The heavy-hydraulics programme asks when a smaller fixed gear pump on an open-centre valve burns less fuel than load-sensing on a small excavator. Over 384 measured cylinder-geometry readings, the answer turns on the gear pump. Earlier today a fifth family, Rexroth's AZPG, showed that one number (efficiency at 150 bar) is not enough. A second number, the pump's power draw at zero pressure, predicted it within 0.005.
That was one held-out family against a rule calibrated on one shape. This run asked the harder question: if every family takes a turn as the one the rule has never seen, how far off is it?
Twenty-five pumps that don't exist
Each family brings two things: the shape of its efficiency map (how efficiency moves with size and pressure) and its no-load drag. Crossing all five shapes with all five drags gives 25 hypothetical pumps. Each was swept in efficiency over the same 384 geometries, and the efficiency at which the fixed pump wins in half of them (E50) was read off.
| range over 25 cells | |
|---|---|
| E50 | 0.8010 to 0.8300 |
| moved by shape (row means) | 0.0096 |
| moved by drag (column means) | 0.0194 |
With the AZPG's low drag (2.58 kW at 85 L/min) in the set, drag now moves the threshold twice as much as the map's shape does.
Leave one family out
For each family, the rules were refitted on the 16 cells that use neither its shape nor its drag. Then they were asked about the family as it is.
| left out | true E50 | one number | efficiency + drag | + shape term |
|---|---|---|---|---|
| Casappa | 0.8277 | +0.0130 | +0.0037 | +0.0015 |
| Danfoss | 0.8297 | +0.0141 | +0.0038 | +0.0049 |
| Marzocchi | 0.8193 | +0.0026 | −0.0002 | +0.0069 |
| bench KP30 | 0.8116 | −0.0073 | +0.0006 | −0.0008 |
| Rexroth AZPG | 0.8040 | −0.0166 | −0.0042 | −0.0056 |
| RMS | 0.0119 | 0.0030 | 0.0046 |
Efficiency plus drag is four times as accurate as efficiency alone, and never off by more than 0.0042. Its slope stays between 0.0162 and 0.0198 per kW whichever family is missing.
A third number for shape, the change in efficiency from 85 to 150 L/min, fits the 25 cells better. But it predicts a left-out family worse. With five folds that is the signature of overfitting, so the rule stays at two numbers. Refitted on all 25 cells:
E50 = 0.824 + 0.018 (kW − 3.57)
where kW is the pump's zero-pressure power at 85 L/min of flow.
What it means for a builder
Two numbers on a gear pump's data sheet matter: overall efficiency at 150 bar, and the power curve at zero pressure. Nothing else in five families' maps moves the answer by more than about 0.004.
Six predictions were written before the sweep. Three held: the slope (0.0179, inside 0.016 to 0.021), the worst held-out error under 0.010, and two numbers beating one. Three failed: the refit's intercept moved 0.0036 from the old Casappa-only one (the bar was 0.003), the shape term did not help, and one cell sat 0.0202 from the pooled 0.821 (the bar was 0.020).
Limits: five families are five folds, so these are five errors, not a distribution. All five are typical catalogue or one-unit curves, and the drags span only 2.58 to 3.64 kW. A family dragging more than 4 kW or less than 2 kW would test the slope where no fold has been.
Code: gear_pump_five.py, results in gear_pump_five_results.txt, pre-registered predictions in the log.