AI deposit discovery: models that reduce drilling risks by 35%

AI geological exploration: where data is lost and drilling risks rise

On a mature asset, the story almost always lies in mixed form: seismic cubes in old projects, well logs in scattered tables, conclusions from past campaigns in reports and presentations.

SEG-Y and LAS formats, different coordinate systems, sampling intervals and sampling depths were never brought to a unified form — each campaign was done for its own task.

The geophysicist merges this manually, and data heterogeneity imperceptibly turns into uncertainty in reservoir prediction.

Then the cost of error comes into play. The reservoir boundary is drawn from one subjective interpretation, while alternative versions remain in someone else's head and are not tested.

If the formation turns out to be dissected or pinches out, the well comes in wet — the risk of a dry well materializes where investments have already been made.

AI geological exploration closes exactly this gap. The model is trained on all available history of the asset — seismic, well logs, past drilling results — and outputs not a single line, but a probability map with uncertainty assessment for each point. Geophysical interpretation ceases to depend on who is looking at the section today.

Drilling targets are ranked by expected productivity and by the stability of reservoir prediction under variations in input data. Below we break down how this is structured and what makes it possible to reduce the risk of a dry well.

What a predictive model provides: accuracy, timelines, failure class

A predictive model does not replace the interpreter, but provides a second independent view of the data. An ensemble of algorithms is trained on seismic cubes, well logs and actual results of already drilled wells, and validated on blind wells that it did not see during training.

For each point, the probability of a productive reservoir and an uncertainty map are output: where the algorithms disagree, the decision goes to additional analysis rather than into the drilling plan.

Preparing a geological model ceases to be a separate multi-day stage. Recalculation on a new cube takes hours, so the model is updated after each well, rather than once a season — the discrepancy between forecast and fact is visible immediately, before it accumulates.

We deliver the result in project formats Petrel, and the geologist remains in the familiar interpretation environment.

Reducing dry wells starts with ranking targets: the model orders targets by probability, and points with maximum agreement between seismic and well logs enter the plan first.

Where data is sufficient for calibration, the share of confirmed drilling targets increases by 15–25 points — this is the difference between an exploration hypothesis and a work plan.

A separate class of avoided failures is interpretation errors during manual merging of data of different resolution: missed pinch-out zones, shifted reservoir boundaries, ambiguous faults.

The model does not create new information, but checks each geologist's version by the same procedure — repeatability of checking eliminates errors that a person makes under tight deadlines.

Approaches in AI geological exploration: pilot, project, support

Before launching the model in the field, it is worth comparing it with what already exists: well log and seismic archives, drilling history, current interpretation regulations.

There are three work formats, and they differ not in volume, but in how deeply the model is embedded in asset decisions.

The formats are not mutually exclusive: a pilot on the archive usually transitions into a model for the asset, and that into support.

The choice depends on whether the data has been brought to a unified form and whether there is a person on the team responsible for updating the model.

Work stages: from data audit to model handover

Work proceeds in iterations, not in one long pass: first we check whether the data even allows setting the task, and only then train the model.

This order captures where an error is easy to miss — in depth matching, in data leakage between neighboring wells — and where the forecast is checked independently.

1. Geodata audit. We inventory seismic, well logs, drilling results and historical reports, bring them to unified coordinates and depth markers, and mark gaps and heterogeneous measurements.
2. Model task definition. We formulate exactly what we predict — reservoir type, probability of a productive interval or formation properties — and the metric by which the geologist accepts the forecast as suitable for deciding on a well location.
3. Training and hypothesis selection. We compare several models on a split by areas, not by individual points: otherwise the model memorizes neighboring wells and shows a good result where there is none.
4. Forecast validation. We check on a holdout sample and on areas with already drilled wells, counting missed objects and false positives separately — this shows the consequences of each error.
5. Pilot area. We run the model over a real area, compare with geological maps, refine the feature set and the trigger threshold.
6. Implementation and model handover. We embed regular recalculation into the workflow, hand over documentation with limitations, and train the customer's team — so that the solution works without us.

Deliverable composition: models, documentation, regulations

The result of the work is not a "black box" with forecasts, but a package that the customer can support and develop without us. Below is what we hand over and in what form, so that the model remains manageable several seasons after launch.

What the customer retains

• Data schema: list of sources (seismic, well logs, geochemistry, historical conclusions), rules for bringing to a unified form, update frequency and the person responsible for loading.
• Feature schema with rationale: which parameters affect the forecast, which are excluded as leakage and why — so that adding new data does not break the logic.
• Trained models in the workflow with a fixed version and the ability to roll back to the previous one without losing forecast history.
• Metrics report: share of correct hits and completeness for prospective areas, separately — type I and type II errors, since missing a reservoir and an extra well have different consequences.
• Model documentation: what exactly it predicts, on what data volume it was trained, applicability boundaries and known biases — by regions and sediment types.
• Retraining regulation: triggers (new wells, metric degradation, change in interpretation methodology), procedure for checking on a holdout sample and the condition for allowing an update.
• Quality control regulation: comparison of forecasts with expert assessment by geologists, version log, stopping criteria and procedure for analyzing discrepancies.
• Instruction for geologists: how to read the prospectivity map and probability assessment so that different specialists interpret one result in the same way.

We hand over each item in a format suitable for internal audit: versions in the log, metrics reproducible, decisions explainable. The team supports the package on its own, and we join at the stage of model revision or expansion to a new area. Describe the task — we will return with an assessment and plan.

Case: reducing drilling risks at a field with complex geology

An area with complex geology — fault tectonics and a thinly laminated reservoir, where neighboring wells penetrate different intervals.

The seismic data for it had been reinterpreted more than once, but the team prepared the decision on the drilling target in 6-8 weeks: merging well logs, seismic and data from past wells into one representation manually could not be done faster. Of 12 drilled targets, 6 were confirmed.

In this geological exploration case, we assembled the archive — seismic cubes, well logs for wells, drilling history with marks on confirmed and unconfirmed targets.

The data was brought to a unified grid, the result of each past well was labeled, and a model was trained on this labeling.

At the output, the geologist receives a probability map and an uncertainty map: it is visible not only "here is prospective", but also where data is objectively insufficient and additional surveying is needed. The results were compared with interpretation in Petrel.

Metrics before and after

Reducing drilling risks here is achieved not by "predicting the reservoir", but by screening out zones where seismic carries no information about the reservoir.

The share of confirmed targets grew because the model explicitly marks disputed areas — decisions on them are made after additional surveying, not by inertia.

How does AI forecasting differ from classical data interpretation?

Classical interpretation is built on a geological model that a specialist assembles manually: seismic, well logs, core, data from neighboring wells.

It explains why the trap formed, but in areas with a sparse observation grid the conclusion runs into personal experience — on the same material, two geologists often give different assessments of prospectivity.

The model works with the same material differently. Training on historical data from already drilled wells makes it possible to test thousands of feature combinations and find dependencies where manual enumeration is unrealistic.

The output is not a binary "drill or not", but a probability map with uncertainty assessment and a ranked list of areas.

Forecast validation is performed on holdout wells that the model did not see during training, and against actual drilling results.

We look at the share of confirmed targets and the stability of the conclusion when the training sample changes: if the assessment "drifts", the area goes not into the work plan, but into the list for further study. Zones where the discrepancy between model and expert is maximal are analyzed manually first.

Therefore, the expert is not removed from the process. The model covers volume and repeatability — it passes through the entire array in the same way and does not miss weak anomalies; the geologist is responsible for geological meaningfulness and the decision on well location.

Abandoning the expert is not planned: discrepancy in conclusions is a signal to recheck the data, not a reason to trust one side.

What is clarified before start: questions on data and models

Before start, we fix data requirements: well logs for wells, seismic cubes, test results and coordinate referencing — in a unified coordinate system and with an edit history.

Separately, we estimate sample size: if confirmed wells with a full set of measurements are fewer than needed for stable training, some conclusions are built on formation physics rather than statistics.

Contractor exports are almost always in different formats: some have well logs in LAS with arbitrary curve names, others have seismic with a different grid step. We bring everything to a unified schema and feature dictionary — otherwise the model trains on incomparable numbers. Below are the doubts that most often surface at this stage.

Each contractor has its own export format — is that a problem?

No, if this is addressed at the input. We write converters for each source and a curve mapping dictionary, and fix rejection rules. At the output — one dataset from which the training sample is assembled reproducibly.

Will a model from one area work on a neighboring one, and who is responsible for the forecast?

As a starting point — yes, as a ready answer — no: we check on holdout wells of the neighboring area and see where accuracy drops. The model gives a probability; the geologist makes the drilling decision.

We are responsible for reproducibility: versions of data, code and model, metrics on the holdout sample, change log.

Let's discuss the task: composition of the first stage and what we need from you

The first stage does not require the full archive: to return with an assessment and plan, it is enough for us to understand the geology of the area, the volume of already collected data and what drilling decisions need to be made. Below is the minimum from which the review begins.

• Area boundaries and target interval: coordinates in your system, depths, stratigraphy.
• Seismic in source form — cubes and profiles indicating polarity, step and completed processing, or an export from your database.
• Well logs and well data: curves, deviation surveys, formation tops, test results.
• Drilling history: where drilling was performed, what was obtained, which intervals turned out empty or had accidents — this is what the model learns from.
• The task in your wording: ranking of prospective zones, reservoir forecast or risk assessment for a specific location.
• Constraints: data export rules, storage and access requirements, who on the customer side confirms the result.

Part of the list is usually missing — then we record the gaps and decide whether to close them by collecting more data or by narrowing the scope.

Next we look at volume and quality, check which forecast horizon is realistic, and return with a first-stage plan: scope of work, area of responsibility of each party, and by which indicators we measure the result.

If there is too little data for training, we will say so before start, not in the middle of the project. Describe the task in a convenient form — we will agree on the starting point of work.