Session: RIP: Predictive Modeling and Characterization of Corrosion Processes in Complex Environments (In Honor of Professor Digby Macdonald) (Part III of IV)
Expert-Guided AI Approach to Improve Remaining Life Assessments Using ILI Data & Situational Context (RIP2026-00092)
Estimating corrosion rates and projecting remaining useful life in pipelines is a difficult problem in practice because several sources of complexity interact simultaneously, from data sensing to the mechanisms that cause corrosion. Multi-year in-line-inspection data alignment and consistent feature matching can introduce significant uncertainty into pit-to-pit growth estimates. When alignment is completed, the resulting growth calculations are affected by measurement inaccuracy arising from changes in ILI tool type, tool vendors, and technology resolution over time. On top of these data issues, standard remaining life methods often treat corrosion defects of one or more types in the same way, even though localized external factors can cause them to grow at very different rates. This combination can lead to both overlooked high-risk locations and over-conservative dig lists that drive up program costs. Current approaches rarely incorporate explicit models of localized corrosive influencers when assigning growth rates and due dates. This paper presents ongoing research of an expert-guided causal AI framework that aims to transform historical ILI runs into situation-specific remaining useful life estimates for optimized dig programs. The approach begins with automated multi-run ILI alignment that performs one-to-one, one-to-many, and many-to-one matching features reported across the years, producing a traceable time series of metal loss at each location. A second step applies statistically grounded, industry informed calibration to reduce run-to-run bias and noise ensuring the calculated growth rates reflect more realistic behavior rather than tool artifacts. Building on this calibrated baseline, a causation-oriented AI model incorporates corrosion agent information and external influencers and AI guardrails derived from expert hypotheses, to assign situation specific growth rates and to project future sizing at the next ILI run. The model is designed to be transparent, so that each projection can be traced back to the situational factors that influenced it, rather than treated as a black box output. Early results on selected pipeline segments show that standard approaches often assign similar growth rates and due dates to features that are exposed to meaningfully different conditions, while the proposed causation-based approach filters through situations and separates growth behaviors more clearly even with limited situational data. The presentation will share methodology, preliminary findings, and planned validation steps across additional assets and environments. Once mature, the causal AI framework has the potential to improve the predictions of corrosion growth rates and estimated remaining life, identifying areas of concern earlier while drastically reducing unnecessary digging expenses.
