Decoding the Subsurface: Well-Log Intelligence and Formation Anomaly Detection

Your production logs catch a spike in wellhead pressure over a few hours. The pressure normalizes. Your team calls it operational noise. Weeks later, the well's effective permeability has dropped, and recovery is in freefall.

That spike was the first whisper of a completion-integrity event — a subsurface signature that manual interpretation catches only in hindsight. By the time permeability shifts are visible, formation degradation is already underway, and the window to intervene cleanly is closing. The WellLogStack instrument reads well logs alongside real-time production telemetry to surface the anomalies that manual interpretation misses — formation damage, completion loss and permeability shifts — with enough lead time for operations teams to intervene before a recoverable field becomes a recovery problem.

⤢ Expand

The Formation Integrity Problem: What Your Data Sees Before Your Wells Feel It

A producing well is a conversation between rock, fluid and completion. The formation holds the resource; the well delivers it. The moment that conversation breaks — a fracture closes, a perforation erodes, pressure seals fail — recovery decays in ways that show up in telemetry hours before they show up in reserve estimates or economic models.

The data is rich but fragmented. Every minute, SCADA historians record flowrate, wellhead pressure, watercut, produced gas-oil ratio, and downhole pressure if sensors exist. Legacy well logs from the drilling campaign sit in subsurface databases — gamma ray, resistivity, sonic, density — months or years old but laden with formation context. The gap: logs are a static picture; telemetry is the live conversation. Tying them together has been the job of geologists with spreadsheets and intuition. Today it is the job of models that read patterns across many wells and many days of production.

Real formation-degradation events have signatures. A loss of confining pressure shows as slow watercut rise weeks before permeability tanks. Completion-zone incursion from an adjacent fault appears as a step change in GOR and pressure signatures that do not match the flowrate. These are not threshold-alarm phenomena; they are contextual — they depend on the well's operating history, reservoir thermodynamics, and what "normal" means for this particular subsurface, not fleet averages.

But the signal is buried in noise. A normal shift sees dozens of micro-transients — flowrate upticks, pressure dips from compressor load changes, separator cycles. None are formation events. Yet to simple alarm rules, they all look identical. After a few false alarms a day, crews stop trusting alerts. After missed real events, wells have already degraded beyond intervention.

The Well-Log Stack: Fusing Subsurface Context with Real-Time Production Flow

The WellLogStack premise is straightforward: formation behavior is encoded in the logs. Gamma ray reveals shale fraction. Resistivity shows water saturation and aquifer boundaries. Sonic indicates mechanical properties. Married to production telemetry, this geological context separates noise from real subsurface events.

For each well, you ingest:

1. The static subsurface model — well logs from drilling (gamma, resistivity, sonic, density), formation top and bottom picks, perforation clusters and open-hole intervals. This is the rock conversation.
2. The production telemetry stream — daily or sub-daily flowrate, pressure, watercut, GOR, temperature, downhole gauges, separator flags. This is the live flow conversation.
3. The operating context — choke settings, compressor speed, recent workovers or interventions that reset the baseline. This is the boundary condition.

The model operates in three parallel tracks:

First, it learns what "normal" looks like for this specific well's geology and operating regime. A tight carbonate with active aquifer drive drifts differently than a layered sandstone already depleting, which drifts differently from a well on mechanical lift held at constant compression. The model builds baselines per well, per formation, per operating mode — it does not assume all wells drift identically.

Second, it detects anomalies in real time. A hybrid ensemble of statistical process control (SPC) and deep-learning autoencoders reads the multivariate pressure-flow-composition space. An autoencoder trained on many normal production days learns the manifold of "healthy operation." When new telemetry falls off that manifold — pressure-watercut relationships break pattern, or GOR-flowrate ratios become geometrically implausible — the autoencoder flags it. The SPC layer catches point anomalies (single pressure spikes) and contextual ones (slow drifts in base watercut with constant rate). Together they catch point, contextual and collective anomalies that either one alone would miss.

Third, it roots the anomaly back to the logs. When a flag fires, the system surfaces formation context: the sand cluster most likely affected, the aquifer or fault boundary most likely breaching, the pressure regime where the event sits. This lets well operators say: "This is real formation damage in the affected sand, not a separator malfunction in the processing train."

Result: 95% detection at under 2% false positives. High enough that crews trust alerts enough to act on them — low enough that they do not abandon trust after a few days of noise.

Process flow · hover a step to trace itExpand

From Detection to Intervention: Why Crews Act on Formation Anomalies

The early-warning window is the entire value proposition. A formation-damage event surfacing as slow permeability drop over a month can be caught early or late. The difference between those two interventions is the difference between a well you save and a well you abandon to accelerated decline.

What makes crews act on alerts is the ability to tell real formation events from processing-train artifacts. A separator upset looks like a watercut spike but flattens within hours; a formation-breach signature is persistent. A compressor surge moves entire pressure baseline; formation damage moves the pressure-watercut relationship. Because the model reads multivariate space, it can distinguish them. Because crews see what to look for, they can verify it in real time.

95%

Detection accuracy

<2%

False positive rate

40%

Less unplanned downtime

4–6 weeks

Assessment phase

Building the Model Across a Fleet: Handling Well-Type Heterogeneity

The most fragile moment in deploying anomaly models at scale is moving from one well to many. Every well has different geology, completion design, operating regime. Some have downhole gauges; some only surface readings. Some are gas-lifted; some naturally flowing. Some are single-layer sandstone; some multi-stack completions with cross-layer communication. The model must work across all without custom baselines per well.

This is where the feature pipeline is the model. Instead of training one autoencoder on raw pressure and flowrate, WellLogStack computes:

• Scaled pressure changes — pressure-change rate normalized by flowrate and the well's estimated formation compliance (from sonic logs). A given pressure rise means something different in stiff limestone versus soft sandstone.
• Composition ratios — watercut and GOR normalized by the well's expected drive mechanism (solution gas, aquifer, pressure maintenance). High watercut is "normal" late in waterflooded fields but "alarming" early in pressure-depletion wells.
• Temporal derivatives at multiple scales — daily, weekly and monthly rolling changes. Rapid watercut spike differs from slow creep; the model sees both.
• Formation-cluster specific terms — for wells with multiple paid intervals, pressure and watercut changes parse per cluster (from well logs). Watercut rise in upper sand is independent data from lower sand changes.

The result is a feature space where "normal" is defined by the well's own geology and operating regime, not fleet averages. One autoencoder, trained on a representative well-type and operating-condition mix, generalizes across new wells because features already encode well-specific context. You deploy to a new well, give it a baseline telemetry window, and it works against that well's own normal — no per-well retraining needed.

The difference between a formation event you see early in the window and one you see late is the difference between a well you save and a well you accept will decline.

Sustain: Keeping Detection Sharp as the Formation Evolves

Every producing well is a moving target. As production continues, pressure declines, water encroaches, reserves deplete. The "normal" telemetry profile learned at deployment differs from later profiles. A model that worked perfectly initially can become blind to real events because its baseline has drifted.

The sustain phase watches for two kinds of drift. Operational drift — well enters different mode, compression increases, choke adjusts — is recalibrated fast. Geological drift — pressure declines, water encroaches — is retrained slower, against the new formation regime. The model must catch a change in decline rate (completion failure), not flag normal depletion as anomaly. The retraining cadence is tuned to keep alert trust high: crews act on alerts when they believe they are real, and stop when false alarms signal a model that does not understand the well's phase transition.

Where to start

Begin with a telemetry and well-log audit across your producing fleet. Map which wells have the richest logs (gamma, resistivity, sonic where available), which have the most complete telemetry (downhole pressure, not just surface readings), and which have shown historical variability in permeability or production stability. Start with the subset that has both good logs and good telemetry, plus the highest economic interest — the wells where timely intervention would have the biggest reserve or NPV upside.

The assessment phase, typically 4–6 weeks, inventories your subsurface and streaming data sources, profiles data quality and completeness, and ranks opportunities by recoverable value and feasibility. The output is a prioritized roadmap: a pilot cohort of wells where the WellLogStack instrument can deploy with confidence, ranked by expected economic impact of early detection.

Transform builds the model on that cohort, validates it against historical events (the anomalies you have already seen, in retrospect), and integrates it with your SCADA and alert workflow so flags land in the operations-center console in real time. Sustain monitors model performance against live telemetry, retrains against pressure and water trends, and tunes alert thresholds to your crew's operational rhythm and tolerance for false alarms.

The result is a layer of formation awareness that reads subsurface stress before it cascades into reserve loss — and gives your teams the lead time to act.