By Saad Iqbal
Every December, somewhere in a reservoir engineering department, someone is staring at a spreadsheet with 800 wells in it and a year-end reserves deadline three days out. Each well needs an Arps decline fit, a type curve grouping, and an EUR that will end up in a filing the SEC actually reads. Fit one well by hand in Excel and you can have a defensible number in twenty minutes. Fit 800 of them one at a time and you’ve spent the whole week — and that’s before anyone asks “what if we group by landing zone instead?” That’s the exact problem AI-assisted decline curve tools were built to solve, and it’s worth being precise about which part of the problem they solve, because it isn’t the curve-fitting.
Decline curve analysis was never the hard part
Fitting an Arps hyperbolic decline to a single well’s production history is a solved problem — it has been since the 1940s. The equation takes an initial rate, a decline exponent, and an initial decline rate, and produces a forecast:
q(t) = qi / (1 + b·Di·t)1/b
where q(t) is rate at time t, qi is the initial rate, Di is the initial decline rate, and b is the hyperbolic exponent (b = 0 gives exponential decline, b = 1 gives harmonic decline). Any engineer with a semester of reservoir engineering can fit that curve to one well’s data in a spreadsheet. What nobody can do by hand is fit it correctly to a thousand wells, catch the ones where the fit is garbage because of a choke change or a shut-in, group them sensibly by completion design and landing zone, and roll the whole thing into a defensible reserves number before the filing deadline. That’s the actual job AI-assisted forecasting tools do — not the math, but the scale and the QC.
What AI tools handle decline curve analysis and production forecasting?
| Tool | Core method | Best for | Access model |
|---|---|---|---|
| ComboCurve | Auto-forecasting engine with 30+ configurable variables, type curves & reserves/economics | A&D teams and reserve analysts evaluating hundreds-to-thousands of wells at once | Cloud SaaS, demo-gated pricing |
| Novi Labs | ML “Forecast Engine” trained on completion + production history across basins | Scenario testing across completion designs & landing zones | Cloud SaaS, demo-gated pricing |
| Python + petropt | Open-source Arps hyperbolic/exponential curve fitting (scipy.optimize under the hood) | Single-well QC, custom workflows, full control over assumptions | Free, open source (pip install) |
ComboCurve — forecasting and reserves at portfolio scale
ComboCurve is built around exactly the year-end-deadline problem above. It bulk-imports well databases from ComboCurve itself, Aries, PHDwin, or public data sources, then runs an auto-forecasting engine across more than 30 configurable variables to fit type curves at scale, flag wells where actual production has drifted from forecast, and roll everything into reserves and economics in the same platform. ComboCurve reports 300MM+ wells evaluated in economic scenarios, 510MM forecasts generated, and 4.4MM type curves created on the platform — company-published usage figures, not independently audited, but indicative of how much of this work has already moved off spreadsheets. It’s used by operators, A&D teams, reserve analysts, and mineral royalty companies; pricing isn’t public and the company gates it behind a demo.
Novi Labs — testing scenarios before you drill them
Novi Labs takes a different angle: its Forecast Engine is a machine-learning model trained on completion design and production history across basins, built to answer “what would this well produce if we changed the lateral length, proppant loading, or landing zone” rather than just fitting history that already exists. The company describes the platform as able to test thousands of development scenarios in hours instead of weeks, which is the real value proposition for a team planning next year’s drilling program rather than closing out last year’s reserves. Novi also ships a Model Engine for self-serve machine learning and a Canada-focused product (Novi EVA), and — like ComboCurve — keeps pricing behind a demo request.
Python + petropt — the free path, if you’re comfortable with code
Not every team needs enterprise software to fit a decline curve, and the open-source path has gotten genuinely good. petropt is an MIT-licensed Python library purpose-built for petroleum engineering calculations — Arps decline, PVT correlations, IPR, volumetrics — installable with a single pip install petropt. Fitting a hyperbolic decline to a well’s production history takes a handful of lines:
import petropt
import numpy as np
from scipy.optimize import curve_fit
# t in months, q in bbl/d — your production history
def arps_hyperbolic(t, qi, di, b):
return qi / (1 + b * di * t) ** (1 / b)
params, _ = curve_fit(arps_hyperbolic, t, q, p0=[q[0], 0.08, 1.0], maxfev=5000)
qi, di, b = params
# forecast the next 60 months
t_forecast = np.arange(len(t), len(t) + 60)
q_forecast = petropt.correlations.arps_decline(qi=qi, di=di, b=b, t=t_forecast)This is genuinely the right tool for single-well QC, sanity-checking a vendor forecast against your own fit, or a custom workflow where you need full control over the curve-fitting assumptions rather than trusting a black-box auto-forecast engine. What it doesn’t give you is ComboCurve or Novi Labs’ bulk import, type-curve grouping UI, or reserves rollup — for 800 wells against a filing deadline, you’re back to writing the orchestration code yourself, and at that point the enterprise tools usually pay for themselves in engineer-hours saved.
Choosing between them
The decision usually comes down to what triggered the need. If you’re closing reserves on an existing portfolio and need bulk forecasting, QC diagnostics, and a rollup to economics in one place, ComboCurve’s auto-forecasting and reserves workflow is built for exactly that. If you’re in a planning cycle and need to test how a lateral-length or completion-design change would move the type curve before you commit capital to drilling it, Novi Labs’ scenario-driven Forecast Engine answers a different question than a historical fit ever could. And if you’re a single reservoir engineer who needs one well fit correctly, wants to sanity-check a vendor’s forecast, or is building a workflow too specific for either platform’s UI, petropt and a few lines of Python will get you there for free — and you’ll understand exactly what assumptions went into the number.
What none of these tools fix
Every tool on this list is still fitting a curve to history, and history doesn’t know about the future. None of them account for an artificial lift conversion, an offset frac hit, a choke change, or a workover that shifts the decline trajectory partway through — those show up as a broken fit after the fact, not a warning beforehand. The b-exponent in particular is notoriously unstable on short production histories; a well with six months of data can support wildly different b-values that all fit the visible curve equally well and diverge badly by month 36. Automated forecasting engines make this worse, not better, if nobody reviews the b-value an algorithm picked — a fast wrong answer at portfolio scale is still wrong, just harder to catch. Treat every auto-generated type curve as a starting point for engineering review, not a finished number.
A quick sanity check for any auto-generated forecast
Whichever tool produced the number, run the same two checks before it goes into a reserves report. First, plot the fit against the raw production history on a log-rate-versus-time axis, not just linear — hyperbolic fits that look reasonable on a linear plot often reveal an obviously wrong b-value once you look at them in log space, especially past the point where real data ends and the forecast begins. Second, compare the resulting EUR per well against a handful of offset analogs with three-plus years of actual history in the same landing zone; if the new forecast’s EUR is more than 20-30% off from what similar wells actually delivered, the fit needs a second look before anyone signs off on it. Neither check takes more than a few minutes per well, and both catch the kind of quietly-wrong forecast that a portfolio-scale auto-forecasting engine will happily generate without complaint.
Related reading
Decline curve analysis is only as good as the production data feeding it — see Best AI Tools for Upstream Production Data Analytics for the platforms that clean and QC that data before it ever reaches a type curve. And when a well doesn’t have a physical meter to validate the rate you’re forecasting from, Virtual Flow Metering: How AI Estimates Well Rates covers how AI models fill that gap.
None of these tools change what decline curve analysis fundamentally is — a fit to history, not a guarantee about the future. What they change is how many wells an engineer can responsibly forecast before a deadline, and how much of that time goes to reviewing the fit instead of building it by hand.
