Skip to content

uq-poiseuille2d

English | Japanese

« Home · « Optimization, inverse problems and uncertainty

Forward uncertainty quantification (UQ) around the plane-Poiseuille LBM forward map. Propagates uncertainty in the kinematic viscosity nu through a REAL body-force-driven D3Q19 BGK+Guo LBM solve (the same kernel as benchmarks/poiseuille_d3q19.cpp) and validates the resulting distribution of the QoI against the EXACT analytic pushforward.

Plots

Time series and distributions of the QoI written by this computation. Where a reference value exists it is overlaid as a horizontal line.

uq-poiseuille2d mc-convergence

uq-poiseuille2d samples

Build & run

cmake -S native -B native/build && cmake --build native/build -j
./native/build/orheo-uq-poiseuille2d

uq_poiseuille2d [N_MC] [seed] [out_prefix]

Cost of the computation that made the figures on this page (Miyabi-G Grace CPU (72 cores, -stdpar=multicore)): 462.6 s wall.

Validation

Tier T0 — order of convergence verified
QoI qoi_mean, qoi_std
Reference forward UQ: MC vs quadrature pushforward of the exact map
Gate — no gate

Source & run details

View source

Item Value
Target orheo-uq-poiseuille2d
Source native/benchmarks/uq_poiseuille2d.cpp
Lattice D3Q19
GPU twin no
Modules lbm_d3q19.hpp
Gallery run 2026-07-27 @ 3456834e
Args 64 20260716 ./uq_
Wall time 462.65 s
CPU time 16489.62 s
Device Miyabi-G Grace CPU (72 cores, -stdpar=multicore)

Computation log

The actual output of this computation (head and tail).

=== UQ Poiseuille (forward MC vs analytic pushforward) ===
channel N=32 H=32  Gx=1.9531e-05  K=G H^2/8=2.5000e-03
input: ln(nu)~Normal(m=-2.99573, s=0.150)  => nu median=0.05000  CV(nu)=15.1%
QoI:   u_max = K/nu  (median u_max=0.05000)  low-Mach target=0.050
analytic pushforward u_max~LogNormal(muL=-2.99573, s=0.150):  mean=0.050566  std=0.007628
seed=20260716  N_MC=64  solver stop: |du|/u<1e-07 or 60000 steps

[calib] LBM(nu=median): u_max=0.049908 vs exact K/nu=0.050000  bias=-0.184%  (steps=31000)  Q=1.0658 vs (2/3)u_max H=1.0667

[collocation] 7-node Gauss-Hermite over ln(nu):
  node 0  nu=0.02849  u_max=0.087590  (w=5.4827e-04)
  node 1  nu=0.03506  u_max=0.071175  (w=3.0757e-02)
  node 2  nu=0.04205  u_max=0.059342  (w=2.4012e-01)
  node 3  nu=0.05000  u_max=0.049908  (w=4.5714e-01)
...

=== MC convergence (block-RMS vs analytic) ===
   Nb  blocks   rms|mean_err|   rms|std_err|
    8       8   2.657994e-03   2.166649e-03
   16       4   1.457116e-03   1.235039e-03
   32       2   1.061922e-03   8.210488e-04
   64       1   9.698756e-04   5.678334e-04

wrote ./uq_samples.csv (64 rows), ./uq_convergence.csv, ./uq_summary.csv
total LBM forward solves: 64 (MC) + 7 (collocation) + 1 (calib) = 72