By Saad Iqbal
Three hundred feet into the lateral, a horizontal well is boring sideways through a reservoir that is maybe forty feet thick, top to bottom. The directional driller is watching a gamma ray curve crawl across a screen, trying to decide, in real time, whether the bit just clipped the top of the zone or is still sitting comfortably in the sweet spot. Get that call wrong by a few feet and the well spends the next thousand feet of lateral drilling through shale instead of pay. That decision, repeated every thirty seconds for the length of the well, is what geosteering actually is — and for twenty years it has lived or died on how fast a human geologist could pattern-match a live curve against an offset type log, by eye, under time pressure.
AI geosteering does not replace that geologist. It gives them a correlation engine that never blinks, never gets tired on hour fourteen of a night shift, and can re-test a bed-boundary call in the time it takes to glance at a second monitor. This is how that correlation actually works, the true stratigraphic thickness math underneath it, and why it is becoming standard practice on unconventional laterals rather than a novelty.
What Is Geosteering, and Why Does It Need Real-Time Correlation?
Geosteering is the practice of adjusting a wellbore’s trajectory while drilling, using logging-while-drilling (LWD) measurements to keep the well inside a target stratigraphic interval. Unlike a vertical well, where you simply drill to a depth and the geology sorts itself out, a horizontal lateral has to stay inside a layer that might be thinner than the length of the drill collar carrying the sensor. The well can wander out of zone from normal structural dip alone — a formation dipping at just two degrees will drift a horizontal well completely out of a 30-foot-thick zone over about 860 feet of lateral if nobody corrects for it.
The correction loop is always the same three steps: measure a log curve at the bit (gamma ray, resistivity, or both), correlate that live curve against an offset well’s type log to work out where in the stratigraphic column the bit actually is, then recommend an inclination change to the directional driller. AI geosteering speeds up the middle step — the correlation — using pattern-matching and dynamic time warping algorithms instead of a geologist eyeballing two curves side by side and reasoning by analogy.
How Does Automated Type Log Correlation Actually Work?
The offset type log is a vertical pilot hole or an analog well that has already penetrated the target formation, usually with a full gamma ray or resistivity suite logged through it. That log becomes the reference pattern: a sequence of peaks and troughs tied to known marker beds, like a barcode for that stratigraphic interval.
As the bit drills, the live LWD curve streams back in near real time, usually with some lag introduced by mud-pulse telemetry. The correlation engine does not just compare the two curves point-by-point on a shared depth axis — it cannot, because the lateral well is seeing the formation at a different apparent thickness than the vertical pilot, stretched or compressed depending on the local dip and the wellbore’s own inclination. Instead it uses a dynamic time warping approach: it searches for the best non-linear stretch of the live curve that lines its character up against the type log, shifting and scaling small windows of the curve until the correlation coefficient between the two is maximized.
Where AI changes this is in how that search is framed. Instead of a geologist manually sliding the type log left and right against a handful of candidate correlation points, a trained model — often a variant of dynamic time warping combined with a classifier trained on thousands of historical correlation picks — proposes a ranked set of correlation hypotheses and a confidence score for each. Platforms like Rogii StarSteer and SLB’s AlphaSight both build on this idea, running structural and stratigraphic interpretation continuously in the background so the directional driller sees an updated best-fit model within seconds of new data arriving, rather than waiting for the next scheduled correlation session.
The Math Underneath: True Stratigraphic Thickness
Every correlation call eventually reduces to one question: how far, vertically through the rock, is the bit from the top or base of the zone? That distance is the true stratigraphic thickness (TST), and it is not the same as the measured depth interval the bit has drilled, because the wellbore is travelling through the formation at an angle.
The working formula geosteering software uses to convert a measured bed thickness along the wellbore into TST is:
TST = MD_interval × cos(θ)
where:
TST = true stratigraphic thickness (ft)
MD_interval = measured depth interval drilled through the bed, along the wellbore (ft)
θ = relative dip angle, i.e. the angle between the wellbore axis and the bedding plane (degrees)Worked example: a directional driller logs 42 feet of measured depth across a marker bed, and the correlation model has just resolved the relative dip angle at that point in the lateral to 71 degrees (a wellbore running nearly parallel to the bedding plane, which is typical deep into a horizontal section). The true stratigraphic thickness is:
TST = 42 ft × cos(71°)
TST = 42 ft × 0.3256
TST = 13.7 ftThat 13.7 feet, not the 42 feet of hole drilled, is what gets compared against the known zone thickness from the type log. This is exactly the kind of arithmetic an AI geosteering model is re-running continuously and silently as new data streams in, because the relative dip angle itself is a moving target — it changes every time the wellbore trajectory or the formation structure changes, and a stale dip estimate will quietly corrupt every TST calculation downstream of it until the next correlation resets it.
Where Does the Automation Loop Actually Close?
It is worth being precise about what is and is not automated today. On the large majority of geosteered wells, the AI model is advisory: it proposes a correlation, a dip estimate, and a recommended inclination or azimuth adjustment, and a human geologist or geosteering specialist reviews and approves that recommendation before it goes to the directional driller. The loop looks like this:
- LWD gamma ray or resistivity data streams uphole via mud-pulse or wired-pipe telemetry.
- The correlation model updates its structural interpretation and TST estimate against the offset type log.
- The model flags whether current trajectory keeps the well in zone for the next planned interval, and by how much margin.
- A geosteering specialist reviews the flagged correlation, confirms or overrides it, and issues a trajectory recommendation.
- The directional driller adjusts inclination or azimuth at the next connection.
Fully closed-loop automated geosteering — where the model’s recommendation drives the rotary steerable system directly, without a human in the loop — exists and is in active field trials, including work from Helmerich & Payne on autonomous directional drilling, but it remains the exception rather than the rule as of today. The bottleneck is not computational; it is confidence calibration. An automated model that is 95% accurate on correlation picks still means a wrong call every twenty decisions, and a wrong call compounded over a 10,000-foot lateral without a human catching it can cost a well its entire economic case.
How Does This Connect to the Rest of the Drilling Optimization Stack?
Geosteering does not run in isolation. The same real-time data backbone that feeds a correlation model also feeds the surface drilling-dysfunction detection covered in our guide to AI drilling optimization software, and the formation pressure and stability estimates that geomechanics platforms produce — the subject of our best geomechanics software for wellbore stability comparison — increasingly feed directly into the geosteering model’s dip and stress assumptions. On managed pressure drilling jobs specifically, the trajectory decisions geosteering makes interact directly with the pressure window the MPD automation systems are holding, since a steeper correction can change the effective circulating density calculations the MPD choke is managing against.
For engineers who want to prototype their own correlation logic rather than rely purely on vendor black boxes, a lightweight dynamic time warping implementation in Python, built with pandas for the curve handling and run interactively in Jupyter, is a genuinely useful way to build intuition for why certain correlation windows fail — particularly across faulted intervals, where the type log assumption of continuous, unbroken bedding simply breaks down and even a well-tuned AI model needs a human override.
What Should an Engineer Actually Watch For?
Three practical checks separate a trustworthy AI geosteering deployment from a vendor demo:
- Confidence scoring, not just a correlation line. Any model worth using should expose a numeric confidence score per correlation pick, not just the best-fit curve. A flat, low-confidence stretch is exactly where you want a human reviewing the call before the directional driller acts on it.
- Dip trend consistency. Watch the relative dip angle the model is reporting over the last several hundred feet. A sudden, unexplained jump usually means the correlation has locked onto the wrong marker bed, not that the structure genuinely changed that fast.
- Telemetry lag awareness. Mud-pulse telemetry can lag real depth by a hundred feet or more at typical rates of penetration. A model that does not explicitly account for that lag when timing its trajectory recommendations will systematically steer late.
None of this makes the directional driller’s job obsolete. It makes the correlation call fast enough, and transparent enough, that the humans in the loop can spend their attention on the decisions that actually need judgment — the faulted intervals, the ambiguous marker beds, the moments where the type log itself might be wrong — instead of burning it on routine pattern matching at two in the morning.
If your team is evaluating a geosteering platform, start by asking the vendor to show you a correlation call it got wrong, and what the confidence score looked like at that moment. That answer tells you more about the tool than any demo of a clean, textbook correlation ever will.
