The numbers, and their limits.
Tournament t001, 1500 held-out simulated two-hour drilling runs (557 with influx) on a 3,000 m well, alarm thresholds frozen at 2 % false flow checks per run on validation; ballooning never seen in training. The bar and the result are on the proof page.
Kick detection volume (bbl)
See the tables below.
Response time (min)
See the tables below.
False flow checks per 100 rig-hours
See the tables below.
Ballooning misread rate
See the tables below.
Same cases, same rules
Scored against, untouched
- Pit-gain threshold
- Delta-flow threshold
- Trend rule
- Connection flow-back fingerprinting
- A lumped observer
- A learned model
Thresholds set on validation data and frozen for every test, including the cases the engine never saw. The truth model differs from the engine's own.
The bar, set before the first run
Less mud gained at the alarm than flow-back fingerprinting and a lumped observer, at 1 false flow check per 100 rig-hours, beyond the interval.
Against flow-back fingerprinting: +5.9 bbl (90 % interval +3.4 to +7.2) and -1.5 min (-2.1 to -0.8) at detection. Against the lumped observer: +10.4 bbl (90 % interval +8.7 to +11.7) and +0.2 min (-0.3 to +0.7) at detection. Not better than both beyond the interval.
First proof
Four to six weeks: single-phase, water-based mud; a U-tube with pump, open flowline and pit; 2,000 scenarios with connections, ramps, ballooning and transfers; detection volume and response time at the frozen rate; truth from an independent second model.
Detection at the same false-alarm rate
| Influxes caught | Gain at alarm, median | Response, median | Caught by 2 bbl | False checks / 100 h | |
|---|---|---|---|---|---|
| Annulus: signature fit on the U-tube | 99 % | 25.1 bbl | 7.2 min | 0 % | 1.01 |
| Pit-gain threshold | 99 % | 24.5 bbl | 7.6 min | 0 % | 1.01 |
| Delta-flow threshold | 95 % | 21.3 bbl | 8.2 min | 0 % | 0.90 |
| Pit trend rule | 96 % | 25.7 bbl | 8.7 min | 0 % | 0.74 |
| Connection flow-back fingerprint | 99 % | 19.2 bbl | 8.8 min | 0 % | 0.80 |
| Lumped observer | 98 % | 14.3 bbl | 7.0 min | 0 % | 0.85 |
| Learned model (GRU) | 97 % | 55.9 bbl | 10.6 min | 0 % | 0.53 |
Every method's threshold set on validation at 2 % false alarms per two-hour run (one false flow check per 100 rig-hours) and frozen. Gain at alarm: the simulated influx volume between its start and the first alarm.
Look-alikes: alarms that would have been false
| Ballooning (held out) | Pit transfer | Sensor fault | Influx caught in a connection | |
|---|---|---|---|---|
| Annulus: signature fit on the U-tube | 1 % | 1 % | 9 % | 99 % |
| Pit-gain threshold | 0 % | 15 % | 0 % | 99 % |
| Delta-flow threshold | 0 % | 0 % | 13 % | 95 % |
| Pit trend rule | 0 % | 11 % | 0 % | 95 % |
| Connection flow-back fingerprint | 0 % | 8 % | 4 % | 99 % |
| Lumped observer | 0 % | 7 % | 5 % | 97 % |
| Learned model (GRU) | 0 % | 0 % | 8 % | 96 % |
Share of look-alike runs that raised an influx alarm. Ballooning: the held-out split, a class no method saw in training or validation.
Against the bar: flow-back fingerprinting and the lumped observer
| Gain at alarm | 90 % interval | Response | 90 % interval | |
|---|---|---|---|---|
| vs flow-back fingerprinting | +5.9 bbl | +3.4 to +7.2 | -1.5 min | -2.1 to -0.8 |
| vs the lumped observer | +10.4 bbl | +8.7 to +11.7 | +0.2 min | -0.3 to +0.7 |
Median difference (Annulus minus baseline) on the test influxes, misses scored as the run's final gain and 120 minutes. Negative is better.
What is proven. What is not.
Proven, in simulation
- Raises an alarm 1.5 minutes sooner than flow-back fingerprinting at the same false-flow-check rate (90 % interval 0.9–2.1 min; simulation, t001)
- Pit transfers read as influx in 1 % of runs; pit-gain threshold 15 %, flow-back fingerprinting 8 %, lumped observer 7 %
- Ballooning never seen in training read as influx in 1 % of held-out runs
Not proven
- Against the bar: 5.9 bbl more gained at alarm than flow-back fingerprinting and 10.4 bbl more than the lumped observer, both beyond the interval
- The lumped observer gains least before its alarm: 14 bbl against 25 bbl
- The influx-rate interval covers 5 % of true rates, not 90 %
- At three times the channel noise it catches 66 % of influxes; the baselines 97–99 %
- Sensor faults read as influx in 9 % of runs; pit-gain threshold 0 %
- No labelled public field kick exists; no gas kick, no oil-based mud, no managed-pressure drilling
The first real number comes from your data.
A month of your exports, replayed offline, gives the figure that matters on real operations. We will show it to you whether it is good or bad.