In both of these
case studies, the goal is to recover oil or oil accumulations that have been
missed by past explorers. Well-trained AI models can do that very well with
better pattern recognition than humans, according to the case studies. However,
in both cases, verification has not yet occurred, but bypassed oil and bypassed
prospective oil accumulations have been identified. Thus, they are basically
leads at this point.
If successful, these cases will verify
that AI’s pattern recognition powers can add oil & gas reserves, in some
cases, better than geologists.
Case Study #1 – Four Bypassed Prospects Identified in
ExxonMobil’s Stabroek Block Offshore Guyana
In May 2026, ExxonMobil Vice
President of Exploration John Ardill noted that the company was expanding its
use of deep learning, machine learning, and high-performance computing to
analyze seismic data and identify hydrocarbon-bearing prospects that were
previously more difficult to evaluate.
According to The Daily
Synapse:
“ExxonMobil has used AI to identify four new exploration
opportunities in Guyana’s Stabroek Block, applying historical discovery data,
drilling results and subsurface information to sharpen prospect evaluation. The
work combines advanced analytics, machine learning, high-performance computing
and next-generation seismic imaging to accelerate exploration, lower costs and
improve discovery success rates.”
Reprocessing and
reinterpreting seismic data are part of the AI-identified prospects. CFO Neil
Hansen and CEO Darren Woods recently commented on identifying the prospects.
According to Rio Times:
“Hansen said the work produced four new discovery
opportunities above and beyond what the company had previously considered
prospects. Chief executive Darren Woods framed the findings as confirmation
that the company is, in his words, not done yet in Guyana.”
Rio Times summarized the
details of the announcement:
Below, Deepwater Exploration
Geoscientist Ryan Christiansen talks more about the AI model in a LinkedIn
post, noting its 90% hindcast accuracy. He points out that the block is very
well-studied, which makes the newly identified prospects intriguing. He also
notes that, due to it being well-studied, AI-model training was enhanced by all
that confirmed data and reserves. He also cautions that the technique may not
work as well where there have been fewer wells drilled (less verified data) and
less subsurface data available.
Case Study #2 – AI-Assisted Review of Old Water Flooded
Fields Identified Potential Unswept Areas – An Eastern Oklahoma AI Screening
Example by Susan Nash
Geoscientist and former
president of the American Association of Petroleum Geologists (AAPG), Susan
Nash, wrote an article on LinkedIn that gives a plausible and practical method
for using AI to assist in analyzing a mature oil field for remaining unswept
oil after a water flood. In this case, the area is Eastern Oklahoma. The field
was originally drilled before 1920 and was later waterflooded in the 1950s and
60s. In this case, there is suspicion that the marginal areas of the oil
reservoir may not have been swept by the water.
First, she notes that early
drilling was not geologically determined, and the only stimulation method available was shooting with nitroglycerine, which is not thought to create a fracture network that extends very far beyond the wellbore. She notes that even though
the field is densely drilled, it may not have been effectively drained. Early
reporting may have misidentified zones and reservoir sands in an area with
suspected reservoir compartmentalization, which is common with fluvial-deltaic
reservoirs. She notes that shooting wells with nitroglycerine could also create
confined induced fracture conduits that miss reservoir compartments or
different sand benches. It also means well-spacing likely won’t be as good in
predicting what may have been bypassed. She gives the following quote
describing how to see the field, not attributed:
“The field was drilled as a collection of wellbores,
flooded as a collection of tracts, and must now be understood as a collection
of flow units.”
Below, she gives four
potential scenarios of oil left behind in terms of apparent opportunity, where
the oil may sit, evidence required, and common false positives.
Below, she explains how her
study was organized and the results of the desktop screening method:
“The source stack combined a 1920s county production map
and geological narrative, state waterflood-unit records, available well and
completion information, modern field boundaries, and current surface-exclusion
layers. AI-assisted methods were used to read and organize long historical
documents, normalize formation and field aliases, parse quarter-quarter legal
descriptions, compare early productive limits with later unit geometry, and
maintain a structured evidence ledger. Human geological review remained essential
wherever a local sand name could represent more than one bench or where a unit
boundary might have followed a genuine reservoir limit.”
Below are the results of the screening process, which ranked three tiers of potentially bypassed or unswept oil.
She notes that AI did not
generate prospects but was able to quickly sift through all the historical data
and make determinations about dates, legal descriptions, reconciling well
duplicates and formation names, and flagging contradictions. These are
functions that remind me of what a colleague called “massaging the data,”
though in a different context.
“AI is also vital in testing data quality, identifying
errors, classifying them accurately, and establishing harmonization.
Furthermore, AI can evaluate waterflood design, implementation, and purpose,
comparing them against available options.”
“Crucially, AI made it easier to test competing
explanations.”
Next, she goes deeper and
considers how agentic AI could be applied to the project, generate
recommendations, and how human interactions fit into the process.
“In an agentic system, a language model manages a
multi-step workflow, selects tools, evaluates results, and routes work to
specialized agents while operating within defined guardrails and human approval
points.”
The table below gives eight
types of AI agents, their functions, and outputs that can be applied to the
project. That work is then examined by geologists, reservoir engineers, and, in
some cases, land, HSE, and regulatory personnel, and tweaked as they see fit.
Finally, as shown below, she gives an outline of how an AI-assisted pilot project can be organized. She calls this a disproof sequence.
References:
AI
Uncovers Four New Prospects in Guyana. Ryan Christianson. LinkedIn Post. August
2026.
The
Unswept Margin: How an AI-Assisted Review of Pre-1920 Fields Identified
Prospective Unswept Areas. Susan Nash. LinkedIn, Article. August 9, 2026. (23)
The Unswept Margin: How an AI-Assisted Review of Pre-1920 Fields Identified
Prospective Unswept Areas | LinkedIn
ExxonMobil
applies AI to identify Guyana prospects. Daily Synapse. August 9, 2026. ExxonMobil
applies AI to identify Guyana prospects – DailySynapse
ExxonMobil’s
AI Breakthrough Signals New Era of Digital Exploration in Guyana, Caribbean Energy
Week. August 7, 2026. ExxonMobil’s
AI Breakthrough Signals New Era of Digital Exploration in Guyana
ExxonMobil
AI Guyana Exploration Finds 4 New Oil Prospects. Lachlan Williams. The Rio Times.
August 3, 2026. ExxonMobil
AI Flags 4 New Guyana Oil Prospects | The Rio Times
No comments:
Post a Comment