RFour Energy
RFour Energy is an independent reservoir-engineering and applied-machine-learning practice working on the upstream subsurface — turning logs, production histories, and core into defensible answers about what is in the ground and how much of it will flow. The throughline is physics first: a model that cannot reproduce something already known about a reservoir is not ready to forecast it.
What we do
We work across reservoir engineering, petrophysics, and applied machine learning for oil, gas, and geothermal. In practice that spans decline-curve and rate-transient analysis, material balance, nodal analysis, rock typing and formation evaluation — including the low-resistivity pay that quick-look logs miss — EOR screening, probabilistic volumetrics, reserves classification under PRMS, and gas- and oil-well surveillance. Each method is applied as a named, traceable technique rather than a black box.
How we work
Five steps, in order. Frame the operational question before reaching for an algorithm. Audit the data that is actually there — well tests, production histories, open-hole logs, the spreadsheets that hold the truth. Prototype a working model fast and test it against events we already remember, not just a train/test split. Document so the next engineer can pick it up without a phone call. And Revisit, because subsurface models age — the ones worth keeping are the ones we are willing to come back to and disagree with.
Selected results
A sample of applied machine-learning work, built with artificial neural network (ANN) and adaptive neuro-fuzzy inference system (ANFIS) architectures. Each is reported on held-out data, not in-sample fit alone.
| Result | Application |
|---|---|
| R² 0.997 | Petrophysics inference for cap-rock evaluation — open-hole logs to petrophysical properties, blind-tested across fields |
| NRMSE 0.032 | Unconventional reservoir properties from log data (coalbed methane) — multi-property inference |
| R² 0.965 | Sorption-parameter prediction (coalbed methane) |
We report performance with the right yardstick — R² paired with NRMSE — and, wherever possible, a blind test, because a single in-sample number is not evidence that a model generalises.
Working stack
Python and the standard ML stack (PyTorch, scikit-learn), ANN and ANFIS, physics-informed methods, and the classical reservoir-engineering toolkit — material balance, decline-curve analysis, reservoir simulation. On top sits an Excel-friendly ingestion layer, interactive web dashboards, and an agentic-AI interpretation layer. Tools are built standalone first, then integrated into the broader workflow.
Tools & writing
We build and maintain a suite of browser-based petroleum-engineering tools — among them an integrated gas-and-oil well-surveillance system and an integrated production-reservoir system — and publish working method notes on the technical blog, organised into collections on production analysis, subsurface characterization, and reserves & reservoir performance.
How we report
Three standards run through everything: performance is stated with uncertainty (R²/NRMSE plus a blind test, never a single number); every analytic claim is traceable to a named method — Archie, Arps, Chan, Coleman-Turner, Waxman-Smits, PRMS — rather than an opaque model; and uncertainty and non-uniqueness are stated plainly, not hidden, such as treating water-drive gas-in-place as an upper bound rather than a single answer.
Get in touch
For project enquiries, reach us at info@rfourenergy.com, or start with the tools and the technical blog.