Corrosion Model Operating Envelope Alignment (CMOEA)

GMI-OG-015Oil and GasIntegration-Required

Definition

Corrosion Model Operating Envelope Alignment, or CMOEA, is the degree of overlap between the operating conditions in the AI corrosion growth model''s training data and the current operating conditions of the pipeline segment, measured across the set of parameters known to influence corrosion. The KPI exists to answer one decision: are the AI''s corrosion growth predictions still valid given that the pipeline''s operating envelope may have moved since the model was trained?

The Measurement Gap

Reza is the corrosion engineering lead for a refined-products pipeline operator running about twenty-two hundred miles of liquids pipe across the upper Midwest. Two and a half years ago the company adopted an AI corrosion growth model from a major IMS vendor, a model that ingests pipe history, environment, CP records, and ILI run data and produces predicted corrosion growth rates per segment that the integrity program uses for remaining-life calculations and reassessment-interval planning. The model''s validation report at deployment showed strong performance: predicted growth rates within fifteen percent of measured growth on the validation dataset across a five-year window.

The vendor''s validation dataset described a pipeline system that looked similar to Reza''s: refined-products service, mid-continent geography, mostly twelve-to-twenty-inch pipe, throughput in the forty to sixty-five thousand barrel-per-day range. Reza''s pipeline was a good fit when the model was deployed. Twenty months later, Reza''s pipeline was no longer the pipeline the model had been trained on. The operator''s commercial team had taken on a new contract that increased throughput on a major spur from fifty-five to eighty-two thousand barrels per day. The product mix had shifted toward a heavier mid-distillate slate that operated at slightly higher temperature. The new pressure profile pushed several segments to operate near MAOP rather than the eighty-percent-of-MAOP average the model had been trained on. The CP program had been retuned to handle the new operating envelope. Every change individually was a managed-change item that had passed the operator''s MOC reviews. Cumulatively, the pipeline''s operating conditions had moved.

The model''s output, though, kept reporting the same shape of corrosion growth predictions. The predicted growth rates against the validation dataset assumptions were still well-behaved. The model''s outputs looked like the outputs from two years ago. The pipeline was no longer the same system.

The next ILI run told the story. Several segments showed corrosion growth at rates between forty and ninety percent higher than the model had predicted for the inter-inspection interval. The specific segments were the ones whose operating envelope had moved the furthest from the training data. The model had not failed in the usual sense: its math was unchanged, its inputs were being processed correctly, its validation metrics on the original dataset would still look fine if re-run. What had happened is more subtle. The pipeline''s operating conditions had moved outside the envelope the model''s training data covered, and the model was extrapolating into a region where its predictions were unreliable, while reporting them with the same confidence as predictions inside the envelope.

Legacy corrosion model metrics do not catch this. Validation accuracy on the original dataset stays good because the original dataset has not changed. Inter-run growth comparisons can flag the discrepancy after the next ILI run, but that is reactive: the operator finds out the model was wrong only after the inspection that the model''s growth predictions informed has been run, and the interval between inspections is exactly where the model''s prediction is supposed to guide the program. Standard model-drift KPIs measure whether the input data distribution has shifted, which is a related but different question. The operating conditions can shift without the input data distribution looking dramatically different to a generic drift detector, especially when the shift is in parameter ranges rather than parameter identities.

What is missing is a metric that asks whether the pipeline the model is currently predicting on is still the pipeline the model was trained on.

Formula

Corrosion Model Operating Envelope Alignment, or CMOEA, is the degree of overlap between the operating conditions in the AI corrosion growth model''s training data and the current operating conditions of the pipeline segment, measured across the set of parameters known to influence corrosion. The KPI exists to answer one decision: are the AI''s corrosion growth predictions still valid given that the pipeline''s operating envelope may have moved since the model was trained?

The structural form is parameter-by-parameter. For each corrosion-influencing parameter (flow rate, product composition, operating pressure, operating temperature, CP potential, inhibitor concentration, water content, sediment loading) compute the overlap between the current operating range on the segment and the training data range on that parameter. The overlap is the fraction of the current operating range that is inside the training data range. Average across parameters. CMOEA of 1.0 means the current operating conditions are entirely inside the envelope the model was trained on. CMOEA of 0.5 means half the current range, averaged across parameters, is in a region the model is extrapolating into.

The parameter set is the part of the metric that determines whether it is honest. The corrosion-influencing parameters depend on the service: refined products, sour crude, sweet crude, natural gas, NGLs, and other products each have a different parameter set that materially affects corrosion. The integrity engineering function has to define the parameter set per service, drawing on NACE guidelines, internal corrosion specifications, and the operator''s own historical corrosion experience. A CMOEA computed on a parameter set that omits a critical driver (say, omitting sulfur content on a crude service line) will look reassuring even when the omitted parameter has moved.

The current operating range is computed from the SCADA history over a defined recent window, typically the most recent six to twelve months, sometimes a full inter-inspection interval. The window should be long enough to capture seasonal variation and operational range, short enough that conditions are representative of the current state rather than ancient history. The training data range is the corresponding range in the model''s validation documentation: what flow rates, pressures, temperatures, and so forth were represented in the data the model learned from. If the vendor does not document these ranges, the metric is asking for information the vendor needs to provide; an evasive vendor here is itself a signal.

Proposed formula, pending validation pilots.

At a Glance

Category
Pipeline & Midstream Integrity Intelligence
← Operational Intensity Lifecycle Adjustment (OILA)ILI Threat Mechanism Coverage Index (ITMCI) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment