Stan Logie and Reid Hasher - BKV Corporation
David Benham, Ryan Benn, Donny Lowry - ROAM-AI
Water carryover - produced liquid entrained in the gas stream leaving the separator instead of reaching the tank - accumulates in gathering lines, raises backpressure, and drives low-pressure wells to underperform. The mechanisms are known: stuck dump valves, failed floats, undersized separators. However, detection is not. Carryover was often found 48 to 72 hours after onset, if not longer.
No instrument directly measured the carryover. Facilities carried water tank levels, gas meter data, tubing and casing pressures, plunger run counts, and compressor data. No liquid meters, water-content measurement, or separator level telemetry was available, and none was added. What signal existed sat across multiple source systems.
METHOD
Procedural AI encodes the operating procedure itself - the ordered rule-out a production engineer would follow - as the model's reasoning instructions, in the field's own language and under version control, rather than training a statistical model to infer the procedure from data. Field adjudications amend the procedure; they do not retune a black box. The primary signal is a deficit: a facility's tank fill-rate slope falling below its own 30-day trailing baseline while gas rate holds, indicating liquid that should have reached the tank and did not. Vessels are aggregated into one equivalent tank and converted to volume using proper tank strapping conversions, since commingling and shared vessels mask slope change.
Two layers run in sequence. A deterministic screen, hand-coded from domain knowledge, rejects known non-carryover explanations: trailing downtime, compressor state, pigging, workovers, shared vessels. A procedural AI routine ranks the survivors, trained on the adjudication record. Each alert carries it’s a written explanation of its evidence, barrel estimates both as a daily rate and a total barrels lost, and a remediation checklist.
RESULTS
Nine events were confirmed in the field over six weeks, seven caught specifically by carryover logic. Each began as a fill-rate deficit and ended in a repair. Alerts requiring field validation fell from an average of 83 per week to 6 per week. The system now surfaces a small number of high-confidence, actionable alerts rather than a large volume requiring manual triage.
The first field session returned a negative result: 17 of 18 candidates were rejected, every rejection traced to a missing input rather than faulty inference, and was treated as an input specification rather than a verdict on the method. Shared-vessel mapping and a partial-downtime rule cut wasted leads from 53.5 to 2.6 percent. After a sub-twelve-barrel carry over event was missed, the alert floor was adjusted with an Off-Trend Watch tab added to surface lower-volume deviations without overwhelming the primary alert list.
ADOPTION
Projects of this kind fail on adoption more often than on inference. That is the model is defensible, but the field does not trust it, and the alerts go unworked. The work followed a deliberate sequence - problem, concept, prototype, adoption - with the field entering at prototype rather than at handover. Twelve twice-weekly sessions over six weeks put candidate alerts in front of named operations personnel, whose accepts and overrides were attributable and not editable after entry.
TRANSFER
The method adds no instrumentation. It needs a measured delivery rate, a trailing baseline, a list of benign explanations, and an attributable record of what the field accepted and rejected. The pattern holds wherever a fluid should arrive somewhere and its absence is the first symptom. Carryover was not the only result of value: downtime coding and gas captured from hung dumps emerged while ruling it out.