot for material case 2, `alpha = 45 deg`, `FOS = 1.525`, is about 0.3 m. A
real nonlinear Mohr-Coulomb trial chain is available through the optional
`--nonlinear-chain` route. The current chain ramps external loads during static
increments, writes cutback/accepted-time/linear-solver diagnostics and can form
a real stable/unstable bracket on coarse conforming probes. A separate
`--nl-write-cloud` route runs one real target-FOS trial and writes a VTK cloud
at `result_00001525.vtk` by default. The cloud route now uses adaptive
continuation, optional displacement prediction, and internal BiCGSTAB/ILU0
preconditioning when requested. It refreshes the VTK after each accepted outer
segment, so interrupted long runs still leave the latest nonlinear
displacement/eqps cloud. That file is the intended artifact for the PDF
material case 2, `alpha = 45 deg`, `FOS = 1.525` displacement/eqps comparison.
Passing the file-structure check is not the same as matching the PDF cloud;
the 0.75 m long-run still has to demonstrate a stable path to full load and
the expected approximately 0.3 m displacement scale.
The production INP parser accepts
`*NONLINEAR CONTINUATION, METHOD=value`; if the keyword is omitted the selected
method is `1` (`TWO_STAGE_LOAD_STRENGTH`). Method 1 now runs a two-stage
load-strength route in the nonlinear chain and target cloud paths: stage A
converges the full gravity/load state at strength factor `F = 1.0`, and stage B
holds the external load fixed while the material strength reduction factor is
continued to the requested trial factor. The nonlinear chain, cloud summary and
cloud progress CSV files record this method id so long runs remain auditable.
Method 2 (`ADAPTIVE_TARGET_NEWTON`) records target/actual/next increment and
adjust reason; method 3 (`TRUST_REGION_DAMPED_NEWTON`) records trust radius,
damping, residual merit ratio and rejection count. Method 5
(`MC_SUBSTEPPING_SEMISMOOTH`) records local Mohr-Coulomb substep and active-set
diagnostics in both summary and row-wise nonlinear outputs. The checker rejects
missing or invalid continuation diagnostics.
The full 0.75 m `FOS = 1.525` cloud still needs a documented longrun to reach
the PDF displacement/eqps target; NC-4/NC-5 and the longrun close-out track the
remaining stabilization work.
Acceptance checks:
- The input must contain `*MOHR COULOMB` and must not use the old elastic-soil
placeholder.
- The strength-reduction search must converge with safety factor in
`[1.20, 1.55]`.
- The history must contain both stable and unstable trials, forming a real
bracket.
- At least two representative slope probes must activate a Mohr-Coulomb
plastic branch at the converged factor.
- The displacement probes must keep the expected crest-down and toe-out signs.
- The input must record the PDF geometry, 0.75 m mesh size, default 45 degree
slope, conforming lower/upper interface, native `*DLOAD, GRAV` gravity route
and `*MOHR COULOMB` material.
- All 8 PDF cases must converge and stay within 0.02 FOS of
`strength_reduction_reference_cases.csv`.
- FOS must decrease monotonically as slope angle increases.
- The associative case must be at least as strong as the Davis-B
nonassociative case at each slope angle.
- Each case must contain a real stable/unstable bracket, plastic-zone probe
activation and crest-down/toe-out displacement signs.
- The checker deliberately validates the FOS curve and diagnostic CSV files
by default. If `strength_reduction_cloud_summary.csv` exists, it also checks
that the nonlinear target cloud VTK exists, contains displacement and eqps
fields, matches the recorded mesh counts, and carries valid MC substep/branch
diagnostic summaries.
The benchmark report is generated by `tools/benchmarks/strength_reduction_benchmark.f90`,
which links the production FemLite core and calls the Mohr-Coulomb and
strength-reduction driver/report modules. The helper is intentionally outside
`src/` because it is benchmark orchestration, not solver production code.
The benchmark report is generated by
`tools/benchmarks/strength_reduction_benchmark.f90`, which links the production
FemLite core and calls the production Mohr-Coulomb material and
strength-reduction search driver. The helper remains outside `src/` because it
is benchmark orchestration, not solver production code. Gravity is now applied
through native `*DLOAD` label `GRAV` as `magnitude + direction`, so the checked
input no longer contains generated equivalent nodal `*CLOAD` rows. The
benchmark executable creates the requested output directory when it does not
already exist, so direct command-line probes do not require a separate `mkdir`
step.
Manual nonlinear-chain probe:
```bat
build-win-gcc-16\strength_reduction_benchmark.exe ^
examples\strengthReduction\strength_reduction_slope_static.inp ^
build-win-gcc-16\strength_reduction_nl_probe ^
--nonlinear-chain --nl-max-trials 10 --nl-increments 100 --nl-max-newton 80
```
Manual target-FOS cloud probe:
```bat
build-win-gcc-16\strength_reduction_benchmark.exe ^
examples\strengthReduction\strength_reduction_slope_static.inp ^
build-win-gcc-16\strength_reduction_pdf_cloud ^
--nl-write-cloud --nl-cloud-factor 1.525 --nl-cloud-step 1525 ^
--nl-increments 100 --nl-max-newton 20 --nl-linear-iter 300 ^
--nl-tol 1.0e-3 --nl-no-line-search --nl-preconditioner ilu0
```
This route is intentionally not part of the default CTest acceptance yet. It is
the real solver audit path used to expose convergence status, load-step
cutbacks, accepted time and linear solve diagnostics while V5.4 continues
toward full 0.75 m continuation and reference displacement/eqps VTK
reproduction. Current 0.75 m probes show real accepted progression through the
full mesh, including a successful jump from `accepted_time=0.003` to
`0.053`, and later to `0.0905`; the displacement scales from `9.7294e-4 m` at
`t=0.003` to `2.9383e-2 m` at `t=0.0905`, consistent with the PDF
approximately `0.3 m` target scale, but a completed `accepted_time=1.0` run is
still open.