Definition
Average normalized deviation between actual measured outcomes on equipment subject to multiple AI models and the sum of those models' independent predictions: detecting whether fouling, corrosion, and process optimization models that share physical systems are producing recommendations that interact adversely without any individual model accounting for the effect.
The Measurement Gap
Heat exchangers, reactor systems, and piping circuits commonly have three AI models acting on them: a fouling model recommending cleaning, a corrosion model recommending inspection, and an APC/RTO optimizing throughput or severity. Each model is evaluated on its own predictions against its own outcomes. None of them is evaluated on whether their combined effect matches reality, because no metric exists for the gap. The most common failure mode is aggressive cleaning that exposes fresh metal to corrosive fluids, accelerating corrosion two to five times the steady rate for thirty to sixty days post-cleaning. The corrosion model does not see the cleaning event. The fouling model does not see the corrosion acceleration. Individual model accuracy looks acceptable. Equipment degrades faster than any individual prediction would suggest. Heat exchanger performance, corrosion rate, and on-stream factor each measure a different surface of the system; none measures the interaction.
Decision Supported:
Are the fouling model, corrosion model, and APC/RTO producing recommendations that are coherent on shared equipment, or is each model's success consuming the operating envelope of the others?
Formula
Plain language: For each piece of equipment touched by more than one AI model, compare what actually happened to the sum of what each model independently predicted would happen. The gap, averaged across equipment items, is the interaction effect.
Notation:
CMII = ( 1 / P ) × Σ | Outcome_actual_p − Outcome_independent_prediction_p | / Outcome_independent_prediction_p
Variables:
| Variable | Definition | Unit |
|---|---|---|
| P | Number of equipment items with multiple AI models operating on them | count |
| Outcome_actual_p | Actual measured outcome for equipment p over the measurement period (wall-thickness loss, heat-transfer degradation, etc.) | engineering unit |
| Outcome_independent_prediction_p | Sum of each model's independent prediction for outcome on equipment p | engineering unit |
| p | Index of shared-equipment item | n/a |
Report CMII per equipment class (heat exchangers, reactors, piping) and per outcome dimension (wall loss, heat-transfer coefficient, fouling resistance).
Proposed formula, pending validation pilots.