Every September, the International Pipeline Conference fills a Calgary convention center with people who spend their careers worrying about things that are, by design, hidden underground. IPC2026 (September 21–25) brings four Irth sessions to the program this year, three papers and one presentation, and they all circle the same question: when a model, a simulation, or an inspection tool tells you a number, how much should you believe it?
That question sounds almost philosophical until you're the one deciding whether to dig, whether to reinspect, or whether to sign off on a risk model that regulators will eventually ask you to defend. Our team, working with C-FER Technologies, TC Energy, and PRCI, put together the evidence below. Here's what we found, and where to catch each one live.
Where to find us at IPC2026:
Tuesday, September 22, 11:40–12:20, Exhibition
The first one doesn't build a model at all. It asks a more basic question about the data everyone else's models depend on: how good is inline inspection at reporting the depth it says it's reporting?
We pulled from Irth's Asset Integrity for Pipelines (AIP) platform, which, as of this analysis, held data from almost 16,000 inline inspections covering 130,000+ miles of pipeline across 33 operators in the US, Canada, and South America, inspected by 30+ different ILI companies. Out of that, we isolated 112,000 metal loss anomalies with both an ILI-predicted depth and a field-verified “truth” depth, and measured certainty (the percentage of predictions landing within a stated tolerance) the way API 1163 defines it.
A few findings should change how people talk about ILI performance. First, depth itself is the single biggest driver of accuracy, more than ILI company, more than technology, more than classification type: certainty within a 10% tolerance falls from roughly 84% for shallow anomalies down to the 35–46% range for the deepest ones, a well-documented pattern for MFL that this dataset confirms at scale. Second, ILI company matters enormously, with certainty within 20% ranging from the mid-40s to the mid-90s depending on vendor, even after excluding the smallest sample sizes and one statistical outlier. Third, and this one surprised us: ultrasonic metal loss (UT ML) technology posts the best overall certainty, but when you isolate just the top-quartile ILI companies running standard-resolution axial MFL, MFL-A outperforms UT. The technology matters less than who's running it.
Wednesday, 11:00 am, Conference
What happens when you train a model on pipeline data and split it the way every introductory ML course tells you to, i.e., randomly? You get a great-looking score, and it's largely fiction.
The problem is autocorrelation. Pipeline integrity data isn't a bag of independent observations. It's hierarchical: system-level attributes like vintage, seam type, and coating repeat across hundreds or thousands of joints, and sub-joint measurements cluster locally on top of that. A random 80-20 split doesn't separate training from testing so much as it shuffles the same system's fingerprints into both piles. The model doesn't generalize. It memorizes and calls it learning.
Anton and Michael quantified this on a real operator dataset using a coating-type imputation task with XGBoost. Under random splitting, macro-F1 came in at 91.5%. Under a validation protocol blocked by system_id, the same model scored 77.3%, a 14-point drop, with individual classes (Paint, in particular) falling by as much as 35 points. Nothing about the model changed between those two numbers. Only the honesty of the evaluation did.
What if the thing making your model look good is the same thing that will make it fail on the next system it sees? That's the mechanism the paper names False Confidence, and it's why the recommendation isn't “collect more data” but “define your unit of independence and block your splits accordingly,” whether that's system, spatial region, or inspection run. They also make a case against a very common piece of hygiene: duplicate removal. In clustered pipeline data, identical-looking rows are often legitimate repeats of system-level attributes across joints, not ingestion errors, and stripping them out compresses real structure rather than cleaning it up.
Thursday, 4:00 pm, Conference
The second paper stays in the same neighborhood, i.e., don't trust a number until you know how much it's allowed to wobble, but applies it to probability of failure (POF) rather than classification accuracy.
POF for pipeline features typically lives in the 10⁻³ to 10⁻⁵ range, which means you're trying to estimate something that, by definition, almost never happens. Standard Monte Carlo simulation handles this by throwing enormous sample counts at the problem, often millions of trials, without ever reporting how precise the resulting estimate is. A fixed sample size can quietly oversample an easy case or undersample a hard one, and you'd have no way to tell from the number alone.
Using the PRCI MAT-8 fracture mechanics model with Paris-law crack growth, the team compared Standard Monte Carlo (SMC) against Latin Hypercube Sampling (LHS). It scrambled Sobol' sequence sampling, run through independent replicas so the confidence interval on the estimate itself becomes explicit rather than assumed. The results were consistent across three POF regimes: Sobol' hit the same estimator-precision target as SMC with 75–94% fewer simulations, and LHS got there with 50–75% fewer. At the rarest end tested (POF ≈ 2.07×10⁻⁵), SMC needed just over 4.1 million simulations to converge; Sobol' converged at roughly 1.05 million.
That's not just a compute-cost story, though the compute savings are real. It's a transparency story. How confident should an integrity engineer be in a POF estimate used to schedule a reinspection? Only as confident as the confidence interval around it, and most published workflows don't report one. The paper's replication-based protocol makes that interval a required output, not an afterthought, and flags the specific failure mode where confidence intervals with too few observed failures can look artificially tight, sometimes producing intervals that mathematically shouldn't include negative probabilities but do.
Friday, 8:30 am, Conference
This one, co-authored with Lowell McAllister, Mario Chima, Michael Murray, Steven Middleton, and Mohammad Al-Amin, is the demonstration case for a broader set of best-practice guidelines that C-FER developed for PRCI on applying machine learning to pipeline integrity. It puts the leakage problem from Anton and Michael's False Confidence work in front of a real regulatory-relevant use case: predicting stress corrosion cracking (SCC) susceptibility.
The team pulled together anonymized attribute and inspection data from six operators across Canada and the US, roughly 30,500 km of transmission pipeline, and aligned it with public soil data (SSURGO in the US, CanSIS in Canada) to build joint-by-joint SCC susceptibility classifiers. Before any of that modeling started, they ran the same leakage experiment described above, on purpose, as a demonstration. Random sampling by joint produced a training AUROC of 0.96 and a validation AUROC of 0.91, with almost no variation across folds (a standard deviation of 0.002). Switching to grouped sampling by region, diameter, and age, the validation AUROC dropped to 0.78 with a standard deviation of 0.12. That's not noise. That's the leakage getting caught.
Read individually, these are four fairly different sessions, i.e., a large-scale ILI performance study, a validation methodology paper, a simulation efficiency paper, and an applied SCC model. Read together, they're one argument: pipeline integrity is accumulating enough data and enough modeling sophistication that the old habit of trusting a headline number, an AUROC, an F1 score, a POF estimate, a certainty percentage, without asking how it was produced is no longer a safe habit. Every one of these found a gap between the number as reported and the number as it would behave in deployment, and every one of them closed that gap with a specific, repeatable methodology rather than a general warning.
If you're at IPC2026 in Calgary, we'd genuinely like to talk through any of this, especially if you're wrestling with your own version of the leakage, precision, or ILI-performance questions above. Reach out at sales@irthsolutions.com, or find us at Booth 218 in the expo.