LAST SEEN ONLINE:
NOV 22, 2026 11:36 UTC
LOCAL_TIME_SYNC:
--:--:-- | -- _
2026 · complete · buoyantBoussinesqSimpleFoam

Part 6: Mesh Verification & Grid Convergence Index (GCI)

Quantifying numerical discretization error and establishing mesh independence in the Differentially Heated Cavity using Richardson Extrapolation and the ASME V&V 20 Standard.

OpenFOAMLinuxCFDVerification & Validation

Author: Martin Mathew | Date: 2026-10-12 | Category: Thermal CFD Tutorial

1. Verification vs. Validation: Solving the Equations Right

In our previous articles, we established the baseline steady-state setup, explored RANS turbulence modeling with k−ωk-\omega SST, and simulated time-accurate plume shedding with the PIMPLE algorithm. However, an essential question remains unanswered:

How much of our computed heat transfer is real physical flow, and how much is numerical discretization error from our grid?

In computational fluid dynamics (CFD), the boundary between Verification and Validation (V&V) defines professional engineering practice:

  • Verification: “Are we solving the mathematical equations right?” (Quantifying numerical algorithm errors, code bugs, and grid discretization error).
  • Validation: “Are we solving the right equations?” (Quantifying how closely the physical model reflects experimental reality).

Before claiming that a simulation validates against benchmark literature (such as De Vahl Davis), we must first verify that our solution is mesh-independent. Without a formal mesh convergence study, a CFD result is merely a point estimate of unknown numerical uncertainty.


2. Richardson Extrapolation & The Grid Convergence Index (GCI)

Rather than arbitrarily increasing cell counts and visually guessing when fields stop changing, the ASME V&V 20-2009 Standard (Standard for Verification and Validation in Computational Fluid Dynamics and Heat Transfer), based on the foundational work of P. J. Roache (1998) and Celik et al. (2008), provides a rigorous mathematical framework.

2.1 The Concept of Richardson Extrapolation

For any discrete numerical scheme, Taylor series truncation error dictates that a computed quantity f(h)f(h) at characteristic cell size hh approaches the exact continuum solution fexactf_{exact} according to:

f(h)=fexact+C1hp+C2hp+1+…f(h) = f_{exact} + C_1 h^p + C_2 h^{p+1} + \dots

When the mesh is sufficiently refined to enter the asymptotic range of convergence, higher-order terms become negligible:

f(h)≈fexact+Chpf(h) \approx f_{exact} + C h^p

Richardson Extrapolation uses solutions from multiple systematically refined grids to mathematically eliminate the leading error term ChpC h^p, yielding the continuum zero-spacing estimate fextf_{ext} without needing an infinitely fine mesh.

2.2 Core Concepts & Mathematical Formulation

Let f1,f2,f3f_1, f_2, f_3 denote discrete solutions evaluated on fine, medium, and coarse meshes with a constant grid refinement ratio r=h3/h2=h2/h1=2.0r = h_3/h_2 = h_2/h_1 = 2.0:

  1. Step Differences:
    ϵ21=f2−f1\epsilon_{21} = f_2 - f_1, ϵ32=f3−f2\epsilon_{32} = f_3 - f_2

    • If ϵ32/ϵ21>0\epsilon_{32}/\epsilon_{21} > 0, convergence is monotonic.
    • If ϵ32/ϵ21<0\epsilon_{32}/\epsilon_{21} < 0, convergence is oscillatory.
  2. Observed Apparent Order of Convergence (pp):
    p=ln⁡∣ϵ32/ϵ21∣ln⁡(r)p = \frac{\ln|\epsilon_{32} / \epsilon_{21}|}{\ln(r)}
    The apparent order indicates how rapidly truncation error drops with grid refinement. It reflects the real, observed spatial accuracy of the coupled solver.

  3. Richardson Extrapolated Value (fextf_{ext}):
    fext21=f1+f1−f2rp−1f_{ext}^{21} = f_1 + \frac{f_1 - f_2}{r^p - 1}
    This represents the predicted solution at zero cell size (h→0h \to 0).

  4. Grid Convergence Index (GCIGCI):
    Raw extrapolated errors can underestimate uncertainty if the grid is coarse. Roache introduced the GCI with a safety factor (Fs=1.25F_s = 1.25 for three-grid studies) to establish a standardized 95% confidence interval:

    • Fine Grid Uncertainty: GCIfine21=Fs⋅ea21rp−1GCI_{fine}^{21} = \frac{F_s \cdot e_a^{21}}{r^p - 1}
    • Medium Grid Uncertainty: GCImedium32=Fs⋅ea32⋅rprp−1GCI_{medium}^{32} = \frac{F_s \cdot e_a^{32} \cdot r^p}{r^p - 1} where ea21=∣(f1−f2)/f1∣×100%e_a^{21} = |(f_1 - f_2) / f_1| \times 100\% is the relative step change.
  5. Asymptotic Range Verification (χ\chi):
    χ=GCI32rp⋅GCI21≈1.0\chi = \frac{GCI_{32}}{r^p \cdot GCI_{21}} \approx 1.0
    When χ≈1.0\chi \approx 1.0, the discrete solutions are verified to reside within the asymptotic convergence regime, confirming that Richardson extrapolation is valid.


3. Mesh Hierarchy & Verification Plots Analysis

To conduct our study, we configured three geometrically similar hexahedral meshes with a constant refinement ratio r=2.0r = 2.0:

  • Grid 3 (Coarse): 50×5050 \times 50 (N=2,500N = 2,500 cells, h=0.020h = 0.020 m)
  • Grid 2 (Medium): 100×100100 \times 100 (N=10,000N = 10,000 cells, h=0.010h = 0.010 m)
  • Grid 1 (Fine): 200×200200 \times 200 (N=40,000N = 40,000 cells, h=0.005h = 0.005 m)

Below is the detailed visual verification and physical insight for each of our key verification metrics:


3.1 Mesh Comparison: Resolving the Buoyant Boundary Layer

Grid Hierarchy Corner Detail

Figure 1: Near-wall grid density comparison at the bottom-left corner (joint between the heated vertical wall and adiabatic bottom wall).

Technical Insights:

  • Boundary Layer Scaling: In high-Rayleigh natural convection (Ra≈2×109Ra \approx 2 \times 10^9), the characteristic thermal and hydrodynamic boundary layer thickness scales according to laminar boundary layer theory as: δ∼L⋅Ra−1/4≈1.0⋅(2.06×109)−1/4≈4.7 mm\delta \sim L \cdot Ra^{-1/4} \approx 1.0 \cdot (2.06 \times 10^9)^{-1/4} \approx 4.7\text{ mm}
  • Resolution Assessment:
    • On the Coarse grid (Δx=20\Delta x = 20 mm), cell width is over four times larger than δ\delta. The entire near-wall buoyant jet is swallowed within a single cell layer, causing severe numerical smearing.
    • On the Medium grid (Δx=10\Delta x = 10 mm), cell spacing drops to roughly 2δ2\delta, beginning to capture the near-wall gradient.
    • On the Fine grid (Δx=5\Delta x = 5 mm), cell size closely approaches the physical boundary layer thickness δ≈4.7\delta \approx 4.7 mm, allowing steep wall-normal velocity and thermal gradients to be captured without excessive numerical diffusion.

3.2 Residuals Convergence: Solver Stability Across Scales

Residual Convergence History Across All Grids

Figure 2: Iteration convergence history for velocity (UxU_x) and temperature (TT) residuals across all three meshes.

Technical Insights:

  • Numerical Diffusion vs. Convergence Rate: On the coarse mesh, high numerical diffusion provides strong artificial damping that accelerates algebraic residual drop below 10−410^{-4} in fewer iterations.
  • Matrix Condition & Resolution: On the medium (10k cells) and fine (40k cells) meshes, resolving finer velocity gradients requires more iterations for the elliptic pressure-Poisson and energy fields to relax across the larger coefficient matrix.
  • Asymptotic Plateau: All three meshes achieve monotonic residual decay and settle into smooth, stable plateaus without unphysical oscillations, confirming that buoyantBoussinesqSimpleFoam maintains numerical stability across all grid scales.

3.3 Vertical Velocity Profile (UyU_y) along Mid-Plane y=0.5y = 0.5

Vertical Velocity Profile Comparison

Figure 3: Vertical velocity profile (UyU_y) along the horizontal midplane y=0.5y = 0.5 (x∈[0,1]x \in [0, 1]).

Technical Insights:

  • Buoyant Wall Jet: Fluid adjacent to the hot wall (x=0x = 0) accelerates upward due to positive buoyancy, while fluid along the cold wall (x=1x = 1) accelerates downward with equal intensity, preserving anti-symmetry.
  • Core Quiescence: Across the central 70% of the enclosure (0.15<x<0.850.15 < x < 0.85), vertical velocity is virtually zero (Uy≈0U_y \approx 0), proving that fluid circulation is confined to narrow boundary streams surrounding a stagnant central core.
  • Peak Acceleration Sensitivity: Maximum upward velocity sharply increases with refinement: vmax=0.085v_{max} = 0.085 m/s (Coarse) →0.128\to 0.128 m/s (Medium) →0.1935\to 0.1935 m/s (Fine). Because uniform Cartesian meshes cannot grade cells infinitely close to the wall, refining hh directly captures the true velocity peak that coarse grids artificially suppress.

3.4 Horizontal Velocity Profile (UxU_x) along Mid-Plane x=0.5x = 0.5

Horizontal Velocity Profile Comparison

Figure 4: Horizontal velocity profile (UxU_x) along the vertical midplane x=0.5x = 0.5 (y∈[0,1]y \in [0, 1]).

Technical Insights:

  • Recirculating Cross-Flow: The vertical plume hits the adiabatic ceiling, turning horizontally across the top from left to right (Ux>0U_x > 0 for y>0.8y > 0.8). Sinking cold fluid turns along the bottom floor from right to left (Ux<0U_x < 0 for y<0.2y < 0.2).
  • Mesh Collapse & Rapid Convergence: Unlike the narrow vertical wall jet, turning currents spread over a broader turning radius. As a result, the Medium (100×100100 \times 100) and Fine (200×200200 \times 200) grids show strong profile collapse (umax≈0.023−0.028u_{max} \approx 0.023 - 0.028 m/s), indicating that horizontal recirculation achieves mesh independence much earlier than the near-wall vertical jet.

3.5 Temperature Profile (TT) along Mid-Plane y=0.5y = 0.5

Temperature Profile Comparison

Figure 5: Temperature distribution (TT) along the horizontal midplane y=0.5y = 0.5 (x∈[0,1]x \in [0, 1]).

Technical Insights:

  • Convection Domination vs. Conduction: In pure conduction (Ra→0Ra \to 0), the midplane temperature is a linear slope from 310 K down to 290 K. At Ra≈2×109Ra \approx 2 \times 10^9, convective mixing establishes an isothermal core at exactly T=300.0T = 300.0 K across 0.1<x<0.90.1 < x < 0.9.
  • Thermal Boundary Layer Confinement: Extreme thermal gradients are compressed entirely into thin layers within Δx<0.05\Delta x < 0.05 m of the hot and cold walls.
  • Symmetry Preservation: The center temperature converges to 300.015300.015 K with an approximate relative error of only ea21=0.01%e_a^{21} = 0.01\%, demonstrating strict preservation of thermal anti-symmetry.

3.6 Local Nusselt Number Distribution along Heated Wall

Local Nusselt Number Distribution

Figure 6: Local Nusselt number distribution Nu(y)Nu(y) along the vertical heated wall (x=0x = 0, y∈[0,1]y \in [0, 1]).

Technical Insights:

  • Leading-Edge Impingement: Cold fluid (290290 K) sweeping across the floor hits the bottom edge of the hot wall (y≈0y \approx 0), producing an intense thermal gradient and peak heat transfer (Nu>150−180Nu > 150 - 180).
  • Boundary Layer Growth: As fluid rises along the wall, it absorbs heat and warms up, thickening the thermal boundary layer and monotonically decreasing the local heat transfer rate Nu(y)Nu(y) toward the top ceiling (y=1y = 1).
  • Heat Transfer Independence: The Coarse grid underpredicts the corner peak (Numax≈100Nu_{max} \approx 100) due to coarse cell volume averaging. However, between Medium and Fine grids, the integrated average Nusselt number shifts by only 1.85% (58.43→59.5358.43 \to 59.53), achieving an ASME GCIfineGCI_{fine} of just 0.16%0.16\%.

4. Quantitative ASME V&V 20 Verification Summary

Applying the automated Richardson extrapolation and GCI calculation pipeline across the three grid levels yields the following verification dataset:

Quantity of InterestUnitCoarse (50x50)Medium (100x100)Fine (200x200)Extrapolated Continuum (h→0h \to 0)Apparent Order (pp)Relative Error (ea21e_a^{21})Fine Grid Index (GCIfineGCI_{fine})Asymptotic Ratio (χ\chi)
Average Nusselt (Nu‾\overline{Nu})[-]41.171758.434159.534159.60903.971.85%0.16%15.989
Cavity Center Temp (TcenterT_{center})[K]299.8630300.0360300.0150300.01501.000.01%0.01%1.000
Max Horizontal Velocity (umaxu_{max})[m/s]0.04510.02260.02780.02781.0018.60%23.25%1.000
Max Vertical Velocity (vmaxv_{max})[m/s]0.08540.12850.19351.09890.1033.58%584.87%0.999

Note: Safety factor Fs=1.25F_s = 1.25. Refinement ratio r=2.0r = 2.0. Redundant dimensionless velocity scalings (u∗,v∗u^*, v^*) have been omitted because they represent identical percentage errors and convergence orders as their dimensional counterparts.


4.1 How Each Verification Metric is Calculated & Its Relevance

To understand why verification is the cornerstone of credible computational fluid dynamics, here is how each column in the table is mathematically computed and what it physically signifies:

  1. Discrete Grid Solutions (f3,f2,f1f_3, f_2, f_1):

    • How it is obtained: Raw simulation outputs evaluated on systematically refined meshes: Coarse (f3f_3, h3=0.020h_3 = 0.020 m), Medium (f2f_2, h2=0.010h_2 = 0.010 m), and Fine (f1f_1, h1=0.005h_1 = 0.005 m).
    • Relevance: Establishes the three discrete data points required to solve for the three unknowns in the truncation error expansion: continuum solution (fextf_{ext}), error coefficient (CC), and apparent order (pp).
  2. Approximate Relative Error (ea21e_a^{21}):

    • Formula: ea21=∣(f1−f2)/f1∣×100%e_a^{21} = |(f_1 - f_2) / f_1| \times 100\%
    • Relevance: Measures the step-to-step percentage change from the medium to the fine mesh. A small value (such as 1.85%1.85\% for Nu‾\overline{Nu}) proves that further refinement is yielding diminishing returns.
  3. Observed Apparent Order of Convergence (pp):

    • Formula: p=ln⁡∣(f3−f2)/(f2−f1)∣ln⁡(r)p = \frac{\ln|(f_3 - f_2) / (f_2 - f_1)|}{\ln(r)}
    • Relevance: Quantifies the actual rate at which numerical discretization error decreases with cell refinement. Comparing pp against the theoretical formal order (pformal=1p_{formal} = 1 for upwind) reveals whether the flow features are truly inside the asymptotic convergence regime (p≥1p \ge 1) or still in a pre-asymptotic state (p<1p < 1).
  4. Richardson Extrapolated Continuum Solution (fextf_{ext}):

    • Formula: fext=f1+f1−f2rp−1f_{ext} = f_1 + \frac{f_1 - f_2}{r^p - 1}
    • Relevance: Analytically eliminates the leading-order truncation error term (ChpC h^p), estimating the exact theoretical value at zero grid spacing (h→0h \to 0) without the prohibitive computational expense of an infinitely fine mesh.
  5. Fine Grid Convergence Index (GCIfineGCI_{fine}):

    • Formula: GCIfine=Fs⋅ea21rp−1GCI_{fine} = \frac{F_s \cdot e_a^{21}}{r^p - 1} (with ASME safety factor Fs=1.25F_s = 1.25 for 3-grid studies).
    • Relevance: Raw extrapolated errors can be overly optimistic. The GCI standardizes numerical uncertainty into an objective 95% confidence interval band. Rather than claiming Nu‾=59.53\overline{Nu} = 59.53 as absolute truth, we report Nu‾=59.53±0.16%\overline{Nu} = 59.53 \pm 0.16\%.
  6. Asymptotic Ratio (χ\chi):

    • Formula: χ=GCI32rp⋅GCI21\chi = \frac{GCI_{32}}{r^p \cdot GCI_{21}}
    • Relevance: Verifies that the grid triplet is within the asymptotic range where Taylor series expansions are mathematically valid. Values near χ≈1.0\chi \approx 1.0 confirm genuine asymptotic behavior.

4.2 In-Depth Analysis of the Verification Results

Examining the numbers across our four physical quantities reveals crucial insights into CFD discretization mechanics:

  1. Global Heat Transfer (Nu‾\overline{Nu}) Reaches Robust Mesh Independence:

    • The average Nusselt number shifts by only 1.85% between the Medium (100×100100 \times 100) and Fine (200×200200 \times 200) grids.
    • The resulting Fine Grid Convergence Index is an exceptionally tight GCIfine=0.16%GCI_{fine} = 0.16\%, bounding the exact continuum heat transfer rate at: Nu‾ext=59.61±0.10\overline{Nu}_{ext} = 59.61 \pm 0.10
    • This proves that for thermal engineering design (such as computing total heat loss across a double-glazed window or electronics enclosure), the 100×100100 \times 100 and 200×200200 \times 200 grids provide mesh-independent answers.
  2. Core Cavity Temperature (TcenterT_{center}) Preserves Exact Physical Anti-Symmetry:

    • In a Boussinesq square cavity with symmetric boundary temperatures (310310 K and 290290 K), the exact center temperature must equal the arithmetic mean: Tcore=(310+290)/2=300.0T_{core} = (310 + 290)/2 = 300.0 K.
    • The computed center temperature converges monotonically from 299.86299.86 K to 300.015300.015 K, with an approximate relative error of only 0.01%0.01\%.
    • This confirms that the numerical solver preserves anti-symmetric thermal stratification across the quiescent core without artificial thermal drift.
  3. Horizontal Recirculation Velocity (umaxu_{max}) Demonstrates Clean First-Order Accuracy:

    • The maximum horizontal turning velocity displays an apparent order of exactly p=1.00p = 1.00 and an asymptotic ratio of χ=1.000\chi = 1.000.
    • This exact match with the first-order upwind differencing scheme (bounded Gauss upwind in fvSchemes) proves that the bulk recirculating currents along the ceiling and floor reside cleanly within the asymptotic convergence regime.
  4. Vertical Peak Velocity (vmaxv_{max}) — Why GCIfine=584.87%GCI_{fine} = 584.87\% Blows Up & The Pre-Asymptotic Trap:

    • While global integrated quantities like Nu‾\overline{Nu} converge with sub-percent precision, the local peak vertical velocity (vmaxv_{max}) yields an apparent order of p=0.10p = 0.10 and an inflated uncertainty index of GCIfine=584.87%GCI_{fine} = 584.87\%.
    • What Does GCIfine=584.87%GCI_{fine} = 584.87\% Actually Signify? In the ASME V&V 20 framework, the GCI defines an objective 95% confidence uncertainty interval around the fine-grid prediction: fcontinuum=f1±GCIfinef_{continuum} = f_1 \pm GCI_{fine} For vmax=0.1935 m/sv_{max} = 0.1935\text{ m/s} with GCIfine=584.87%GCI_{fine} = 584.87\%, the uncertainty interval evaluates to: vmax,continuum=0.1935±(5.8487×0.1935)=0.1935±1.13 m/s  ⟹  [0,  1.325] m/sv_{max, continuum} = 0.1935 \pm (5.8487 \times 0.1935) = 0.1935 \pm 1.13\text{ m/s} \quad \implies \quad [0,\; 1.325]\text{ m/s} This signifies that the numerical discretization uncertainty is nearly six times larger than the computed value itself. The fine grid solution for this metric has not established mesh independence.
    • The Mathematical Mechanism Behind the Blowout: Recall the ASME GCI formulation: GCIfine=Fs⋅ea21rp−1GCI_{fine} = \frac{F_s \cdot e_a^{21}}{r^p - 1} Between Medium and Fine grids, the step change was large (ea21=33.58%e_a^{21} = 33.58\%). More critically, because truncation error was not decaying at the theoretical rate, the apparent order collapsed to p=0.10p = 0.10. When p→0p \to 0, the denominator (rp−1)=(20.10−1)≈0.0718(r^p - 1) = (2^{0.10} - 1) \approx 0.0718 collapses toward zero. Dividing an already large relative error by 0.07180.0718 (and multiplying by the safety factor Fs=1.25F_s = 1.25) magnifies the index by a factor of 17: GCIfine=1.25×33.58%0.0718≈584.87%GCI_{fine} = \frac{1.25 \times 33.58\%}{0.0718} \approx \mathbf{584.87\%}
    • The Physical Cause — Accelerating Step Changes (ϵ21>ϵ32\epsilon_{21} > \epsilon_{32}): Richardson Extrapolation assumes the solutions reside within the asymptotic range, where successive refinements produce monotonically smaller differences (ϵ21<ϵ32\epsilon_{21} < \epsilon_{32}). However, examining our step differences reveals:
      • Coarse to Medium difference: ϵ32=0.1285−0.0854=0.0431 m/s\epsilon_{32} = 0.1285 - 0.0854 = 0.0431\text{ m/s}
      • Medium to Fine difference: ϵ21=0.1935−0.1285=0.0650 m/s\epsilon_{21} = 0.1935 - 0.1285 = 0.0650\text{ m/s} Because ϵ21>ϵ32\epsilon_{21} > \epsilon_{32}, the velocity peak is steepening faster as cell width drops from 10 mm to 5 mm. This occurs because the physical thermal boundary layer thickness is δ≈4.7 mm\delta \approx 4.7\text{ mm}. On the coarse grid (Δx=20 mm\Delta x = 20\text{ mm}) and medium grid (Δx=10 mm\Delta x = 10\text{ mm}), cell size was larger than the entire boundary layer; the jet was completely smeared across 1 or 2 cells by numerical diffusion. Only on the fine grid (Δx=5 mm\Delta x = 5\text{ mm}) did cell spacing begin to resolve the boundary layer, allowing the steep physical peak to emerge.

4.3 Why Don’t All Metrics Converge Simultaneously in CFD?

A common pitfall in CFD verification is assuming that a simulation is either “converged” or “not converged” as a blanket state. In reality, different flow metrics operate on fundamentally different spatial scales and physical mechanisms:

  1. Global Integrated Quantities (Nu‾\overline{Nu}) vs. Spatial Error Cancellation:

    • Metrics obtained by integrating over a surface or volume—such as the average Nusselt number Nu‾\overline{Nu}, total drag CdC_d, or mass flow rate—benefit from spatial error averaging.
    • Local discretization errors and discretization noise along the wall average out over the full surface integral (y∈[0,1]y \in [0, 1]). Consequently, global heat transfer achieves an exceptional GCIfine=0.16%GCI_{fine} = 0.16\% and enters the asymptotic regime even on coarse-to-medium grids.
  2. Macro-Scale Core Quantities (TcenterT_{center}, umaxu_{max}) vs. Large Length Scales:

    • The central core of the cavity and the recirculating turning currents along the ceiling and floor have large characteristic dimensions (L∼0.5−1.0 mL \sim 0.5 - 1.0\text{ m}).
    • Even a coarse mesh (Δx=20 mm\Delta x = 20\text{ mm}) provides 50 cells across the cavity, which is more than sufficient to resolve smooth, macroscopic core stratification (Tcenter=300.015 KT_{center} = 300.015\text{ K}, ea21=0.01%e_a^{21} = 0.01\%) and bulk recirculating flow (p=1.00p = 1.00, χ=1.000\chi = 1.000).
  3. Localized Gradient Peaks (vmaxv_{max}) vs. Boundary Layer Thinning:

    • Point extrema situated inside steep boundary layers (such as maximum wall shear stress τw\tau_w, local skin friction CfC_f, or peak upward buoyant velocity vmaxv_{max}) are governed by the smallest hydrodynamic scales (δ∼Ra−1/4≈4.7 mm\delta \sim Ra^{-1/4} \approx 4.7\text{ mm}).
    • Resolving a point maximum without numerical damping requires resolving the second derivative of velocity (∂2Uy/∂x2\partial^2 U_y / \partial x^2) at the peak, demanding at least 10−1510 - 15 cells clustered inside δ\delta (Δxwall<0.5 mm\Delta x_{wall} < 0.5\text{ mm}).
    • On uniform Cartesian meshes, refining from 20 mm→10 mm→5 mm20\text{ mm} \to 10\text{ mm} \to 5\text{ mm} only just begins to uncover the peak. The metric remains pre-asymptotic, triggering the GCIfine=584.87%GCI_{fine} = 584.87\% blowout.

Key Verification Principle: Mesh independence for global integrated quantities (heat flux, lift, drag) does not imply mesh independence for localized point metrics (peak velocities, hot spots, separation points). A rigorous ASME V&V 20 verification study must always evaluate and report both.


5. Key Engineering Conclusions

  1. Global Heat Transfer is Mesh-Independent:
    The global heat transfer across the cavity is verified to high accuracy. With an ASME Fine Grid Convergence Index of GCIfine=±0.16%GCI_{fine} = \pm 0.16\%, the continuum thermal solution is established at: Nu‾ext=59.61±0.10\overline{Nu}_{ext} = 59.61 \pm 0.10
  2. Computational Efficiency Tradeoff:
    The Medium grid (100×100100 \times 100) completed in 35 seconds and came within 1.85% of the Fine grid (200×200200 \times 200), which required 169 seconds. For engineering sweeps, the 100×100100 \times 100 mesh offers the optimal accuracy-to-cost ratio.
  3. Local Velocity vs. Global Thermal Sensitivity (The Multi-Scale Lesson):
    While integrated energy flux (Nu‾\overline{Nu}) achieves robust mesh independence (GCIfine=0.16%GCI_{fine} = 0.16\%), capturing the localized peak of the vertical wall jet (vmaxv_{max}) produces an uncertainty blowout (GCIfine=584.87%GCI_{fine} = 584.87\%). This demonstrates why mesh convergence is metric-dependent: global integrals benefit from spatial error cancellation, whereas local boundary layer peaks require near-wall mesh grading (Δxwall≪δ≈4.7 mm\Delta x_{wall} \ll \delta \approx 4.7\text{ mm}) rather than uniform Cartesian grids.

6. Reproducibility

The complete 3-grid verification workflow is automated under 04_mesh_independence_gci/:

# Clone the repository
git clone https://github.com/martinmat13/OpenFOAM-Differentially-Heated-Cavity.git
cd OpenFOAM-Differentially-Heated-Cavity/04_mesh_independence_gci

# Run all 3 grids and generate GCI verification tables and plots
./Allrun

All verification metrics and comparison figures are saved directly into 04_mesh_independence_gci/results/.

Related Projects