How AI Is Transforming Petrophysical Analysis in West African Basins

For four decades, petrophysical interpretation in West Africa has been a manual craft. A senior geoscientist sits with a well-log montage — gamma ray, density, neutron porosity, resistivity, sonic — and applies cross-plot heuristics built up over a career. The work is slow, expensive, and non-reproducible: ten experts looking at the same data routinely produce ten subtly different reservoir models. As mature Niger Delta fields enter late-life and frontier deepwater plays face brutal capital scrutiny, that craft economy is breaking down. Operators need interpretations they can audit, version, and reproduce — and they need them faster.

Over the last three years, a new generation of domain-adapted machine learning models has moved from the literature into production interpretation workflows. The models do not replace the petrophysicist. They do something more useful: they compress the routine 70% of the work — lithology assignment, fluid typing, porosity estimation — into a defensible first pass that the human reviews, edits, and signs off on. Hwodye Energy has deployed this pattern across Niger Delta acreage and the Anambra Basin since 2024. The numbers below are drawn from a portfolio of 17 wells where we ran the AI workflow in parallel with a senior petrophysicist's manual interpretation and measured both outputs against subsequent core data.

Why petrophysics is unusually well-suited to deep learning

Most domains that AI has attacked in the last decade share three features: large amounts of labelled data, stable input distributions, and a clear loss function. Petrophysics has all three — but the data lives behind proprietary walls, the input distribution shifts every time you cross a basin boundary, and the loss function is sometimes argued about for years. That is precisely why generic, off-the-shelf models perform badly here, and why domain-trained models work so well.

Modern petrophysical AI typically uses one of three architectures. The first is a convolutional neural network treating the well log as a 1D image — depth on the long axis, log curves as channels. The second is a recurrent or transformer-based sequence model that explicitly models depth as time, well-suited to thin-bed and laminated reservoirs. The third — and the one that has moved fastest into production over the last 18 months — is a graph neural network that treats the wells in a field as nodes in an analogue-well graph, propagating information about lithofacies between geologically similar wells. Each of these has strengths and failure modes; the Hwodye Energy stack uses an ensemble that combines all three with a calibrated weighted vote.

The data problem — and how the industry is solving it

The single biggest obstacle to deploying petrophysical AI is data. West African well logs sit in heterogeneous proprietary formats — DLIS, LAS, WITSML — across decades, vintages, and tool generations. Most operators have no clean library of (log → core → outcome) triples to train on. The first six months of any production deployment is spent on data engineering, not modelling. That work — format ingestion, depth alignment, environmental correction, mud-cake compensation — is unglamorous but it is where the accuracy comes from. The model architecture is, in our experience, a distant second.

Public datasets help but only marginally. The SPE Equinor Volve dataset released in 2018 is a remarkable open contribution — a complete North Sea field with logs, seismic, production, and core. It is also nothing like the Niger Delta. Models trained purely on Volve transfer poorly to deltaic sequences. The most useful open contribution in recent years has come from the FORCE Machine Learning competition series, which produced a clean, labelled facies classification benchmark from the Norwegian Continental Shelf. It is now the standard testbed for new architectures even though, again, the target deployment domain is usually elsewhere.

What the AI workflow actually looks like in production

When a new well is logged, the raw curves arrive at the platform in DLIS or LAS format. The first stage of the pipeline runs deterministic quality control — depth shift, splice, environmental correction — using the same algorithms an interpreter would apply manually. The output is a clean, harmonised log suite. The ML stage then runs three parallel tasks: lithology classification, porosity and saturation estimation, and fluid typing. Each model emits both a prediction and a calibrated uncertainty estimate, which is more valuable than the prediction itself. A petrophysicist can ignore a confident lithology call; they cannot ignore a 60/40 sand-shale call in the pay zone.

The third stage is where the workflow stops looking like a research demo and starts looking like a regulated industrial process. Every prediction is logged, every model version is pinned, and every interpretation that goes to the petroleum engineering team includes a complete provenance trail — input curves, model version, hyperparameters, training data hash, and human override history. This is not optional. Petrophysical interpretations are regulated submissions; the NUPRC Drilling and Production Regulations require reservoir engineering submissions to be reproducible, and AI-assisted workflows have to clear the same bar.

The model architecture is, in our experience, a distant second to the data engineering. The first six months of any production deployment is spent on format ingestion, depth alignment, and environmental correction.

What it changes — and what it doesn't

The economic effect of moving from manual to AI-assisted interpretation is substantial but easily overstated. The single biggest gain is throughput: a workflow that takes a senior petrophysicist three weeks per well now takes 90 minutes of compute plus a four-hour human review. For an operator producing reservoir models across 200 wells, that is the difference between an annual interpretation cycle and a monthly one — which means the reservoir model can actually keep up with production reality.

The accuracy improvement matters more than the speed. Manual lithology interpretation across complex Niger Delta sequences typically achieves 70–80% agreement with core when reviewed against ground truth. Hwodye Energy's ensemble pushes that to roughly 94% on the same benchmark wells, with most of the remaining 6% concentrated in thin laminated intervals where the underlying log resolution is itself the limiting factor. That improvement compounds. A 15-point lift in lithology accuracy translates downstream into measurably better porosity-permeability estimates, better OOIP/OGIP volumetrics, and — most consequentially — better infill drilling target selection.

What it does not change is the role of the petrophysicist. The expert judgement that turns a clean log suite into a reservoir narrative — recognising a sequence boundary, identifying a fluid contact, calling a pay zone in a marginal sand — is not going away. The AI workflow simply removes the repetitive 70% of the work that was preventing experts from spending time on the 30% where their judgement actually matters. That is the change we believe operators across the Niger Delta should be preparing their teams for now.

What the next 18 months will bring

Three developments will shape the next 18 months of petrophysical AI in West Africa. First, foundation models for geoscience — analogous to large language models for text — will start to appear. Microsoft Research's GraphCast is the closest analogue from an adjacent earth-science domain: a single very large model trained on broad data that fine-tunes cheaply for specific tasks. Several major operators have research programs targeting the same idea for subsurface data; expect public results from 2026 onwards.

Second, the regulatory environment will tighten. The Nigerian Petroleum Industry Act and its implementing regulations are creating audit requirements for reservoir interpretations that will force operators to choose between maintaining manual workflows (slow and expensive) or deploying AI workflows with full provenance (faster but requiring infrastructure investment). The operators that move first will have a meaningful capital efficiency advantage in the late 2020s acreage round.

Third, the Energy transition will reshape what counts as a 'good' reservoir. CCUS site screening, geothermal play assessment, and underground hydrogen storage all use the same petrophysical primitives — but optimise for different properties. Caprock integrity matters more than reservoir quality; thermal conductivity matters more than porosity. The flexibility of a learned, multi-task model — fine-tuned to each application — is a much better foundation for that transition than a portfolio of cross-plot heuristics designed for hydrocarbons. That is the case Hwodye Energy is making to West African operators today, and it is the case we expect the industry to converge on by 2027.

If you operate or invest in West African subsurface assets and want to see this in action on your data, Hwodye Energy runs a four-week pilot programme that benchmarks our ensemble against your in-house workflow on a subset of your wells. The pilot is structured so that you keep the model, the training data, and the provenance trail at the end of it — even if you choose not to extend the engagement.