What would make the prediction accurate?#
Suppose a calculation predicts a sideways force of 100 N. Do we need to know it within 20 N, 2 N, or 0.2 N? There is no useful claim of “accurate” without saying what quantity matters and what error is acceptable.
An observable is a quantity we can compare, such as cloth shape, total force, local pressure, or frequency, the number of oscillations per second (Hz = 1/s). Different observables can have different errors in the same run.
Verification asks whether the program solves its chosen mathematical model correctly. Validation asks whether that model and calculation predict measurements of the physical system. NASA distinguishes these in its verification and validation overview. Solving the wrong physical approximation more precisely is not validation.
Choose comparable physical conditions#
Airflow depends on size, speed, density, and viscosity. Before a useful comparison, define:
\(\rho\): air density (kg/m³); choose 1.225 for the example.
\(U\): incoming speed (m/s); take 6 for the example.
\(c\): representative chord, the width along the flow (m); choose 1 m as a comparison length, not the whole changing outline’s width.
\(\eta\): dynamic viscosity (Pa s); use the illustrative \(1.8\times10^{-5}\) from Chapter 3.
The Reynolds number, written \(\mathrm{Re}\), compares acceleration effects with viscous effects on this speed and length scale:
The units cancel. Our example gives
NASA’s Reynolds-number explanation derives this ratio. It helps compare experiments, but the same Reynolds number does not guarantee the same cloth shape or tell us precisely where separation begins. Separation is air peeling away from cloth; turbulence is irregular motion over many sizes of swirling region. Predicting them requires suitable equations, resolution, and possibly additional turbulence models.
Measure rather than invent the material#
Record the manufactured cut and seams, where the cloth is attached, its mass per area, and force versus stretching, shear, and bending. Pretension is tension applied before wind loading. Permeability is ease of air passage through cloth. Measure these if relevant. Warp and weft are the two crossing thread families; testing both matters even if a teaching example chooses equal stiffness.
Calibration chooses parameters to fit measured behavior. To test predictive value, calibrate on some conditions and validate on different ones. Fitting every experiment is not an independent prediction. Record air speed, direction, fluctuations, and properties too. Chapter 1’s illustrative stiffnesses are not measurements of this sail.
Make the numerical calculation more precise#
Spatial resolution is how finely space is divided. A refinement study repeats the same physical problem with smaller mesh cells or patches. Refine cloth and air separately as well as together. Resolve the boundary layer, the thin air region with rapid velocity change near cloth, free edges, and the disturbed downstream wake. Move the artificial outer boundaries farther away to check their influence.
Temporal resolution is the size of the time step. Reduce it while keeping other errors small. More image frames do not necessarily mean smaller simulation steps.
Iterative resolution concerns repeated attempts to satisfy equations within a step. Tighten permitted equation mismatch and increase air/cloth subiterations. Convergence means predictions approach a limiting value under these changes, rather than merely remaining finite.
For a simple comparison, let \(Q_f\) and \(Q_c\) be the same scalar observable on a finer and coarser calculation. Subscripts name “fine” and “coarse,” not different physics. Both have the same units, for example N for force. Let \(e\) be their relative difference, with no units. Vertical bars mean absolute value, ignoring sign. If \(Q_f\) is nonzero,
A finer value of 100 N versus a coarser 95 N gives \(e=0.05\), or 5%. This formula is just a normalized difference, not an error estimate against nature. Near zero, compare the absolute difference against a documented scale instead.
NASA’s spatial-convergence tutorial explains more rigorous refinement assessments, and its temporal-convergence tutorial treats time steps. Two grids alone do not establish a reliable convergence rate. Refinement cannot recover viscosity, wake history, or cloth behavior that the chosen model leaves out.
Compare with an experiment and its uncertainty#
Use a fixed mast and boom in controlled airflow. Measure three-dimensional shape, total forces, support reactions, local pressures where feasible, and air velocities or flow visualization. A support reaction is the force needed to hold an attachment still. Repeat conditions and estimate measurement uncertainty, how much the measured value could differ from the unknown true value.
For illustration, a prediction of 100 N and a measurement of 99 N with uncertainty ±3 N are compatible with that measurement, but do not establish 1% prediction accuracy. Agreement could also result from compensating errors. NASA’s validation assessment discusses comparison with experimental evidence.
Keep errors distinct: wrong material properties, missing physics, finite mesh and time-step errors, coupling lag, and measurement uncertainty. A residual is the amount by which the current numbers fail to satisfy their chosen equations. A tiny linear-system residual shows that one algebraic solve was accurate; it does not show the model predicts reality. This is why accuracy must be assessed before interpreting an attractive browser scene.