Written by an AI agent working in this repository.
The heavy-hydraulics programme asks when a small excavator should carry a cheap fixed gear pump instead of a load-sensing piston pump. Its rule takes two numbers from a pump catalogue: the efficiency at 85 L/min and 150 bar (the index) and the idle drag in kW. From those it predicts the index efficiency at which the gear pump wins on fuel in half of 384 measured cylinder geometries.
Earlier today a third number read the pump family's efficiency at 57 L/min against its index. Left out one maker at a time, it cut the rule's error by 18%. We had registered 20% as the bar. That run's next step asked a plainer question: which pump sizes does the fuel optimum actually pick?
Where the picks fall
For each geometry the model sweeps fixed-pump sizes and keeps the one with the best fuel break-even. Read at each family's own crossing, over six pump families (2304 picks):
| flow band, L/min | share of picks |
|---|---|
| below 50 | 0.0% |
| 50 to 70 | 41.5% |
| 70 to 90 | 47.4% |
| 90 to 110 | 11.1% |
| above 110 | 0.0% |
The median is 72 L/min, and the 10th and 90th percentiles are 58 and 90.5 L/min. So the 85 L/min index sits above most picks, and the earlier 57 L/min reading sits near the bottom of the range.
The test
Before computing any of the numbers above, we registered six predictions in the log (7e). The new feature averages each family's efficiency over those picks and divides by its index. A second feature counts the share of picks above a family's largest size. Scoring matched step 7d exactly: 14 pump readings, 196 shape-by-drag cells, each maker family's readings left out together. Both of the earlier rules reproduced their numbers to four places.
| rule | left-out RMS, 14 readings | held families only | geared pumps |
|---|---|---|---|
| efficiency + drag | 0.0084 | 0.0022 | 0.0051 |
| + shape at 57 L/min (7d) | 0.0068 | 0.0022 | 0.0040 |
| + shape over the picks (7e) | 0.0069 | 0.0017 | 0.0042 |
| + share of picks above the largest size | 0.0093 | 0.0027 | 0.0049 |
- The weighted shape number also cut the error by 18%, so the headline prediction (20%) is refuted. It did not beat 57 L/min either (refuted).
- Its coefficient is negative, as the mechanism says (held).
- On real catalogue pumps it gives the lowest error the programme has had, 0.0017 (held). It fixes the Casappa and Danfoss misses the 57 L/min number left or overshot.
- The share above the largest size does not help (held, as registered). It is a constant for each family, so it cannot tell a pump run as sold from one spun faster through a gearbox.
- The geared Rexroth AZPN still misses, by 0.0082 (refuted).
Three of six predictions held.
What it tells us
Two shape numbers chosen for different reasons gave the same 18%. That looks like the most a shape read at 150 bar can buy here. The geared AZPN's miss sits somewhere else. One candidate stands out: the picks centre on 72 L/min, but the rule reads the drag at 85 L/min. A gearbox adds a fixed loss and spinning raises drag per litre, so reading drag at the wrong size should hurt a geared pump most. That is the next registered test.
The paper keeps the two-number rule as its headline. For a builder comparing catalogue gear pumps, read the efficiency curve between about 60 and 90 L/min, not only at one point.
Code: gear_pump_chosen.py, results: gear_pump_chosen_results.txt. The catalogue curves are typical figures, not tests of individual pumps. This is a modelling result, not advice about any machine.