By Saad Iqbal
The compressor gauge on the pad reads 2.1 MMscf/d, and that number is the whole problem. Three gas-lifted wells are tied into it, each one hungry for more injection gas than the field can spare, and the production foreman has to decide, by Thursday, how to split a budget that never quite stretches far enough. Add gas to Well A and its liquid rate climbs a little. Pull gas from Well C and its rate falls a lot more than anyone expected. Somewhere in that trade-off is the combination that squeezes the most oil out of a fixed gas supply — and finding it by trial and error, well by well, is how a lot of gas lift allocation still gets done today.
Gas lift allocation optimization is the discipline of answering that question properly: given a limited volume of lift gas and a set of wells competing for it, how do you split the supply so total liquid production is as high as it can be? The math behind the answer is decades old. What has changed is how often, and how accurately, that math gets solved — because machine learning can now keep the performance curve behind every well current enough to re-run the allocation every day instead of every quarter.
What a gas lift performance curve actually shows
Every gas-lifted well has a performance curve: liquid production rate plotted against gas injection rate, holding everything else constant. Pump in too little gas and the fluid column in the tubing stays too heavy to lift efficiently. Pump in too much and the extra gas just adds friction without adding much lift — the curve flattens, then can even turn down. Somewhere between those two extremes sits the point of diminishing returns, where one more Mscf/d of gas buys the smallest incremental barrel the well will ever give up.
That curve is built from nodal analysis: matching the well’s inflow performance relationship against its vertical lift performance at a series of assumed injection rates, then reading off the liquid rate at each one. Done by hand in well-test software, it is accurate the day it is built and stale a few weeks later, because water cut, reservoir pressure, and tubing condition all keep moving. A machine learning model trained on the well’s own injection and production history — rather than a single static multiphase-flow run — can re-forecast that curve continuously, picking up drift in water cut or productivity index before anyone schedules a new well test.

The equal-slope method: the classic answer to gas lift allocation optimization
Once every well has a performance curve, allocating a fixed gas supply across them is a constrained optimization problem, and it has a clean, well-established solution: the equal-slope method, sometimes called the equal-marginal-value method. The rule is simple to state. At the optimal allocation, the slope of every well’s performance curve — the incremental barrels of liquid gained per incremental Mscf/d of gas — is the same across every well in the system.
The logic is pure economics. If Well A’s curve is steeper than Well B’s at their current injection rates, the next Mscf/d of gas is worth more at Well A. Moving a unit of gas from B to A raises total liquid production, full stop. You keep shuffling gas from the flattest slope to the steepest until no such move helps anymore — at which point every well’s marginal return is identical, and the fixed gas budget is doing the most work it possibly can.
- Build (or forecast) a performance curve for every well sharing the gas supply.
- Pick a starting allocation and compute the local slope — the derivative of liquid rate with respect to injection rate — at each well’s current point.
- Shift a small increment of gas from the well with the lowest slope to the well with the highest slope.
- Recompute the two affected slopes and repeat until every well’s slope converges to the same value, or the gas budget runs out first.
- Check that no well has been pushed past its mechanical or safety limits — maximum injection pressure, dome setting, or surface choke capacity — along the way.

Worked by hand on three or four wells, that’s an afternoon with a spreadsheet. On a pad of thirty wells sharing one compressor train, it’s a linear or quadratic program, and most integrated production modeling packages will solve it for you once the curves are loaded in. The equal-slope result itself hasn’t changed since the method was formalized decades ago — what’s changed is how fast and how often those input curves can be refreshed, which is where the optimization actually lives or dies in the field.
Where the classic method quietly breaks down
Equal-slope allocation is only as good as the curves feeding it, and curves go stale fast. A well that waters out gains a steeper decline in its lift-gas response than its neighbors; a well recovering from a workover temporarily looks far more gas-hungry than its long-run behavior justifies. If the performance curves were last built from a well test run two months ago, the allocation being executed today is optimizing against a field that no longer exists. Engineers compensate by re-running tests more often than the budget really allows, or by padding allocations toward wells they suspect have drifted — which quietly defeats the purpose of solving the problem rigorously in the first place.
Closing the loop: how ML keeps the allocation current
The fix is to stop treating the performance curve as a one-time deliverable and start treating it as a live model. A gradient-boosted or neural regression model trained on each well’s historical gas injection rate, wellhead pressure, water cut, and resulting liquid rate can re-estimate the local shape of the performance curve every time new SCADA data lands — typically daily, sometimes hourly. Feed those refreshed curves straight into the equal-slope solver, and the allocation itself becomes a standing, automated recommendation rather than a quarterly workshop exercise.

Operators running this as a genuinely closed loop add one more step before anything touches a surface choke or dome regulator: a constraint check against each well’s safe operating envelope, and a rate-of-change limiter so the controller never swings an injection rate hard enough to trigger casing-heading instability. That last part matters — an allocation engine that is technically optimal but physically unstable is worse than no optimization at all, which is exactly the failure mode covered in our look at detecting gas lift instability and casing heading.
Does equal-slope allocation still work when wells decline at different rates?
Yes, and that’s actually the case where refreshing the curves matters most. Equal-slope allocation makes no assumption that wells decline together — it only assumes you know each well’s current performance curve. A well declining faster than its neighbors will show a changing slope at any fixed injection rate, and the method will naturally shift gas away from it over time as its marginal return falls. The method breaks only when the curves it’s fed are outdated, not when the underlying wells behave differently from each other. That is the whole argument for re-forecasting curves continuously instead of per well test.
Tools built for this today
A handful of production-optimization platforms now package this loop rather than leaving operators to assemble it from a well-test database and a spreadsheet solver. SLB’s OptiLift is built specifically around unconventional lift optimization, connecting wells and operations teams to recommend lift changes across a field rather than one well at a time. Weatherford’s ForeSite platform layers asset-wide physics-based modeling with machine-learning-based predictive failure management to flag bottlenecks and plan interventions before they cost production. Neither tool eliminates the engineering judgment behind the equal-slope method — they just mean the curves feeding it don’t go stale between well tests.
If gas lift allocation at the field level is new territory, it’s worth reading backward to the well-level problem first: our earlier piece on AI gas lift optimization and injection-rate forecasting covers how a single well’s optimal setpoint gets found before it ever competes for gas with its neighbors, and our roundup of gas lift optimization software compares the platforms built to run both jobs at scale.
The allocation is only as good as the curve underneath it
None of this replaces engineering judgment on the pad. A production engineer who knows that Well C just came back from a workover, or that Well A’s water cut jumped after an offset frac, should still override what a model recommends until the curve catches up. What changes is the default: instead of re-optimizing the gas budget once a quarter on hand-built curves, the equal-slope allocation runs continuously against curves that are themselves refreshed from the data the field is already collecting. The method hasn’t changed. How often you get to use it properly has.
Looking to put this into practice on your own pad? Start by pulling six months of injection-rate and liquid-rate history for every well sharing a compressor train — that history is what any performance-curve model, machine-learned or not, has to be built on.
