Table of Contents
Virtual flow metering (VFM) estimates the oil, gas and water rates of each well from measurements that are already available, such as pressures, temperatures and choke opening, using a physical model, a data-driven model or both. It runs continuously between well tests and is calibrated against test separator or multiphase meter results.
Key takeaways
- A VFM replaces extra metering hardware with a numerical model fed by readily available field measurements such as pressure and temperature.
- Reference measurements are sparse: well tests are usually performed 1-2 times a month, and multiphase meter uncertainty is often reported at around 10%.
- First-principles models dominate commercial VFM systems. Data-driven and hybrid models are an active research field.
- Accuracy varies widely between wells and studies. One study on 10 wells found oil rate errors from 1.8% to 40.6%.
- A VFM degrades as the reservoir changes. Plan how it will be recalibrated before it goes live.
What is virtual flow metering?
A production system routes many wells through chokes and flowlines to a common separator. For economic operation, operators need the oil, gas and water rate of each well to optimise production, allocate rates and manage the reservoir (Bikmukhametov and Jäschke, 2020).
A virtual flow meter computes those per-well rates. According to the same review, its inputs usually include:
- Bottomhole pressure and temperature
- Wellhead pressure and temperature upstream of the choke
- Wellhead pressure and temperature downstream of the choke
- Choke opening
Where wells use gas lift or electric submersible pumps (ESP), lift gas rates and pump data are added. In other words, a VFM is a soft sensor whose target is the flow rate of each phase.
Why not measure every well directly?
The test separator routes one well at a time to a dedicated separator and measures each phase with single-phase meters. The NFOGM handbook states that a well-instrumented test separator can measure gas with an uncertainty better than 5%, and potentially as low as 2% and 1% for the gas and oil phases if the instrumentation is optimised and separation is ideal. But a test takes hours, and the results are used until the next test. With this setup, a well's rates are typically measured every 7-60 days.
The multiphase flow meter (MPFM) measures at the wellhead without separating the phases. It gives continuous readings, but MPFMs are expensive, need an intervention when they fail, lose accuracy outside their operating range and can degrade from sand erosion (Bikmukhametov and Jäschke, 2020). They also suffer from sensor drift and need regular calibration, so they are not installed on all wells.
| Method | How it works | Strengths | Limits |
|---|---|---|---|
| Test separator | One well at a time through a separator with single-phase meters | Most accurate reference | Hours per test; rates assumed constant between tests |
| Multiphase flow meter | Hardware at the wellhead, no separation | Continuous, per well | Costly; drift, erosion, failures; often around 10% uncertainty |
| Virtual flow meter | Model fed by existing pressure, temperature and choke data | No new hardware; continuous | Depends on sensor health, model quality and regular tuning |
A VFM does not remove the need for the first two. It fills the gaps between them and is calibrated by them.
How does a first-principles VFM work?
First-principles models are the most widely used in commercial VFM systems (Bikmukhametov and Jäschke, 2020). A typical system has three parts:
- A fluid properties model: a black-oil model or an equation of state, tuned to lab analyses of fluid samples and stored as PVT tables.
- Production system models: reservoir inflow (inflow performance relationship), thermal-hydraulic models of the well and flowline, a choke model and, where relevant, an ESP model.
- Data validation and reconciliation: an optimiser that adjusts rates and tuning parameters until the model matches the measured pressures and temperatures and the total measured at the separator.
The review describes the workflow in eight steps, condensed here:
- Build and tune a fluid properties model from PVT lab data.
- Choose the sub-models that the available measurements can support.
- Validate the measurements: remove outliers and filter noise.
- Select tuning parameters, such as choke discharge and heat transfer coefficients.
- Simulate and compare the outputs with measured pressures and temperatures.
- Adjust the parameters to minimise the mismatch, then report each well's oil, gas and water rates.
PVT data is one of the most critical inputs, and fluid properties change over the life of a field, so the PVT model needs recalibration too.
How does a data-driven VFM work?
A data-driven VFM learns the relation between the same measurements and the flow rates recorded by well tests or MPFMs, without a detailed physical description of the well. Once trained, it is fast to run and easy to update. Its main weakness is that it is reliable only within or near the range of its training data (Bikmukhametov and Jäschke, 2020).
The review describes FieldWare Production Universe, a data-driven VFM built from deliberately disturbed well tests, in which an operator changes a well's settings step by step while it flows to the test separator. By 2011 it had been running on around 60% of Shell's producing fields.
Recent research uses data from many wells to make these models more robust:
- Modelling 55 wells from four assets, a multi-task model reduced errors by 25-50% on average for the assets where single-well models struggled.
- On 80 wells, a model pretrained on many wells and calibrated on a new well reached a median error below 5% after only three data points.
Physics, data or both?
Hybrid, or gray-box, models keep a physical structure and learn the uncertain parts from data. On historical data from 10 wells, the mechanistic model achieved an error of 17.2% against 10.3% for the best hybrid model. Separately, adding physics-based models to machine learning on Equinor field data made the models both more accurate and more interpretable.
| First principles | Data-driven | Hybrid (gray box) | |
|---|---|---|---|
| Built from | PVT, inflow, well, flowline and choke models | Historical measurements and reference rates | Physical structure plus learned terms |
| Needs | Deep production engineering knowledge, good PVT data | Enough well tests or MPFM history | Physical insight and field data |
| Strengths | Long field track record; can estimate unmeasured variables | Fast; easy to update with new data | Can reduce a mechanistic model's error; stays physically consistent |
| Weaknesses | Tuning is not straightforward; computationally heavier | Unreliable outside the training range | Balancing physics and data is hard to generalise |
Process surrogates apply a related idea, fitting a fast model to a slow physical one; see our guide to surrogate models for chemical processes.
How accurate is virtual flow metering?
Published numbers vary with the field, the instruments and the method:
- One widely used first-principles system was reported to run on more than 300 wells by 2011, with average errors usually below 5%. The same review cites a neural network on Brazilian fields with errors of 4-7%.
- A study testing VFM simulators against field data reported an average absolute error of 20%, which its authors considered acceptable.
- On 60 wells from five assets, the best half of the models reached errors of 4-6% on historical test data and 8-13% on future data.
Two cautions apply. Error is measured against a reference with its own uncertainty: the same paper notes that models tested on MPFM data cannot reveal the MPFM's own systematic errors. And performance drops on future data, so always test on a later period.
Sensitivity studies summarised in the review add practical points: redundant measurements improve accuracy, wellhead pressure matters more than bottomhole pressure, oil wells are sensitive to the water cut input and gas wells to the gas-oil ratio (GOR), and accuracy falls as the choke opens wider. Drift in pressure and temperature sensors can create a large systematic error in the estimated rates.
Why does a VFM drift, and how do you keep it accurate?
Wells change. In a dataset of 80 wells, water cut typically starts close to zero and rises over the life of the well, often approaching 100%. As the reservoir depletes, operators open chokes to hold production, so later data can differ from the training period. In the 10-well gray-box study, predicting only one week ahead greatly increased accuracy compared with three months.
A study of recalibration on 10 wells found that, with frequent measurements, frequent model updating sustains an excellent prediction performance. With infrequent measurements, frequent updating plus expert knowledge is essential. Operating experience from Total, summarised in the review, also shows that VFM still needs skilled people to tune and calibrate it (Bikmukhametov and Jäschke, 2020).
In practice:
- Compare the VFM with every new well test and track the error per well over time.
- Recalibrate after workovers, changes of lift method, and large changes in water cut or GOR.
- Check input instruments first when estimates jump. A stuck transmitter looks like a change in flow. Our comparison of anomaly detection methods covers ways to catch faulty inputs.
Where is VFM used?
The review describes several roles:
- Back-up to multiphase meters, taking over when a physical meter fails.
- Checking physical meters: VFMs have identified erroneous separator measurements and inconsistent topside meters.
- Standalone metering on a small tie-back field in Norway, where an MPFM would have significantly increased capital and operating costs.
- Production monitoring and optimisation, using continuous rates instead of the last well test.
For measurement and allocation, API MPMS Chapter 20.5 provides recommendations for applying well flow modelling and virtual flow metering, including the use of well tests to validate and update them. Whether a VFM figure is acceptable for allocation or any fiscal purpose on a given licence depends on the regulator, contracts and partners. Operators in Algeria should confirm this with their metering specialists and legal counsel.
How do you start a VFM project?
- List the instruments on each well and note which ones are unreliable.
- Collect the reference data: well test history with test times and conditions, MPFM data where available, and PVT reports.
- Choose the approach from what you have: a calibrated well model favours first principles; long test histories on many similar wells favour data-driven or hybrid models.
- Test on the latest period: train on older data and evaluate on the most recent months.
- Run in parallel with well tests for several months before using the VFM for decisions.
- Agree the recalibration rule and a named owner before go-live.
For the connection to the historian and control system, see our guide to integrating AI with SCADA and OT systems.
What we have seen at KABAS AI
We have not published a virtual flow metering project. Our published oil and gas work is at ADFC (Sonatrach Group), where Virtual Lab supports drilling fluid formulation, a different problem. Two lessons carry over: comparing simulated with actual results drives refinements, and historical data becomes useful only once it is structured and traceable. For VFM, that means:
- Start with the well test archive. Clean, time-stamped tests with recorded choke settings are the basis of any calibration.
- Decide early where data and models will run. Well rates and water cuts describe a field's performance; on-premise deployment keeps both under your control.
- Plan for people. Calibration needs a production engineer and a data scientist working together; our article on building a first industrial data team covers the roles.
Our data science team prepares industrial data, builds predictive models and deploys them with monitoring on your own infrastructure. See our oil and gas solutions and the broader list of AI use cases in North African oil and gas.
Sources
- Bikmukhametov and Jäschke, First Principles and Machine Learning Virtual Flow Metering: A Literature Review, Journal of Petroleum Science and Engineering 184, 2020 (author preprint)
- Bikmukhametov and Jäschke, Combining machine learning and process engineering physics towards enhanced accuracy and explainability of data-driven models, Computers and Chemical Engineering 138, 2020
- NFOGM and Tekna, Handbook of Multiphase Flow Metering, Revision 2, 2005
- Grimstad et al., Bayesian Neural Networks for Virtual Flow Metering: An Empirical Study, Applied Soft Computing 112, 2021 (arXiv 2102.01391)
- Hotvedt et al., On gray-box modeling for virtual flow metering, Control Engineering Practice 118, 2022 (arXiv 2103.12513)
- Sandnes, Grimstad and Kolbjørnsen, Multi-task learning for virtual flow metering, Knowledge-Based Systems, 2021 (arXiv 2103.08713)
- Hotvedt, Grimstad and Imsland, Passive learning to address nonstationarity in virtual flow metering applications, arXiv 2202.03236, 2022
- Løvland, Grimstad and Imsland, Multi-task and few-shot learning in virtual flow metering, Nordic Machine Intelligence 5(1), 2025 (arXiv 2309.15828)
- Mokhtari and Waltrich, Performance Evaluation of Multiphase Flow Models Applied to Virtual Flow Metering, WIT Transactions on Engineering Sciences 105, 2016
- API MPMS Chapter 20.5, Recommended Practice for Application of Production Well Testing in Measurement and Allocation, first edition 2017, reaffirmed 2023 (catalogue description)
Frequently asked questions
Can a virtual flow meter replace a multiphase flow meter?
Sometimes. Published field cases include VFMs used as the only metering on a small tie-back where a multiphase meter would have raised capital and operating costs. More often the VFM runs as a back-up to physical meters and as a check on them. Either way it still needs reference measurements from well tests or meters to calibrate it.
What measurements does a virtual flow meter need?
Typically wellhead pressure and temperature upstream and downstream of the choke, choke opening, and, where installed, bottomhole pressure and temperature. Gas lift rates or pump data help where the well uses artificial lift. It also needs fluid property (PVT) data and a history of well tests or multiphase meter readings for calibration.
How often should a virtual flow meter be recalibrated?
There is no fixed interval. Research on field data shows that accuracy degrades as reservoir and operating conditions change, and that frequent updating keeps it high. In practice, compare the VFM with every new well test, recalibrate when the error grows, and always after a workover, a new lift method or a large change in water cut or gas-oil ratio.
Can virtual flow metering be used for production allocation?
VFM is used in production measurement and allocation, and API MPMS Chapter 20.5 gives recommendations for well flow modelling and virtual flow metering in that context. Whether a VFM figure is acceptable for allocation on your field depends on the regulator, your contracts and your partners, so confirm with them and with your metering specialists before relying on it.
Is a virtual flow meter a soft sensor?
Yes. A VFM is a soft sensor whose target is the oil, gas and water rate of each well. Like other soft sensors, it estimates a hard-to-measure variable from easy measurements, depends on healthy input instruments, and needs a reference measurement, here the test separator or a multiphase meter, to stay accurate.



