By Saad Iqbal
The pump truck operator sees it first: the treating pressure line on the monitor, which had been climbing in a lazy, predictable arc for the last four minutes, suddenly steepens. Nobody says anything for a few seconds — pressure wobbles happen constantly during a frac stage. Then it steepens again. By the time someone calls it out loud, the proppant concentration in that stage is already bridging across the perforations. The pumps get throttled back, the stage is killed early, and the completions engineer spends the next morning explaining a lost stage to a client who is paying by the day.
That sequence — missed signal, late reaction, expensive cleanup — is why AI screenout prediction has become one of the more closely watched applications of machine learning in completions. A screenout isn’t a mystery; the pressure data usually contains the warning several minutes before a human eye catches it. The question engineers actually care about is whether a model can read that warning reliably enough to act on it, and what happens when it’s wrong.
What Actually Happens Inside a Screenout?
A screenout happens when proppant being pumped downhole can no longer pass through the fracture or perforations and bridges, choking the flow path. Surface treating pressure spikes because the pumps are still pushing fluid and sand into a passage that just got narrower. Left unchecked, that pressure spike can damage the wellhead, sand up the tubing, or force an early, incomplete stage.
The standard warning sign engineers have used for years is the slope of surface pressure, normalized against the fluid flow rate. When that ratio starts climbing faster than the pump schedule would predict, it means the fracture is accepting less proppant than it should — the first sign that a bridge is forming. The problem isn’t that this signal doesn’t exist. It’s that it’s noisy, it looks different stage to stage, and a human watching six screens at once will catch the obvious spikes and miss the subtle ramps.
How Is AI Actually Predicting Screenouts?
The most detailed public account of this comes from Noble Energy’s work in the Niobrara-DJ Basin, presented by data scientist Dr. John Sun at the 2020 URTeC conference and later written up by Drilling Contractor. The team didn’t rely on a single black-box model — they built an ensemble that blends a physics-based “inverse slope” model (essentially a smarter, automated version of the pressure-ratio heuristic engineers already use) with a CNN-LSTM neural network trained on time-sliced pressure sequences.
- Surface pressure normalized against fluid flow rate, cut into 1,280-second sequences
- Cumulative flow rate and cumulative proppant volume pumped so far in the stage
- Geological area and target formation, since screenout behavior varies by rock
Those features fed a CNN-LSTM — convolutional layers to pick out local pressure shapes, LSTM layers to track how the signal evolves over the stage — trained and tested against 65 confirmed screenout stages plus normal stages from the same wells.
What Do the Results Actually Show?
This is the part worth reading carefully, because the results aren’t a straight line to “AI wins.” On the validation set, swapping a baseline LSTM for the CNN-LSTM pushed the true positive rate — the share of real screenouts actually caught — from 57% to 74%, and the true negative rate from 89% to 94%. That’s a real improvement.
But on the held-out test set, the CNN-LSTM’s true positive rate actually fell, from 86% (baseline LSTM) to 76%, while its true negative rate dropped from 100% to 90%. That’s the classic sign of a model that memorized patterns in its validation data rather than learning something that generalizes — a result every engineer evaluating a vendor’s model should know to ask about.
The ensemble model — blending the physics-based slope model with the CNN-LSTM rather than trusting either alone — is what actually delivered on the test set: a 95% true positive rate and a 96% true negative rate, beating every individual model on true positives, true negatives, and false positives together. The lesson isn’t “neural networks beat physics.” It’s that combining a transparent, physics-grounded signal with a learned one gave the most reliable result, and that nobody should trust test-set numbers a vendor hasn’t shown separately from validation numbers.
Earlier work on screenout prediction, including a 2022 SPE Journal paper on physics-informed screenout models, reached a similar conclusion from a different angle: pure data-driven models struggle on wells or formations they haven’t seen before, while models that bake in frac mechanics generalize better — which is exactly why the “ensemble” approach keeps showing up as the practical answer, not a single algorithm.
Building a Lightweight Early-Warning Layer Yourself
Not every operator has a data science team and 65 labeled screenout stages sitting around. A smaller, rule-plus-model layer built on top of your existing SCADA feed can still catch most of the obvious cases, and it’s a reasonable first step before committing to a full vendor platform.
In practice that means pulling 1–5 Hz treating data into Python with pandas, computing a rolling slope of pressure-over-rate against a short lookback window, and setting an alert threshold tuned on your own historical stages — a handful of known screenouts is usually enough to calibrate a first-pass rule. For engineers who want to go further without building everything from scratch, the open-source petropt package wraps common petroleum-engineering optimization routines in Python, and a dashboard in Power BI turns the rolling feature into something the whole completions team can watch live, not just the person at the data terminal.
Where This Still Falls Short
None of this removes judgment from the job. The inverse-slope model alone produced false positives on two of three pressure spikes in Noble’s own data — pressure ramps that look like trouble but aren’t always one. And a model trained on one basin’s rock and pump schedules won’t automatically transfer to another; the geological-area and formation features in the Noble model exist precisely because screenout behavior is basin-specific. Any model you adopt — vendor-built or homegrown — needs validation against your own stages before you let it quietly override a pump operator’s judgment.
If you’re building out a broader pressure-analysis workflow, it’s worth pairing this with the fundamentals: our guide to hydraulic fracturing breakdown pressure and net pressure covers the Hubbert-Willis side of the same treating-pressure story, and proppant settling velocity walks through the Stokes’ Law math that governs whether proppant stays suspended in the first place. If frac hits are also on your radar, we’ve covered AI frac hit detection on multi-well pads as well.
A screenout is still, at bottom, a pressure signal moving faster than a person can act on it. AI doesn’t change the physics — it just reads the gauge faster, and more consistently, than six screens and one coffee-deprived engineer ever will.
What’s your take — are physics-informed ensembles the right call, or should completions AI lean harder into pure deep learning as more labeled stage data becomes available? Tell us what’s worked on your pads.
