RFour Energy · Field Notes

Reservoir Simulation — What the Model Knows

Q: What does a reservoir simulator actually solve?

It solves conservation of mass for each fluid component over a set of connected control volumes, with Darcy's law supplying the flow between them and an equation of state or black-oil PVT table supplying the fluid properties. Combining conservation with Darcy's law produces a set of coupled non-linear partial differential equations that have no general analytical solution, so the reservoir is divided into grid blocks and time into steps, the equations are written in discrete form for every block, and the resulting large non-linear system is solved iteratively at each timestep.

Q: Why is the well-block pressure not the flowing bottomhole pressure?

Because the simulator computes one average pressure for an entire grid block that may be a hundred metres across, while the flowing pressure exists at a wellbore a few centimetres in radius, and pressure varies logarithmically in between. Peaceman showed that the calculated well-block pressure is essentially equal to the actual flowing pressure at an equivalent radius of about one-fifth of the block dimension for a square grid, and later generalised this to rectangular grids and anisotropic permeability. The well index built from that equivalent radius is what connects block pressure to bottomhole pressure, and getting it wrong misprices every rate the model produces.

Q: What is history matching and why is it not enough?

History matching is the process of adjusting uncertain model parameters until simulated production and pressure reproduce what the field actually did. It is an inverse problem, and it is ill-posed: many different parameter combinations reproduce the same history. Published work using a deliberately simple model has shown that the objective function has multiple minima, that a good fit to production data does not guarantee good estimates of the underlying parameters, and that a well history-matched model can produce a bad forecast. A match is therefore a necessary condition for a credible forecast, not a sufficient one.

Q: How does simulation differ between a green field and a mature field?

In a green field there is no production history to match, so the model is a pure hypothesis and the forecast range is governed almost entirely by static uncertainty: volumes, connectivity, aquifer strength and relative permeability. The right output is a wide, honestly stated range of scenarios rather than a single case. In a mature field there is abundant dynamic data, so the range narrows, but a new risk appears: a model tuned to reproduce the past can lock in a reservoir description that is wrong in ways the past could never reveal, and that error surfaces only when the field is asked to do something it has not done before.

A simulator is not a measuring instrument. It is a hypothesis that computes — and the discipline is in keeping track of which parts of it are constrained by data and which parts are simply things somebody decided.

RFour Energy · Reservoir Engineering Series

Numerical simulation is the only tool that can carry geology, fluid physics, well hydraulics and a development plan in one place and produce a number a company can spend money against. That is why it is used, and it is also why it is dangerous: the output arrives as a smooth curve with four significant figures, and nothing in that presentation distinguishes the parts that follow from measured data from the parts that follow from an assumption made in an afternoon. This article walks the whole path — the equations, the discretisation, the grid, the well model, uncertainty, history matching, and forecasting in green and mature fields — with particular attention to where the confidence is real and where it is manufactured.

What the Simulator Actually Solves

Underneath every commercial package is a short physical statement: mass is conserved. For each component — oil, water, gas in the black-oil formulation; individual hydrocarbon components in a compositional one — the mass entering a volume of rock, minus the mass leaving, equals the change in mass stored inside it.

Flow between volumes comes from Darcy's law, with relative permeability and viscosity setting how each phase moves, and capillary pressure setting the difference between phase pressures. Fluid properties come from PVT tables or an equation of state. Put those together and the result is a set of coupled, non-linear partial differential equations with no general analytical solution — which is precisely why the problem is handed to a computer rather than to a formula.^[1][2]

FIG. 01The whole numerical method in one block. Conservation is written over a finite control volume, flux across each face comes from Darcy's law using the potential difference to the neighbour, storage is the time derivative of what the block holds, and wells enter as source or sink terms. Everything else — solver choice, timestep control, linearisation — is machinery for solving this statement quickly enough to be useful.^[1][2][3]

Two solution strategies dominate, and the choice is an engineering trade rather than a preference. IMPES solves pressure implicitly and updates saturations explicitly — cheap per timestep, but stable only for small steps. Fully implicit solves everything simultaneously — expensive per step, stable at much larger ones, and the default wherever there is coning, high-rate gas or strong non-linearity. A model that runs slowly is often a model asking the wrong solver, not a model that needs a bigger machine.

Building the Model

The build is a chain, and every link degrades information from the one before it.

Structural and stratigraphic framework. Surfaces, faults, zonation. This decides connectivity, and connectivity decides more forecasts than permeability does.
Property population. Facies, porosity, permeability, saturation — geostatistically distributed between wells. Between control points this is a plausible field, not a measured one.
Upscaling. A geological model with tens of millions of cells becomes a flow model with far fewer. Averaging permeability is not neutral: the arithmetic mean flatters layered flow, the harmonic mean punishes it, and the right answer is directional.
Rock and fluid physics. Relative permeability and capillary pressure curves per rock type; PVT per fluid region. Relative permeability is usually the single most influential and least constrained input in the entire model.
Initialisation. Equilibration to a consistent initial state — contacts, transition zones, initial pressure. A model that will not initialise cleanly has a contradiction in it, and running it anyway just hides the contradiction.
Wells and completions. Trajectories, perforations, lift, constraints — and the well model discussed below.

Grid design, and the error nobody sees

Coarse grids do not merely lose detail; they actively distort flow. Numerical dispersion smears sharp fronts, so a water or gas front that should arrive as a step arrives as a ramp, and breakthrough looks earlier and gentler than it is. Grid orientation effects can bend a displacement toward the grid axes. Neither error announces itself — both simply produce a result that looks reasonable.^[3] The practical defence is a sensitivity run at finer resolution on a representative segment: if the answer moves materially, the coarse grid is part of the answer.

The well model — a small equation with large consequences

A simulator computes one average pressure for a block that may be a hundred metres across. A well is a few centimetres in radius, and pressure near it varies logarithmically. Connecting the two is not a detail; it decides every rate the model reports.

FIG. 02Peaceman's result, and the reason well rates are believable at all. For a square grid the calculated well-block pressure equals the actual flowing pressure at a radius of about 0.2 Δx;^[4] the later generalisation to rectangular grids and anisotropic permeability gives r₀ = 0.14(Δx²+Δy²)^½, with the exact constant e^−γ/4.^[5] The two are consistent to four figures: 0.14036 × √2 = 0.1985. Everything a model says about drawdown, lift and coning passes through this conversion.

Where the Uncertainty Actually Lives

Simulation uncertainty is usually discussed as though it were spread evenly across the inputs. It is not. It concentrates, and knowing where lets a study spend its runs usefully.

Source	Typical influence on forecast	How well data constrains it
Volumes in place	Scales almost everything	Moderate — but a match can be achieved at several volumes
Connectivity, faults, baffles	Very high — decides sweep and drainage	Poor between wells; often the real forecast driver
Relative permeability	Very high — sets breakthrough and recovery	Poor: few measurements, upscaled to blocks they never described
Aquifer strength	High in water-drive systems	Only after enough depletion to see the response
Permeability field	High locally, moderate at field scale	Well control only; interpolated elsewhere
PVT	Moderate	Usually good — this is measured
Numerical choices (grid, solver)	Underestimated, sometimes large	Not data at all — an analyst decision

The last row deserves the emphasis. Grid resolution, upscaling method and timestep control are modelling decisions, yet their effect on a forecast can rival a genuine geological uncertainty. They rarely appear in an uncertainty study because they are not thought of as parameters.

Method follows from this. Screen many parameters cheaply with experimental design to find the few that matter; carry those few through a proper probabilistic treatment. What matters more than the technique is honesty about the prior: a range set by what the team is comfortable defending is not an uncertainty range, it is a comfort range, and it will be too narrow in exactly the direction that hurts.

History Matching — An Ill-Posed Problem, Treated Honestly

History matching adjusts uncertain parameters until the model reproduces what the field actually did. Stated that way it sounds like calibration. It is not calibration; it is an inverse problem, and it is ill-posed: the data do not contain enough information to determine the parameters uniquely.^[6][7]

This is not a philosophical caveat. It has been demonstrated deliberately, on a model simple enough that the truth is known. Using a two-dimensional layered cross-section with a fault and just three unknown parameters — high permeability, low permeability and fault throw — the objective function turns out to have multiple minima rather than one; a good fit to production data does not guarantee good estimates of the parameters; and a well history-matched model can produce a bad forecast. The authors go further: the true model is not necessarily the one conventional history matching will find.^[6][7] A companion paper on the same work carries the blunt title Our calibrated model has no predictive value.^[8]

FIG. 03The central fact of history matching, drawn. Five parameter sets reproduce the measured history identically — the black curve through the data points — and then separate as soon as they are asked to predict, ending a factor of about 2.2 apart. Nothing in the history period distinguishes them. This is why a single history-matched case presented as the forecast is a statement about the analyst's confidence rather than about the reservoir.^[6][7]

What good practice looks like

Match what you intend to forecast. If the decision is about water breakthrough in a new well, matching field-total oil rate proves very little. Match at the level the decision lives.
Resist over-parameterisation. Freeing more parameters always improves the fit and usually worsens the forecast; the optimum level of parameterisation is itself a classical result in this field.^[9] Multipliers with no geological meaning are the clearest symptom.
Change things for reasons. Every adjustment should have a defensible physical story. "It improved the match" is not one.
Keep several matched models. An ensemble of matched cases is a more honest object than one polished case, and modern assisted and ensemble-based methods are built around exactly that idea.^[10][11][13]
Hold data back. Match on part of the history and test on the rest. It is the only cheap way to find out whether the model predicts or merely interpolates.

A history match is a necessary condition for a credible forecast. It has never been a sufficient one.

Forecasting — Green Field versus Mature Field

The same software does two genuinely different jobs, and confusing them is a common source of misplaced confidence.

FIG. 04Two different jobs. Without production history the forecast band is wide because static uncertainty — volumes, connectivity, aquifer, relative permeability — is unconstrained, and a wide honest band is the correct deliverable. With a long history the band narrows, but narrowness is not accuracy: a model tuned to reproduce the past can be confidently wrong about anything the past never tested, and that error appears only when the field is asked to do something new. One simplification is worth naming: both bands are drawn opening from a single point, because early rate is largely set by well count and facility capacity. Early-rate uncertainty is not literally zero — it still carries the uncertainty in kh — but it is small compared with the divergence that volumes, connectivity and aquifer strength produce later.

Green field

There is no history to match, so the model is a hypothesis in its entirety. The useful output is not a number but a structured set of scenarios: low, mid and high cases built on stated geological alternatives rather than on parameter multipliers, plus the answer to the question that actually matters at sanction — which uncertainty, if resolved, would change the decision? That question turns a simulation study into an appraisal plan. A single-deterministic green-field forecast carried into an economic model is the most expensive form of false precision in this discipline.

Mature field

Here the model is constrained, and it earns its keep on questions that cannot be answered by decline analysis: infill well placement and expected interference, water or gas injection patterns, EOR screening, workover and conformance targets, and the field-wide effect of a facility constraint. The specific risk is subtler than in a green field — the match may be excellent and the description still wrong, and the failure will not surface until the field is asked to do something it has not done before. Which is, of course, the exact circumstance in which the forecast was commissioned.

What Simulation Cannot Do

It cannot create information. A model interpolates and extrapolates what it was given. Between wells it produces plausibility, not knowledge.
It cannot resolve an ambiguity the data does not contain. If two geological interpretations fit the data equally, no amount of running resolves the tie — only new data does.
It cannot validate itself. A converged run is a solved system of equations, not a correct description of the reservoir.
It cannot make a decision robust. That is what scenario ranges and value-of-information analysis are for.

None of which argues against simulation. It argues for reporting it correctly: with the ranges attached, the assumptions named, and the parts that came from an afternoon's decision clearly marked as such.

Validation

The formulation of the flow equations, their discretisation and the standard solution schemes follow Aziz and Settari^[1] and Ertekin, Abou-Kassem and King;^[2] the practical treatment of model construction, gridding and numerical error follows Mattax and Dalton.^[3] The well model in Fig. 02 is Peaceman's: the 1978 result that the well-block pressure equals the actual flowing pressure at a radius of about 0.2 Δx for a square grid,^[4] and the 1983 generalisation to rectangular grids and anisotropic permeability giving 0.14(Δx²+Δy²)^½ with the exact constant e^−γ/4.^[5] The non-uniqueness demonstrated in Fig. 03 — multiple minima in the objective function, a good production fit with poor parameter estimates, and a good match giving a bad forecast — is Tavassoli, Carter and King's,^[6][7] with the companion paper stating the conclusion in its title.^[8] The over-parameterisation argument follows Shah, Gavalas and Seinfeld.^[9] Modern history-matching methodology, including ensemble approaches and the case for carrying multiple matched models, follows Oliver, Reynolds and Liu^[10] and the review by Oliver and Chen.^[11] Benchmarking of simulator behaviour against a common problem follows the SPE comparative solution projects.^[12]

The simulator is honest by construction — it solves exactly the problem it was given. Whether that was the right problem is entirely the engineer's responsibility.

Related tool

IPRS — Integrated Production Reservoir System

Three coupled physics layers on one canvas — reservoir material balance, well performance and the surface network.

A working example of coupled layers, at a scale where you can still see what each one is doing.

Open IPRS →

References

Aziz, K. & Settari, A. (1979). Petroleum Reservoir Simulation. Applied Science Publishers, London; ISBN 0-85334-787-5. Long known as “the green book” of reservoir simulation. Formulation of the flow equations, finite-difference discretisation, IMPES and fully implicit schemes, solution of the resulting non-linear systems.
Ertekin, T., Abou-Kassem, J.H. & King, G.R. (2001). Basic Applied Reservoir Simulation. SPE Textbook Series Vol. 7, Society of Petroleum Engineers, Richardson; ISBN 978-1-55563-089-8. (Some secondary reference lists cite this as Volume 10; the publisher and library records give Volume 7.)
Mattax, C.C. & Dalton, R.L. (1990). Reservoir Simulation. SPE Monograph Vol. 13, Henry L. Doherty Series, Society of Petroleum Engineers. Practical model construction, grid design, upscaling and the recognition of numerical error.
Peaceman, D.W. (1978). Interpretation of Well-Block Pressures in Numerical Reservoir Simulation. SPE Journal 18(3), 183–194; SPE-6893-PA. Establishes that the well-block pressure is essentially equal to the actual flowing pressure at a radius of 0.2 Δx for a square grid.
Peaceman, D.W. (1983). Interpretation of Well-Block Pressures in Numerical Reservoir Simulation with Nonsquare Grid Blocks and Anisotropic Permeability. SPE Journal 23(3), 531–543; SPE-10528-PA. Gives r₀ = 0.14(Δx²+Δy²)^½, with the exact constant e^−γ/4 where γ is Euler's constant.
Tavassoli, Z., Carter, J.N. & King, P.R. (2004). Errors in History Matching. SPE Journal 9(3), 352–361; SPE-86883-PA, doi:10.2118/86883-PA. A good fit to production data does not guarantee good parameter estimates and may lead to a bad forecast; the true model is not necessarily the one conventional history matching finds.
Tavassoli, Z., Carter, J.N. & King, P.R. (2005). An Analysis of History Matching Errors. Computational Geosciences 9(2), 99–123. Multi-modal objective functions and consequently non-unique history matches; a well history-matched model can give a bad forecast.
Carter, J.N., Ballester, P.J., Tavassoli, Z. & King, P.R. (2006). Our Calibrated Model Has No Predictive Value: An Example from the Petroleum Industry. Reliability Engineering and System Safety 91(10–11). First presented at SAMO 2004; the journal version carries the 2006 date, and both years appear in secondary citation lists.
Shah, P.C., Gavalas, G.R. & Seinfeld, J.H. (1978). Error Analysis in History Matching: The Optimum Level of Parameterization. SPE Journal 18(3), 219–228; SPE-6508-PA.
Oliver, D.S., Reynolds, A.C. & Liu, N. (2008). Inverse Theory for Petroleum Reservoir Characterization and History Matching. Cambridge University Press, Cambridge; ISBN 978-0-521-88151-7.
Oliver, D.S. & Chen, Y. (2011). Recent Progress on Reservoir History Matching: A Review. Computational Geosciences 15(1), 185–221.
Odeh, A.S. (1981). Comparison of Solutions to a Three-Dimensional Black-Oil Reservoir Simulation Problem. Journal of Petroleum Technology; SPE-9723. The first of the SPE comparative solution projects, in which independent simulators are run on a common specified problem.
Aanonsen, S.I., Nævdal, G., Oliver, D.S., Reynolds, A.C. & Vallès, B. (2009). The Ensemble Kalman Filter in Reservoir Engineering — a Review. SPE Journal 14(3), 393–412. The methodological basis for carrying an ensemble of matched models rather than a single tuned case.

Frequently Asked Questions

What does a reservoir simulator actually solve?

Conservation of mass for each fluid component over a set of connected control volumes, with Darcy's law supplying flow between them and PVT or an equation of state supplying fluid properties. That combination gives coupled non-linear partial differential equations with no general analytical solution, so the reservoir is divided into grid blocks and time into steps, and the discrete system is solved iteratively at each step.

Why is the well-block pressure not the flowing bottomhole pressure?

Because the simulator computes one average pressure for a block that may be a hundred metres across, while flowing pressure exists at a wellbore centimetres in radius, with logarithmic variation in between. Peaceman showed the block pressure equals the actual flowing pressure at an equivalent radius of about 0.2 Δx for a square grid, later generalised to rectangular grids and anisotropic permeability. The well index built from that radius is what links block pressure to bottomhole pressure.

What is history matching and why is it not enough?

Adjusting uncertain parameters until the model reproduces what the field did. It is an ill-posed inverse problem: many parameter combinations reproduce the same history. Published work on a deliberately simple model shows multiple minima in the objective function, good production fits with poor parameter estimates, and good matches producing bad forecasts. A match is necessary for a credible forecast, never sufficient.

How does simulation differ between a green field and a mature field?

In a green field there is no history to match, so the model is a hypothesis and the forecast range is governed by static uncertainty — volumes, connectivity, aquifer, relative permeability. The correct deliverable is a stated range of scenarios. In a mature field the range narrows, but a model tuned to the past can lock in a wrong description, and the error surfaces only when the field is asked to do something new.