Puckworks · a model-composition story

More Physics Made It Worse

Scientific components do not automatically become a better model when they share a variable.

EXPLORATORY_SIMULATION evidence: qualitative

Loading the generated data…

Summary (JavaScript is off)

This page normally lets you add model components one at a time and watch the fit change. With JavaScript disabled, the complete result — every number and caveat — is in the generated static text summary and the summary image below. You can also download the generated data.

Three flow-versus-time curves over one espresso shot. The observed flow rises and levels off. An extraction-only model tracks that rise closely. Adding a swelling branch makes the composite prediction a flat line at the wrong level that misses the rise. A lower panel shows the shared porosity: extraction opens it slightly above its reference, while the composite over-closes it well below the reference. The exact numbers are in the static text summary linked above.

1 · The target and the simplest baseline

An espresso shot's flow rises as the puck opens. The observed 9-bar flow is our target. A constant is deliberately the simplest possible prediction — the minimum benchmark a more complex mechanism should beat. Its error over the ·–· s window is · g/s at a level of · g/s.

Build the model, one component at a time

Which prediction to show against the observed flow

Predicted flow (selected model) vs the observed 9-bar trace. The shaded band is the scored window.

·

Show the numbers behind the flow plot (text table)

2 · Extraction-only improves the fit

The time-varying extraction branch follows the rising target far more closely: its error is · g/s — about ·× better than the constant, with · free parameters. An independent producer (degeneracy_rmse) reproduces it at · g/s.

A lower residual does not prove that this mechanism is uniquely identified or physically correct. This is verification against one trace, not independent validation.

3 · Add the tested swelling composition

Both branches modify the same porosity state. Adding an imported swelling branch over-closes that state — it drops to · (below the · reference). The dissolution-driven opening term floors, so the composite prediction decouples from the rise and its error climbs to · g/s — ·× the constant baseline and ·× the extraction-only model.

This does not mean the swelling component is universally invalid. It means this composition — these two branches through one shared state on a saturated pre-wet rig — is incompatible.

4 · Model components need contracts

Before two components are combined, the composition should answer:

More equations are not automatically more explanation. A model composition must earn its interfaces.

What this does not prove

Scope. ·

Caveat. ·