Numerical Methods and Engineering Simulation Exercises

These exercises practise numerical methods and engineering simulation as decision evidence. They focus on the checks that determine whether a computed result is credible enough to support design, troubleshooting, optimization, validation, or release.

The goal is not to run software. The goal is to understand whether the model, discretization, solver, time step, boundary conditions, convergence history, uncertainty, and validation evidence are strong enough for the engineering decision. A simulation can converge numerically and still represent the wrong physics or the wrong operating condition.

Assume simplified screening models unless an exercise states otherwise. Real simulation work should also check units, geometry source, material data, boundary conditions, solver settings, mesh quality, element or cell type, time integration method, software version, hardware precision, validation data, and documented limits of use.

How to Use These Exercises

For each exercise, state:

1. the quantity of interest and decision threshold;
2. the numerical method and discretization;
3. the error, residual, convergence, or stability measure being checked;
4. the validation or benchmark evidence needed beyond the calculation;
5. the action if the numerical evidence is weaker than the decision margin.

The common mistake is treating a plot, a solver “converged” message, or one mesh as proof. Engineering simulation needs traceable numerical checks and physical validation.

Release Evidence Notes

Numerical simulation evidence should begin with the decision it supports. Record the quantity of interest, release threshold, physical boundary, model scope, governing equations, discretization, solver settings, material data, units, coordinate system, load case and validation target before interpreting any numerical result. A converged plot is not enough if the decision variable was never defined.

Discretization evidence should separate mesh, time-step, interpolation, integration and boundary-condition effects. Mesh convergence, observed order, CFL number, explicit stability, residual scaling, solver tolerance and conservation closure answer different questions. Passing one check does not prove the others, and a quantity near a release threshold should carry a numerical margin.

Solver evidence should include the residual definition, scaling, norm, iteration history, relaxation strategy, stopping rule and quantity-of-interest history. A small residual can be misleading when units are mixed, the matrix is ill-conditioned, nonlinear contact is changing state, or the engineering output is still drifting.

Validation evidence should connect the model to the physical system. Benchmark cases, measurement data, calibration status, uncertainty, sensor placement, boundary-condition realism, geometry source, material traceability and version control should be documented before a simulation is accepted as release evidence. Simulation precision is not the same as validation accuracy.

The practical release question is whether discretization, solver behavior, physical assumptions, uncertainty and validation data all support the same engineering action. If one layer conflicts, the result should trigger mesh refinement, time-step reduction, solver review, boundary-condition correction, validation testing, guard-band increase or refusal to release the design.

Engineering Boundary Notes

These exercises use simplified numerical-methods and simulation verification screens. They do not replace code verification, solution verification, model validation, uncertainty quantification, configuration management, solver qualification or formal simulation credibility assessment. A simulation margin applies only to the stated model, discretization, solver settings, boundary conditions, quantity of interest, validation basis and release criterion.

Separate numerical convergence from physical credibility. Residuals, mesh refinement, time-step stability, conservation closure, solver tolerance, GCI, surrogate error, FFT resolution and Monte Carlo precision each control a different credibility claim. A stable-looking solution can still be numerically invalid, nonconservative or outside validation.

Common Release Mistakes

• accepting low residuals while the quantity of interest is mesh or time-step dependent;
• refining singular point stress instead of using a physically meaningful averaged or regularized quantity;
• treating explicit stability as accuracy or conservation proof;
• using surrogate optimization after validation error consumes release margin;
• quoting Monte Carlo or sensitivity results without enough samples or boundary-condition evidence;
• releasing a simulation result when validation case, solver version or configuration differs from the decision case.

Scenario Map

Scenario	Exercises	Primary check	Engineering decision
Discretization and convergence evidence	1, 2, 4, 13, 17	Finite difference, observed order, mesh refinement margin, mesh-dependent stress and GCI	Decide whether the quantity of interest is numerically stable enough to use.
Solver behavior and numerical conditioning	3, 7, 8, 9, 15	Residual scaling, Newton iteration, condition number, solver tolerance and under-relaxed update control	Separate numerical convergence from physical credibility and release impact.
Time-step and conservation controls	5, 6, 10, 18	Explicit stability, CFL screen, mass-conservation closure and transient time-step refinement	Hold simulations that are stable-looking but numerically invalid, nonconservative or under-refined in time.
Uncertainty and sensitivity for release	11, 12	Monte Carlo precision and boundary-condition sensitivity	Add samples, guard bands or validation when uncertainty is close to the decision limit.
Spectral numerical evidence	14	FFT bin spacing, record duration and Nyquist frequency	Choose a record long enough to separate the frequencies that support the decision.
Surrogate and model-reuse validation	16	Withheld high-fidelity error, maximum error guard and optimization margin	Block optimized candidates when model-reduction error consumes the release margin.

Validation Package Checklist

• model version, solver, mesh, time step, boundary conditions, quantity of interest and release limit are documented;
• residual scaling, convergence history, conservation closure and conditioning are reviewed before using results;
• mesh refinement, observed order, GCI, time-step refinement and localized singular behavior are separated;
• uncertainty, sensitivity, Monte Carlo precision and validation evidence are visible near limits;
• surrogate, reduced-order and reused models have withheld validation against the governing quantity of interest;
• solver tolerance, under-relaxation, CFL, record length and numerical stability choices are configuration controlled;
• final decision states accept, refine mesh, reduce time step, revise model, add validation, restrict surrogate use or hold.

Exercise 1: Finite Difference Heat-Flux Estimate

A one-dimensional temperature profile is measured at three equally spaced locations:

Position	Temperature
x=0.00\ \text{m}	80^\circ\text{C}
x=0.02\ \text{m}	74^\circ\text{C}
x=0.04\ \text{m}	69^\circ\text{C}

Estimate the temperature gradient at x=0.02\ \text{m} using a central difference. Then estimate conductive heat flux using:

with:

Solution

Central difference:

Here:

so:

Heat flux:

Engineering Comment

The result is a local estimate based on three points. It assumes one-dimensional conduction, steady behavior, uniform material, and temperatures measured at representative locations. A real simulation comparison should also check contact resistance, lateral heat spreading, sensor placement, calibration uncertainty, and whether the grid spacing resolves the thermal gradient.

Plausibility Check

The temperature drops by 11^\circ\text{C} across 0.04\ \text{m}, so a gradient of -275\ \text{K/m} is consistent. With positive conductivity, heat flux is positive in the direction of decreasing temperature, giving 4125\ \text{W/m}^2.

Exercise 2: Observed Mesh Convergence Order

A finite element model predicts peak stress in a rounded bracket. The same quantity of interest is extracted on three successively refined meshes:

Mesh	Peak stress
h	102.4\ \text{MPa}
h/2	98.1\ \text{MPa}
h/4	97.0\ \text{MPa}

Estimate the observed convergence ratio:

and compare it with the expected value for a second-order asymptotic response:

Solution

Compute the two differences:

Ratio:

This is close to:

so the sequence is consistent with second-order convergence for this quantity.

Engineering Comment

The convergence behavior is encouraging, but it applies only to the extracted quantity and modeling setup. It does not prove that the boundary conditions, material model, contact assumptions, or load path are correct. Convergence should be reported for the engineering quantity of interest, not only for global residuals.

Plausibility Check

The first refinement changes stress by 4.3\ \text{MPa} and the second by 1.1\ \text{MPa}. The ratio 3.91 is close to the expected second-order value of 4, so the convergence interpretation is plausible.

Exercise 3: Residual Scaling Across Mixed Units

A coupled simulation reports residual components:

Equation block	Residual	Scale
Force balance	120\ \text{N}	5000\ \text{N}
Moment balance	0.8\ \text{N m}	20\ \text{N m}
Mass conservation	0.004\ \text{kg/s}	0.2\ \text{kg/s}

Compute normalized residuals and the infinity norm. The acceptance criterion is:

Solution

Normalized residuals:

The infinity norm is the largest absolute normalized residual:

Compare with the criterion:

The scaled residual criterion passes.

Engineering Comment

Scaling prevents one unit system from dominating the convergence check. The result still needs a quantity-of-interest check. A small residual means the numerical equations are internally consistent; it does not prove that the equations, boundary conditions, or model form represent the real system.

Plausibility Check

The normalized residuals are 0.024, 0.040 and 0.020. The infinity norm is therefore the largest of those values, 0.040, and it is below the 0.05 criterion.

Exercise 4: Mesh Convergence Decision Margin

A pressure-vessel support model predicts maximum structural stress:

Mesh	Stress
Coarse	142\ \text{MPa}
Fine	151\ \text{MPa}
Extra fine	154\ \text{MPa}

The team requires the change from the last refinement to be less than:

of the extra-fine result. Check whether the final refinement passes.

Solution

Change from fine to extra fine:

Relative change:

As a percentage:

Since:

the final refinement passes the stated mesh-convergence screen.

Engineering Comment

The coarse-to-fine change was larger, so using the coarse mesh alone would have been weak evidence. Passing this convergence screen does not automatically justify the stress value. The support constraints, load introduction, local stress concentration, material data, and code acceptance method still need review.

Plausibility Check

The final refinement changes the result by 3\ \text{MPa} relative to 154\ \text{MPa}. That is only 1.95\%, so it passes a 5\% mesh-change screen.

Exercise 5: Explicit Heat-Equation Time-Step Stability

An explicit one-dimensional heat-conduction simulation uses thermal diffusivity:

and grid spacing:

For the simple explicit scheme, use the stability screen:

Find the maximum stable time step and check whether:

passes.

Solution

Rearrange:

Convert:

Then:

The proposed step is:

so it fails the stability screen.

Engineering Comment

An unstable explicit simulation can produce oscillations or divergence that are numerical, not physical. Reducing time step, refining the method, or using an implicit scheme may be needed. Stability is not the same as accuracy; even a stable time step may be too coarse to capture the thermal transient that matters.

Plausibility Check

The stability limit scales with \Delta x^2, so a 2.0\ \text{mm} grid permits only a small explicit time step. The computed maximum is 0.1429\ \text{s}, making the proposed 0.25\ \text{s} step clearly too large.

Exercise 6: CFL Screen for an Advective Transient

A one-dimensional flow model has characteristic velocity:

and cell size:

Use the first-pass CFL screen:

Check whether:

passes.

Solution

Maximum time step:

So:

The proposed time step is:

CFL number:

The proposed step fails the screen.

Engineering Comment

This check is especially important for wave propagation, fluid transients, moving fronts, queue simulations with time buckets, and sampled control models. A time step can make a model look smooth while hiding the fast event that controls the decision.

Plausibility Check

At 12\ \text{m/s}, a disturbance crosses a 0.15\ \text{m} cell in 12.5\ \text{ms}. A 20\ \text{ms} time step is longer than that crossing time and gives C=1.6, so the screen fails.

Exercise 7: Newton Iteration for a Nonlinear Equation

Use Newton’s method to solve:

starting from:

Perform two iterations.

Solution

Derivative:

Newton update:

First iteration:

Second iteration:

The exact positive root is:

Engineering Comment

This example behaves well because the derivative is not small and the starting point is close. Engineering nonlinear solvers can fail or converge to the wrong branch because of contact, buckling, material softening, discontinuous controls, poor initial guesses, or badly scaled variables. Residual history and physical plausibility should both be reviewed.

Plausibility Check

The first Newton step moves from 3.0 to 3.1667, close to \sqrt{10}. The second step gives 3.1623, essentially the positive root, which is expected for a smooth function with a good starting point.

Exercise 8: Conditioning and Error Amplification

A linear solve has estimated condition number:

The input data have relative uncertainty:

Estimate the possible relative solution perturbation using the first-pass bound:

Solution

Substitute:

As a percentage:

Engineering Comment

A 0.02 percent input perturbation can become a 20 percent solution perturbation under this bound. The bound may be conservative, but it is a warning. Poor conditioning can come from scaling, weak constraints, nearly dependent equations, extreme stiffness ratios, bad mesh quality, or penalty parameters. Scaling, preconditioning, model reformulation, or better constraints may be more important than tightening the residual tolerance.

Plausibility Check

The condition number is large at 1000. Multiplying it by the small input uncertainty 0.0002 gives 0.20, so a seemingly tiny input perturbation can plausibly become a large solution perturbation.

Exercise 9: Solver Tolerance and Quantity of Interest

A simulation is run with three solver tolerances. The engineering quantity of interest is reaction force:

Relative residual tolerance	Reaction force
10^{-2}	101.8\ \text{kN}
10^{-3}	102.4\ \text{kN}
10^{-4}	102.5\ \text{kN}

The release rule requires the quantity of interest to change by less than:

between the two tightest tolerances. Check the rule.

Solution

Change:

Relative change using the tightest result:

As a percentage:

Since:

the tolerance study passes for this quantity.

Engineering Comment

The result says the reaction force is insensitive to this tolerance range. It does not say every output is insensitive. Peak contact pressure, local stress, energy balance, or flow separation may require tighter settings. Solver tolerance should be tied to the quantity used in the decision.

Plausibility Check

The two tightest tolerances change the reaction force by only 0.1\ \text{kN} out of 102.5\ \text{kN}. That is about one tenth of one percent, safely below the 0.5\% rule for this quantity.

Exercise 10: Conservation Check in a Flow Simulation

A steady flow simulation has inlet mass flow:

Two outlet flows are:

and:

The model should have negligible accumulation. Compute mass imbalance as a percentage of inlet flow. The acceptance limit is:

Solution

Total outlet flow:

Imbalance:

Relative imbalance:

As a percentage:

Since:

the conservation check fails.

Engineering Comment

A failed conservation check is stronger evidence than a visually plausible contour plot. The cause might be unconverged residuals, boundary setup, compressibility assumption, interpolation across report surfaces, leakage through unintended boundaries, or post-processing error. The simulation should not be released until conservation is reconciled.

Plausibility Check

The two outlets sum to 11.7\ \text{kg/s}, which is 0.3\ \text{kg/s} below the inlet. Relative to 12.0\ \text{kg/s}, that is 2.5\%, well above the 0.5\% acceptance limit.

Exercise 11: Monte Carlo Numerical Precision

A Monte Carlo simulation estimates failure probability:

from:

independent samples. Estimate the standard error:

and an approximate 95 percent interval:

The requirement is:

Solution

Standard error:

95 percent half-width:

Interval:

so:

The upper end is below:

but only narrowly.

Engineering Comment

The point estimate alone would look comfortable. The numerical precision shows that the evidence is just inside the requirement. If the consequence is high, the team may need more samples, variance reduction, importance sampling, better input distributions, or a guard band before claiming compliance.

Plausibility Check

With 20000 samples and a probability near 0.018, the standard error is below one tenth of one percent. The 95 percent upper estimate is 0.01984, below the 0.020 requirement but with very little margin.

Exercise 12: Boundary-Condition Sensitivity

A thermal-stress simulation predicts peak stress:

with nominal support stiffness. A support-stiffness sensitivity study gives:

Support stiffness case	Peak stress
-30\% stiffness	171\ \text{MPa}
nominal stiffness	186\ \text{MPa}
+30\% stiffness	205\ \text{MPa}

The allowable stress is:

Find the worst-case margin and decide whether the nominal result alone is adequate evidence.

Solution

Worst stress in the sensitivity set:

Worst-case margin:

Relative margin:

As a percentage:

Engineering Comment

The nominal simulation gives a margin of 34\ \text{MPa}, but the boundary-condition study reduces that margin to 15\ \text{MPa}. The nominal result alone is not adequate evidence if support stiffness is uncertain. The engineering action may be to measure stiffness, improve the support model, add a guard band, change the design, or validate with strain-gauge data.

Plausibility Check

The highest stress in the sensitivity set is 205\ \text{MPa}, not the nominal 186\ \text{MPa}. Comparing 205\ \text{MPa} with the 220\ \text{MPa} allowable leaves only 15\ \text{MPa}, or 6.82\%.

Exercise 13: Interpreting a Mesh-Dependent Stress Peak

A finite element model reports maximum point stress near a sharp re-entrant corner:

Mesh	Point stress near corner	Section-average stress
Coarse	180\ \text{MPa}	120\ \text{MPa}
Fine	260\ \text{MPa}	123\ \text{MPa}
Extra fine	410\ \text{MPa}	124\ \text{MPa}

Decide which result is more credible for a release screen and what additional review is needed.

Solution

The point stress changes strongly with mesh refinement:

It is not converging.

The section-average stress changes much less:

Change from fine to extra fine:

The section-average stress is much more stable than the point stress.

Engineering Comment

The point stress may be a mathematical singularity caused by an ideal sharp corner, point load, rigid constraint, or contact edge. It should not be used blindly as a material-failure stress. The release screen should use a stress measure appropriate to the failure mode and code basis, such as averaged stress, structural stress, notch stress with a justified radius, fracture mechanics, or a refined local model with validated geometry.

Plausibility Check

The point stress grows from 180 to 410\ \text{MPa} as the mesh is refined, so it is not converging. The section-average stress changes only from 123 to 124\ \text{MPa} on the final refinement, a 0.81\% change, making it the more stable release-screen quantity.

Exercise 14: FFT Bin Spacing and Record-Length Requirement

A vibration simulation exports a time record sampled at:

The current FFT block length is:

Two spectral components near an operating order are expected at:

and:

The release rule requires at least three FFT bins across the frequency separation. Check the current record, then estimate the minimum record duration and the next power-of-two block length that satisfies the rule.

Solution

Record duration is:

FFT bin spacing is:

Nyquist frequency is:

The two components are below Nyquist, but separability is controlled by record length. Their separation is:

Number of current bins across the separation:

The current record fails the three-bin rule.

Required bin spacing is:

Required record duration is:

Required sample count is:

If the workflow uses power-of-two FFT blocks, the next practical block is:

with:

and:

Bins across the separation become:

The 8192-sample block satisfies the release rule.

Engineering Comment

Frequency resolution comes from record duration, not from making a plot look denser. Zero padding can interpolate the displayed spectrum, but it does not add physical observation time. A defensible FFT review should also state window type, leakage risk, amplitude scaling, sample-clock accuracy, anti-alias filtering, stationarity, averaging, and whether speed variation would make order tracking more appropriate than a single stationary FFT.

Plausibility Check

The current 4096-sample record at 2048\ \text{Hz} lasts 2.0 seconds, so 0.50\ \text{Hz} bin spacing is expected. A 1.2\ \text{Hz} separation spans only 2.4 bins and fails a three-bin rule. Extending to 8192 samples doubles the record duration and halves bin spacing to 0.25\ \text{Hz}, giving 4.8 bins across the separation.

Exercise 15: Under-Relaxed Nonlinear Update Acceptance

A coupled thermal-structural simulation uses a fixed-point coupling iteration. At the current iteration, the state variable is:

The full solver correction would be:

The review rule limits any accepted iteration increment to:

The nonlinear residual before the update is:

The solver proposes an under-relaxation factor:

After applying that damped update, the residual is:

The iteration acceptance rule requires at least a 25\% residual reduction. The final release tolerance is:

Check whether the full update is acceptable, whether the under-relaxed update is acceptable for continuing the solve, and whether the simulation is ready for release.

Solution

The full update would give:

But its increment is:

This exceeds the accepted-increment limit:

So the full update fails the iteration control rule.

The largest allowable relaxation factor from the increment limit is:

The proposed value is:

So the under-relaxed increment is within the limit:

The accepted state becomes:

The required residual after a 25\% reduction is:

The observed residual is:

Because:

the damped iteration satisfies the residual-reduction rule.

The residual ratio is:

Residual reduction is:

or about:

The damped step is acceptable for continuing the nonlinear solve. It is not enough for final release because:

The correct decision is to accept the damped iteration, continue solving, and withhold release until the absolute residual and the quantity of interest both satisfy the release criteria.

Engineering Comment

Under-relaxation is not a way to hide convergence trouble. It is a controlled way to limit an update when the full nonlinear correction is too aggressive for the model state, contact condition, coupling interface, or load step. A defensible review records the full update, the accepted damped update, the residual history, the quantity-of-interest history, and why the relaxation factor is still physically credible. A small residual reduction is not enough if the stress, force, temperature, or displacement driving the engineering decision is still moving.

Plausibility Check

The full correction is almost three times the allowed increment, so accepting it would be inconsistent with the review rule. Multiplying by \alpha=0.35 gives 0.28\ \text{mm}, just below the 0.30\ \text{mm} limit, which is plausible for a cautious nonlinear step. The residual reduction is about one third, so the iteration is improving, but 1.05\ \text{kN} remains far above the 0.20\ \text{kN} release tolerance.

Exercise 16: Surrogate-Model Validation for an Optimized Candidate

A reduced-order surrogate model is used to optimize a bracket geometry. Before accepting an optimized candidate, the team compares surrogate peak-stress predictions with withheld high-fidelity simulations:

Validation point	High-fidelity stress	Surrogate stress
A	142\ \text{MPa}	138\ \text{MPa}
B	156\ \text{MPa}	151\ \text{MPa}
C	171\ \text{MPa}	165\ \text{MPa}
D	188\ \text{MPa}	176\ \text{MPa}

The surrogate can be used for automatic optimization release only if:

and:

The optimizer proposes a new candidate with surrogate-predicted peak stress:

The allowable stress is:

and the release rule requires a guarded stress margin of at least:

using the maximum withheld validation error as the surrogate guard band. Compute the validation errors, e_{max}, RMSE, guarded candidate stress and release decision.

Solution

Prediction errors using surrogate minus high-fidelity stress are:

Maximum absolute validation error:

Root-mean-square validation error:

The surrogate fails both validation gates:

and:

Guarded candidate stress using the maximum validation error:

Guarded margin:

The candidate fails the required guarded margin because:

The optimized candidate should not be released automatically. The team should run a high-fidelity check at the candidate, retrain or restrict the surrogate, and review whether point D is outside the validated design domain.

Engineering Comment

This is a model-reuse and optimization gate. The surrogate is fast, but it underpredicts stress at the high-stress validation point by 12\ \text{MPa}. An optimizer will tend to move toward regions where model error is useful to the objective. A defensible workflow guards the optimized result with withheld validation error and requires high-fidelity confirmation when the guard consumes the release margin.

Plausibility Check

The largest error occurs at the highest-stress validation point, which is exactly where an optimized lightweight bracket is most likely to be risky. Adding the 12\ \text{MPa} guard to the 178\ \text{MPa} surrogate prediction reaches the 190\ \text{MPa} allowable, leaving no margin. Blocking automatic release is therefore consistent with both the validation metrics and the candidate stress screen.

Exercise 17: Richardson Extrapolation and Grid Convergence Index

A finite element model predicts maximum service displacement of a support bracket. The release limit is:

The same quantity of interest is extracted from three uniformly refined meshes with refinement ratio:

Mesh	Displacement
h	4.82\ \text{mm}
h/2	4.46\ \text{mm}
h/4	4.37\ \text{mm}

Use the observed-order estimate:

Use Richardson extrapolation from the two finest meshes:

and a fine-grid Grid Convergence Index screen:

with:

Calculate observed order, extrapolated displacement, fine-grid GCI, absolute GCI in millimeters, guarded upper displacement and release decision.

Solution

Mesh-to-mesh differences are:

Observed order:

Richardson extrapolated displacement:

Fine-grid relative difference:

Fine-grid GCI:

Absolute fine-grid GCI:

The fine-grid displacement alone appears to pass:

But the guarded upper displacement is:

Compare with the release limit:

The guarded release decision is therefore hold. The mesh sequence is converging, but the remaining discretization uncertainty is large enough to consume the fine-grid margin.

Engineering Comment

Grid convergence evidence is strongest when it is tied to the exact quantity being released. Richardson extrapolation and GCI do not prove that the model physics, boundary conditions or material data are correct; they only screen numerical discretization error for the chosen setup. When the fine-grid result is close to a limit, the guarded value should drive the release decision rather than the unguarded mesh output.

Plausibility Check

The changes shrink from (0.36\ \text{mm}) to (0.09\ \text{mm}), exactly a factor of four, so an observed second-order response is plausible. The extrapolated value moves slightly below the finest result because the sequence is decreasing toward the limit value. The fine-grid value has only (0.03\ \text{mm}) of unguarded margin, while the absolute GCI is about (0.038\ \text{mm}), so the guarded hold decision is internally consistent.

Exercise 18: Time-Step Refinement and Transient Peak Release Gate

A transient thermal simulation predicts a short-duration peak temperature at a protected component. The release limit is:

The same model is run with three uniformly refined time steps:

Time step	Peak temperature
4\ \text{ms}	90.10\ \text{deg C}
2\ \text{ms}	91.00\ \text{deg C}
1\ \text{ms}	91.225\ \text{deg C}

Use the observed-order estimate:

Use Richardson extrapolation from the two finest time steps:

and a GCI-style time-step guard:

with:

Calculate observed temporal order, extrapolated peak temperature, time-step GCI, absolute temperature guard, guarded upper temperature and release decision.

Solution

Time-step differences are:

Observed temporal order:

Richardson extrapolated transient peak:

Fine time-step relative difference:

Time-step GCI:

Absolute temperature guard:

The finest time-step result alone appears to pass:

But the guarded upper peak is:

Compare with the release limit:

The guarded release decision is hold. The time-step sequence is converging, but the remaining temporal discretization uncertainty consumes the unguarded margin.

Engineering Comment

A stable transient simulation can still be too coarse for release. Explicit stability and CFL screens prevent certain numerical failures, but they do not prove that the peak value used for a thermal, pressure, stress or control decision has converged in time. When the time-step-refined peak is close to a limit, the release package should preserve time-step sequence, observed temporal order, peak extraction method, interpolation basis, event timing and a numerical guard before accepting the result.

Plausibility Check

The peak temperature increases as the time step is refined, so the coarse result is nonconservative for this release limit. The differences shrink from (0.90\ \text{deg C}) to (0.225\ \text{deg C}), a factor of four, so second-order temporal behavior is plausible. The finest result has only (0.075\ \text{deg C}) of unguarded margin, while the absolute guard is about (0.094\ \text{deg C}), making the hold decision internally consistent.

Review Checklist

When reviewing numerical-method or engineering-simulation evidence, ask:

• Is the quantity of interest tied to a design, troubleshooting, validation or release decision?
• Are the physical boundary, governing equations, geometry source, material data, loads and units stated?
• Is the discretization described: mesh or grid, element or cell type, time step, interpolation and integration method?
• Are mesh convergence, observed order, time-step stability, CFL number or spectral resolution checked where relevant?
• Are Richardson extrapolation or GCI guards used when a mesh-dependent quantity is close to a release limit?
• Are time-step refinement and temporal guards used when a transient peak controls the release decision?
• Is the residual definition clear, scaled consistently and compared with the quantity-of-interest change?
• Are solver tolerance, nonlinear update size, under-relaxation, conditioning and iteration history recorded?
• Are conservation, balance closure or benchmark checks included before trusting local fields or peaks?
• Are boundary-condition sensitivity, model-form uncertainty and numerical precision separated from measurement uncertainty?
• Are validation data traceable to sensor location, calibration state, operating condition and uncertainty?
• Are surrogate, reduced-order or reused models checked against withheld high-fidelity or physical validation data before optimization release?
• Is the release margin large enough after numerical error, model uncertainty, validation error and guard bands?
• Does the result imply mesh refinement, time-step reduction, solver review, retest, guard-band increase or release?

Common Mistakes

• Treating a solver “converged” message as proof that the modeled physics and boundary conditions are correct.
• Comparing two meshes without checking whether the quantity of interest is in the asymptotic convergence range.
• Treating a fine-grid result as a release pass when the GCI or numerical guard consumes the margin.
• Reporting residuals without units, scaling or a link to the engineering output being released.
• Using a stable-looking transient result after violating the explicit time-step or CFL condition.
• Accepting a transient peak from one stable time step without checking whether the peak has converged in time.
• Tightening solver tolerance while ignoring mesh error, material uncertainty or boundary-condition sensitivity.
• Accepting a stress peak that keeps growing with mesh refinement as a material release value.
• Using Monte Carlo output precision as if it were validation accuracy.
• Choosing an FFT record too short to separate the frequencies that drive the decision.
• Applying under-relaxation to force a solve through an unstable load step without documenting physical credibility.
• Releasing an optimized surrogate-model result without guarding it by withheld validation error or high-fidelity confirmation.
• Releasing a simulation result after geometry, load case, contact state, solver version or material data changed without rerun evidence.

Engineering simulation is strongest when the numerical checks, physical assumptions, validation data and decision margin all describe the same quantity of interest.