Huanxia Wei

Huanxia Wei

HKUST, NUS, Tongji

Groetzbach Verification Output Guide

This guide tells an agent how to interpret outputs from run_groetzbach_validation_standalone.py.

Primary Files

Given an output directory, read these first:

  • groetzbach_summary.json: authoritative run metadata, statistics, failure fractions, refinement suggestions, and visualization file paths.
  • groetzbach_metrics.csv: compact table of scalar statistics for ratio_max, ratio_i, ratio_j, ratio_k, eta, epsilon, and delta_max.
  • groetzbach_worst_subcells.csv: worst local subcells ranked by ratio_max.
  • check_fields/metadata.json: shape, array names, and meaning of complete subcell fields.
  • visualization_3d/: full-domain isosurface .vtp files and PNG views.
  • plots_2d/: histograms, directional CDFs, profiles, and projected max-ratio maps.
  • groetzbach_validation_*.run.log: console transcript from the run.

Method To Report

State the method explicitly:

  • The script validates one instantaneous Nek field, not an averaged field, unless the input file itself is an averaged field.
  • It reads velocity U; coordinates X come from the same file when present, otherwise from the sequence head or an explicitly supplied coordinate field.
  • Velocity gradients are evaluated spectrally on GLL nodes inside each spectral element.
  • The physical velocity gradient is obtained by inverting the coordinate Jacobian.
  • Dissipation is epsilon = 2 * nu * Sij * Sij.
  • Kolmogorov length is eta_K = (nu^3 / epsilon)^0.25.
  • Each GLL subcell is checked with local edge lengths:
    • ratio_i = Delta_i / eta_K
    • ratio_j = Delta_j / eta_K
    • ratio_k = Delta_k / eta_K
    • ratio_max = max(Delta_i, Delta_j, Delta_k) / eta_K
  • The refined Groetzbach SEM DNS threshold is ratio <= pi.

Use the words local i, local j, and local k. Do not call them global x, y, or z directions unless a separate geometric mapping analysis has been done.

Verdict Fields

In groetzbach_summary.json, use:

  • run.field: input field.
  • run.coord_field: coordinate source.
  • run.time: nondimensional time from the Nek header.
  • run.nu and run.re_r: viscosity and Reynolds number assumption.
  • run.elements_processed: processed spectral elements.
  • accumulator.valid_subcells: valid tested GLL subcells.
  • accumulator.bad_jacobian_nodes: mesh/metric quality warning.
  • accumulator.fail.max.threshold.count_fraction: fraction of subcells with ratio_max > pi.
  • accumulator.fail.max.threshold.volume_fraction: volume fraction with ratio_max > pi.
  • accumulator.metrics.max.p99_volume: volume-weighted p99 of ratio_max.
  • accumulator.metrics.max.max: p100/max of ratio_max.

Recommended conclusion format:

SEM DNS threshold: Delta_max / eta_K <= pi.
The case fails/passes by the strict p100 criterion.
At pi, failing count fraction is ... and failing volume fraction is ....
ratio_max p99_volume is ..., and p100/max is ....
Bad Jacobian nodes: ....

Refinement Suggestions

Use refinement_suggestions in groetzbach_summary.json.

For a p99 recommendation, prefer p99_volume_plan unless the user explicitly asks for count-weighted p99. Report:

  • continuous_direction_factors.i
  • continuous_direction_factors.j
  • continuous_direction_factors.k
  • integer_direction_factors
  • estimated_spectral_elements_integer
  • estimated_stored_gll_points_integer

For strict p100, use p100_plan.

Interpretation:

  • A factor of 1.0 means no extra refinement is required in that local direction for the selected target.
  • A factor of 2.0 means the local spacing should be roughly halved in that local direction.
  • Integer estimates multiply the current spectral-element count by the integer directional factors. They are conservative and can become very large.

Complete Field Arrays

check_fields/metadata.json defines the complete subcell arrays.

The normal shape is:

[element, k_subcell, j_subcell, i_subcell]

Important arrays:

  • x.npy, y.npy, z.npy: subcell center coordinates.
  • epsilon.npy: dissipation.
  • eta.npy: Kolmogorov length.
  • delta_i.npy, delta_j.npy, delta_k.npy: local subcell edge lengths.
  • ratio_i.npy, ratio_j.npy, ratio_k.npy, ratio_max.npy: Groetzbach ratios.
  • volume.npy: subcell volume approximation.
  • dominant_direction.npy: 0=invalid, 1=local_i, 2=local_j, 3=local_k.

Use these arrays when the user asks where the failure is located, which direction dominates, or wants new custom plots.

3D Visualization

The main full-domain 3D outputs are in visualization_3d/.

The isosurface levels are:

  • half_threshold: 0.5 * pi
  • threshold: pi
  • double_threshold: 2 * pi
  • quad_threshold: 4 * pi

The rendered PNGs use the Q-criterion suite style:

  • white background
  • ground plane
  • domain grid
  • hemisphere body when inside bounds
  • full domain box outline
  • parallel projection
  • small native axes viewport
  • semi-transparent nested threshold surfaces

The first three full-domain views normally to inspect are:

  • groetzbach_ratio_max_threshold_levels_parallel.png
  • groetzbach_ratio_max_threshold_levels_parallel_head_to_tail.png
  • groetzbach_ratio_max_threshold_levels_parallel_top.png

Also check side/upstream/downstream views when explaining spatial distribution.

Caveats

  • The 3D surfaces are generated from the voxelized ratio_max grid saved as groetzbach_ratio_max_voxel_grid.npz and .vti; the complete original subcell data are in check_fields/.
  • Histogram p99 values are binned estimates. The p100/max statistic is from the streaming exact maximum.
  • If the input field was produced with regularization, this validation checks the resolved field against the Groetzbach spacing criterion; by itself it does not prove the simulation is an unregularized DNS.
  • If a user asks whether the data are instantaneous or averaged, answer from the actual input file and run.method; do not infer it from file names alone.

Example From The 260606 Run

For the run on E:\2_Cases\1_CFD\1_Basic_Flow_Mechanics\Hemisphere\5_Results\260606_CPU_SentToArrhenius_1370\hemi_restart0.f00000, the completed output in Z:\Hemisphere\0_Scratch reported:

  • time: t* = 550
  • spectral elements: 189,114
  • valid subcells: 64,866,102
  • bad Jacobian nodes: 0
  • ratio_max > pi count fraction: 12.054433%
  • ratio_max > pi volume fraction: 10.561990%
  • ratio_max volume p99: 6.2194216
  • ratio_max p100/max: 16.226545
  • p99-volume refinement factors: i x1.816, j x1.666, k x1.000
  • p100 refinement factors: i x5.165, j x2.560, k x2.821