Modeling and Discretization

Explains how to transform a continuous physical system into a computable, verifiable discrete model.

Conceptual context

Physics describes continuous systems through fields, geometries, and laws containing infinitely many points. A computer can store only finite data. Modeling selects the phenomena, variables, and scales considered essential. Discretization replaces the continuum with nodes, cells, elements, and time steps. $h$ controls resolution, $N$ and $N_dof$ set algebraic size, $E_h$ measures lost fidelity, and $Co$ links spatial and temporal resolution. A representation is useful when it preserves relevant behavior without impossible computational cost.

🟢 Essential level

A computational model does not copy reality point by point. It selects variables, assumptions, and boundaries, then divides the domain into manageable pieces.

$h$ sets the size of those pieces. As it decreases, $N$, nodes, and degrees of freedom usually increase, raising memory and solution time. The benefit is that $E_h$ may decrease and gradients, waves, or thin layers are represented more faithfully. A finer mesh, however, cannot repair an incorrect physical model. In time-dependent problems, reducing spatial size may also require reducing $\Delta_{t}$ to preserve stability. Good discretization balances detail, stability, cost, and physical fidelity. It also requires repeatable refinement evidence and checks that target quantities change little under refinement while physical balances remain coherent.

The same domain can be discretized in very different ways depending on which scale is deemed relevant: properly resolving a thin boundary layer requires a much smaller $h$ than the rest of the domain, even when the overall geometry is simple.

🔵 Formal level

Basic formulas

For a uniform one-dimensional mesh, cell count is estimated by dividing the $L$ by the chosen cell size:

Number of cells in a one-dimensional discretization

The algebraic problem size is approximated by:

Estimate of degrees of freedom

In the asymptotic regime, representation error follows:

Order-p discretization error

Here $p$ describes reduction rate, while $C$ collects method, geometry, regularity, and norm.

Real student questions

Do more cells always mean a better result? No. They must resolve relevant scales and the model must be correct.

Are nodes and degrees of freedom identical? No. Degrees depend on variables and constraints.

Why can a fine mesh be unstable? The time step may no longer be compatible with spatial size.

When should refinement stop? When target error is met and another source starts to dominate.

Is adaptive refinement always better? It helps only when the indicator identifies physically important regions.

Final synthesis

Modeling decides which physics to preserve. Discretization decides how to represent it with finite resources. Reliable computation balances resolution, error, stability, and cost without mistaking numerical detail for physical truth.