Machine learning for upstream oil and gas
Machine learning for upstream oil and gas
See which wells need attention, improve production forecasts and assess operating options with evidence your engineers can review. NeoBram scopes upstream machine learning around one field decision, your available data and your petroleum specialists.
Build forecasts that help production planning
A production forecast should help someone plan capacity, investigate a shortfall or update an operating outlook. We assess well and field histories, test frequencies, downtime and operating changes before selecting a method. Start with the forecast horizon and the decision it supports, then compare machine learning with the engineering or statistical baseline your team already uses.
Tests use later time periods and, where relevant, wells excluded from training. Outputs should show uncertainty and flag conditions outside the tested range. A forecast learned from operating history does not replace a reservoir study.
Make an illustrative daily surveillance list useful
A production engineer begins by screening wells for unexpected changes. A bounded workflow can compare recent measurements with the relevant operating context and prepare a review list. A planned shut-in, delayed measurement or allocation change should be visible before the engineer treats a difference as a production problem.
The engineer receives the well identity, period, observed change, supporting records and missing context. They can record whether the item needs investigation, is explained by a known event or should remain under observation. That feedback helps test whether the list improves the work rather than simply generating more alerts.
- Prepare the evidence: align well identifiers, units, timestamps and operating events.
- Prioritize for review: separate unavailable inputs and known events from unexplained changes.
- Review the decision: petroleum specialists judge the significance and next action.
- Measure usefulness: count important missed changes, distractions and total screening effort.
Focus well surveillance on changes worth investigating
Pressure, temperature, flow and operating-state records can help surface unexplained production changes. A useful review list distinguishes missing measurements, planned shut-ins and changed operating conditions from patterns that need investigation. Each item should identify the well, the time period, the evidence and the appropriate reviewer.
The practical benefit to test is less time screening routine data and faster review of important deviations. Count missed events and distracting alerts alongside any time released.
Compare artificial-lift performance and operating options
For a suitable project, lift-performance analysis can examine how an ESP, gas-lift or rod-pump system behaves under different conditions. Your specialists define feasible changes, well constraints and the evidence needed to judge a recommendation. The initial scope can rank candidates for engineering review or compare operating scenarios within a validated range.
Production optimization and equipment failure prediction require different tests. This page covers well performance and advisory operating decisions; pump-health warnings and maintenance planning have a separate scope.
Start with the records your team already trusts
We review production allocations, well tests, historian or SCADA extracts, downtime records and operating events. Consistent well identifiers, units, timestamps and measurement quality matter before model complexity. Virtual flow estimates are an option to assess where reference measurements support validation; they must remain distinguishable from measured rates.
Approved extracts may be enough for an initial study. A maintained application needs agreed interfaces, refresh frequency and a clear response to stale data. On-premises, edge or offline operation depends on local inputs, hardware and software dependencies.
Define reservoir and drilling work separately
Reservoir characterization, history matching and drilling optimization need specialist models, domain expertise and their own validation. We discuss the assessment requirements and whether the right expertise and evidence are available before proposing that scope. We confirm the specialist expertise, suitable data and validation approach before accepting that work.
Ask for outputs that support the field decision
A first project can provide a data and event assessment, a comparison with the current engineering or statistical method, uncertainty estimates and a prototype review output. Agree the horizon, wells and operating regimes included. Explain where measurements are sparse or estimated so the team understands the strength of the result.
For forecasting, assess later periods and bias as well as overall error. For surveillance, assess which review decisions improved. For an advisory operating option, document physical constraints and the evidence needed before an approved intervention. Keep forecast quality, time released and additional production separate in the business case; changing field conditions can affect each differently.
More clarity
Questions and answers
Can this work with our existing SCADA or historian?
Potentially. We first assess approved exports or interfaces, tag quality and update frequency. A first study can use extracts without changing the live control system.
How much production history do we need?
There is no universal minimum. The records must cover the operating conditions, forecast horizon and events being tested. Gaps and changes in measurement practice affect what is feasible.
Will AI automatically optimize well settings?
The proposed starting point is engineer-reviewed advice. Automatic control requires a separately engineered and approved scope, with suitable protection and fallback arrangements.
Can one model work across all our wells?
That needs testing. Well behaviour, lift systems and operating regimes differ. We compare shared and well-specific approaches rather than assuming performance transfers.
What does a first project deliver?
Agree a focused data assessment, baseline comparison, tested model or prototype, and findings on operational usefulness. Integration, support and rollout are defined separately from the initial evaluation.
Can a first study use approved data exports?
Yes, if the exports contain the timing, identifiers and context required by the decision. A recurring application needs separately agreed refresh, interface and failed-input arrangements.
How can we compare the result with our engineers’ current method?
Use the same decision horizon and information available at that time, then test later observations. Review bias, important operating regimes and usefulness with the engineers, rather than comparing unlike datasets or only model scores.
Start with one business problem