There is a specific, well-documented way to make an oil or gas well look more valuable than it is, and it hides inside one innocent-looking button: auto-fit.

Decline-curve analysis is the most-used reserves method on earth. You take a well's production history, fit a curve through the dots, extrapolate it to an economic limit, and the area under that curve is the well's estimated ultimate recovery — its EUR, the number that drives its reserves and its value. The math is Arps' 1945 rate-time relations, and for a conventional well in boundary-dominated flow it works beautifully.

The trouble starts when you point it at a shale well and let the software pick the parameters for you.

The b-factor is where the bodies are buried

Arps' hyperbolic decline has three knobs: an initial rate, an initial decline rate, and an exponent called b. That b-factor controls how fast the decline flattens. And it has a physical meaning: for a real reservoir in boundary-dominated flow, b lives between 0 and 1. Exponential decline is b = 0; harmonic is b = 1; everything physical is in between.

Here is the problem the literature has been shouting about for fifteen years. When you let an algorithm find the "best fit" to a modern shale well's early production, it very often comes back with a b-factor greater than 1 — because unconventional wells spend years in transient flow, not the boundary-dominated flow Arps assumed, and a b above 1 is what it takes to force a hyperbolic curve through that transient data. The fit looks gorgeous. The R² is 0.99. And it is a trap.

Why a trap? Because Lee and Sidle proved in 2010 what is now textbook: when b ≥ 1 and you extrapolate the Arps hyperbolic curve to infinite time with no constraint, cumulative production goes to infinity. The well never stops producing. That is, of course, physically impossible — but the software doesn't know that. It just integrates the curve you gave it. The result, in the words of one SPE paper after another, is that "unconstrained use of the Arps hyperbolic relation, particularly where b > 1, can and almost always does yield significant overestimates of reserves."

That's not a fringe opinion. It's the consensus position across Rushing, Lee, Sidle, Okouma, Blasingame, and the rest of the modern DCA literature. Auto-fitting an unconstrained hyperbolic to a shale well systematically overstates EUR. Sometimes by a lot.

What the careful engineers actually do

The fix has been known for years, and it is the opposite of "let the software decide."

The standard discipline is the modified hyperbolic: let the curve be hyperbolic in early time, then splice on a terminal exponential decline — a minimum decline rate (a "Dmin") below which the well is forced to decline exponentially to a finite end. That single constraint turns an infinite, impossible EUR into a bounded, defensible one. It's the difference between a number you can book and a number that will embarrass you at the next reserves audit.

The alternative is to abandon Arps for models actually built for fracture-dominated flow — Duong's power-law model, the stretched-exponential (SEPD), Fulford-Blasingame's transient hyperbolic, logistic growth. They help. But note the honest catch the papers keep flagging: these models are non-unique in the hands of most users, and some (Duong especially) can overestimate in their own way — one study found Duong EURs running up to three times the Arps or SEPD estimate. There is no model that removes the need for judgment. There is only the choice between judgment applied with constraints, and a black box applied without them.

The common thread in every responsible workflow is the same: a human looks at the fit, applies physical constraints, and refuses to extrapolate a curve the reservoir can't actually follow. The b-factor is set with knowledge of the play, not just least-squares. A terminal decline is enforced. The forecast is checked against analog wells and diagnostic plots, not accepted because the curve passed through the dots.

Why "AI" makes this worse, not better

The current wave of "AI-powered" forecasting tools is, under the hood, often doing exactly the thing the literature warns against at scale: finding the parameters that minimize error against the historical data. But a curve that fits the past perfectly and is unconstrained in the future is precisely the failure mode — a high-b hyperbolic with a 0.99 correlation and an infinite tail. Optimizing harder for historical fit can make the EUR more wrong, not less, because the error isn't in the fit. It's in the extrapolation. No amount of training data fixes a model that's allowed to produce a physically impossible answer.

Speed and automation are genuinely useful — nobody wants to hand-decline ten thousand wells. The danger is automation that hides the one decision that matters (where and how the curve is constrained) behind a confident single number.

Where this connects to what we build

This is the exact problem Workollab's Auto-DCA engine was built around. It fits the same Arps curves — validated against a known-good open-source reference (Equinor's library) to a fraction of a part per million — but it does three things a black box won't. It lets you watch the fit happen and override it. It enforces a terminal-decline guard so the curve can't run away to an impossible EUR. And it keeps every assumption visible and auditable, because the b-factor and the Dmin you chose are the whole ballgame and an engineer (and an auditor) needs to see them. It runs in your browser, on your data, with no backend.

The point isn't to remove the engineer. It's the opposite: to put the one decision that drives the number back in front of the person responsible for it — instead of letting "auto-fit" quietly book reserves that aren't there.

A decline curve is a forecast, not a measurement. The least the tool can do is stop it from lying about infinity.

Sources

  • Arps decline relations and the b-factor's physical range (0–1); b > 1 → unbounded cumulative production: J.J. Arps (1945); Lee & Sidle, "Gas Reserves Estimation in Resource Plays," SPE 2010.
  • Unconstrained hyperbolic (b > 1) almost always overestimates EUR; the "modified hyperbolic" / terminal-decline remedy: SPE 162910, Okouma et al. (Blasingame), "Practical Considerations for Decline Curve Analysis in Unconventional Reservoirs"; Rushing et al. (2007).
  • Alternative models (Duong 2011 power-law; stretched-exponential/SEPD; Fulford-Blasingame transient hyperbolic; logistic growth) and their non-uniqueness / over-estimation caveats: modern DCA literature, 2011–2019 (incl. SPE 185589).
  • Workollab Auto-DCA validates against Equinor's open-source decline-curve-analysis library (MIT).

This article is commentary on reserves-forecasting methodology, not reserves advice. For booked reserves, engage a qualified reserves evaluator. One of Workollab's oil & gas engineering pieces; related reading covers our browser-native decline-curve (Auto-DCA) and PVT engines.

Need to turn this engineering logic into a production-ready tool? Explore custom AI automation for oil and gas operators.