LEARN / OCEAN WAVES / CHAPTER 20

Numerical Model Principles

Full original scientific lesson and experiments, reorganized into a dedicated chapter.

Original physics and annotationsInteractive scientific demonstrationsScientific teaching examples

Scientific scope: This lesson retains the original scientific explanations, diagrams, equations and interactive demonstrations. All numerical examples are educational and subject to their stated assumptions; they are not independently field-validated engineering design tools or operational forecasts.

CHAPTER 20 / DISCRETIZATION AND CONVERGENCE

Turn equations
into a calculation.

A model stores a finite set of numbers. Geographic cells describe where waves are; frequency and direction bins describe which waves are present. Time steps advance those numbers. Each choice can change the answer, even when the physical equations stay the same.

Four resolutions, different jobs.

A geographic grid must represent the bathymetry and the spatial scales of propagation. Frequency bins must resolve spectral peaks and tails. Directional bins must distinguish spreading and crossing systems. Time steps must resolve evolution and satisfy the requirements of the numerical method. Refining one coordinate does not repair inadequate resolution in another.

For a structured calculation, a rough storage count is Nx Ny Nf Nθ values per field. Doubling both horizontal dimensions multiplies that count by four; doubling the frequency and directional counts as well multiplies it by sixteen. More time steps and more source iterations add work without necessarily adding stored fields. These counts describe scaling, not a hardware benchmark.

Stable does not mean accurate.

A stable scheme can still smear a narrow swell peak or diffuse a packet in geographic space. Convergence means that a sequence of appropriately refined calculations approaches a consistent solution. It is not established by one fine-looking plot or by a small iteration residual on one grid.

Compare at the same physical time and with the same forcing and geometry. Record the outputs that matter: wave height, period, directional spread, energy flux and local extremes. A mesh that resolves significant height adequately may still miss a narrow directional feature.

Separate spatial and spectral resolution.

The upper plot repeats Chapter 19's periodic packet test at t = 200 s (one complete circuit of the domain), using Courant number 0.5 on each grid. The lower plot samples a narrow Gaussian teaching spectrum with known variance 0.36 m². Its peak is deliberately not aligned with most bin centres. No numerical renormalization hides a missed peak.

Blue: upwind cell averages; coral: analytical translated cell averages. Compare the same 2 km domain, speed 10 m/s and elapsed time.

Blue: midpoint-bin samples joined for visibility; coral: continuous reference. The frequency interval is 0.04–0.40 Hz. Peak frequency is 0.123 Hz and standard deviation is 0.006 Hz.

These controls change different calculations. Spatial refinement reduces packet diffusion; spectral refinement resolves the narrow peak. Neither is a physical growth or dissipation process.

What is being measured?

The packet error is the domain mean of the absolute difference between numerical and exact cell averages. The table compares all four grids after 200 s. Each grid uses Δx = 2000/N and Δt = 0.5Δx/10, hence 2N updates. A smaller time step accompanies a smaller cell to keep the Courant number fixed.

The spectrum is S(f) = 0.36 exp[−(f − 0.123)²/(2 × 0.006²)]/(0.006√(2π)). Its full-real-line integral is 0.36 m²; the omitted negative-frequency and displayed-band tails are negligible at these parameters. The numerical estimate is ΣS(fi)Δf. Peak period uses the largest sampled value and can change in steps as the grid changes. Zero-padding an FFT would not supply this missing physical or observational resolution.

Boundaries are part of the problem.

Specify incoming spectra at open boundaries, with units and direction conventions checked explicitly. Land boundaries, reflections and outgoing conditions depend on the model and configuration. Place artificial lateral boundaries far enough from the area of interest, and test their influence. Finer interior cells cannot repair incorrect boundary forcing.

A stationary solver seeks a balance under fixed conditions; a nonstationary solver also needs an initial state and an appropriate spin-up or initialization procedure. Iterative convergence, time-step sensitivity and grid convergence are separate checks. Source terms can impose additional restrictions beyond simple propagation.