Product · Tieback
Subsea. The state you cannot see.
Subsea tree pressure estimated from topside readings, with an interval. Severe slugging flagged early. Largest stable choke opening per well. To be scored on public well events against a gauge the estimator never sees.
ConceptA proposal with references. Not built.
Engine · Two-phase observer bank on reduced well-flowline-riser models
1
Subsea pressure from topside channels, with an interval
2
Severe slugging flagged early (target; lead time unmeasured)
3
Largest stable choke opening per well
Needs
- Topside choke pressure and temperature, choke opening, gas-lift rate
- Well and flowline geometry, or a normal-operation window to identify it
- Separator trip and choke logs for the score
- A flow-assurance engineer to check field realism
Is not
- Not an anti-slug controller; never writes to the choke
- Not a transient multiphase simulator
- No hydrate margin yet (phase 2)
- No riser-base gauge in the public data
- Not built. No field result.
Where it stops
Wavepin estimates single-phase pipelines. Tieback estimates two-phase wells, flowlines and risers. Shares the tournament, ledger and hypothesis layer; later one two-phase module with Annulus. Soft sensors on real well data exist; the claim is narrower: a public split, a hidden gauge, calibrated intervals.
Built from
What exists, what does not
pipeintel: tournament protocol (frozen thresholds, out-of-distribution splits, failure capture), ledger, measured-against-inferred types, signature-bank fit for choke-restriction and sensor-bias families, the safety-validator pattern for proposed choke openings; about 12 %, scaffolding only. New: the reduced well-flowline-riser model, the observer bank, an oscillation detector, the Hopf-point choke report, a public well-event loader, black-box soft-sensor baselines. Phase 2: thermal model and hydrate curve; two-phase truth shared with Annulus.
First proof
The experiment
Three to four weeks, one person. A public labelled well-event dataset, real instances only, counts per class. Subsea tree pressure estimated from topside choke pressure and temperature, with choke opening and gas-lift rate as inputs; tree pressure hidden and scored. Parameters identified per instance on a normal window, scored on a held-out window and on unseen wells. Observer bank against ARX and NARX models, gradient-boosted trees, recurrent soft sensors and the published event classifiers.
Tournament
Scored against
Real split: a public labelled well-event dataset, real instances only, with severe slugging, flow instability and hydrate-in-line as a confuser. Estimator sees topside choke pressure and temperature; choke opening and gas-lift rate are inputs. Hidden and scored: subsea tree pressure. Simulated truth from a different-structure model for slug libraries across geometry, gas-oil ratio, inflow and choke. Against: oscillation threshold, published event classifiers, ARX and NARX, gradient-boosted trees, recurrent and convolutional soft sensors, extended and unscented Kalman, constrained moving-horizon. Scored: tree-pressure error on the hidden gauge, 90 % interval coverage, slug lead time at a frozen false-alarm rate, false alarms per day, stable-choke error against simulated truth.
Kill rule
The observer bank does not match black-box soft sensors on unseen wells with calibrated intervals: the engine is not the moat, and the ledger says so.
Methods
- Reduced well-flowline-riser models with per-well parameter identification
- Extended and unscented Kalman filters; constrained moving-horizon estimation with holdup bounds
- Controllability limits of topside-only measurement
- Largest stable operating point from the linearised model's Hopf point
Rules and standards
Also
See Tieback on your asset.
A proposal with references. Not built. A conversation first; a data pilot when it is ready.