Ensemble-Based Geosteering Under Uncertainty: A Coal-Seam Retrospective Case Study

There is a timely need to address and incorporate geological uncertainty into geosteering. To account for it, we have designed a workflow that utilises a predefined ensemble of geomodels and updates model probabilities as new real-time observations arrive from the steered well, to guide every next steering decision.

The starting point is an uncertainty ensemble of equiprobable sector geomodel realisations along with defined operational constraints. Geosteered drilling is segmented into discrete decision intervals, where admissible inclination and azimuthal changes are represented through a decision tree; each path leading to a full candidate trajectory. Target lithotype footage is evaluated across all realisations for every leaf. The option with the highest expected footage is then selected. Monte Carlo sampling is utilised for large decision trees. After each decision interval, we compare the true lithotype sequence along the drilled path to date with every realisation’s predicted response and estimate realisation probabilities from mismatches via Boltzmann weighting. Once realisations are penalised for failing to predict the observed lithotypes, remaining decisions are reviewed.

The workflow has been applied to a real gas field geosteering case retrospectively, where two existing horizontal wells were used as a hindsight reference. The prior ensemble accounts for the offset vertical wells and incorporates two alternative structural interpretation concepts as baselines: (i) a basement-to-top conformable framework based on well markers only, and (ii) a framework driven by seismic-reflection trends. Baseline structures were varied within agreed uncertainty limits. Ensemble models include four lithotypes (sand, silt, poor coal, and coal) with fractions fixed according to the base case. Geosteering aimed for a coal-seam target.

A hindsight model was constructed by integrating the interpreted geosteering objects, well markers and horizontal well logs. This standalone model serves as the source of the “truth” lithotype log and the post-run evaluation reference; it is not available for decision-making. The outcome is a virtual geosteering replay, and at each step decisions are made automatically to maximise expected coal footage, with the resulting trajectory evaluated against the hindsight model. Cumulative coal footage, reservoir-simulation metrics, and net present value versus the pre-drill trajectory were assessed, showing that the final trajectory outperformed the pre-drill plan and most single-model trajectories.

This case study proves that ensemble-based updates have practical value and can be delivered as steering recommendations and matching hints to assist manual real-time geosteering decisions. In the next steps, we plan to implement ensemble updates between the decision steps.