RFour Energy is where petroleum engineering meets machine learning for smarter subsurface and production decisions.
Deep field experience in reservoir engineering, production surveillance, and applied AI research — across multiple basin and reservoir types. The list below is what we've actually worked on long enough to break, fix, and explain to someone else.
Forecasting, anomaly detection, physics-informed models — when analytical methods run thin.
Neural networks, ANFIS, physics-informed models — on real wells, not toy datasets.
Material balance, DCA, economics — auditable, transparent, defensible.
Production data, well tests, simulation — for the morning meeting, not the demo.
Start with the operational question, not the algorithm. Most subsurface problems are misframed long before they're modelled — the time spent here saves the rest.
Inventory what's actually there: well tests, production histories, open-hole logs, simulation runs, the Excel tabs that somehow hold the truth. Surface what's signal, what's swamp.
Build a working model fast — tested against historical events we actually remember, not just train/test splits. If it can't reproduce a known outcome, it isn't ready.
Notebooks, an Excel-friendly tool, a serialized model with notes — whatever lets the next engineer pick it up without a phone call.
Subsurface models age. The ones we keep are the ones we're willing to come back to in six months and disagree with.
Stability-aware classifier — methodology informing GOWIS.
Open-hole logs → petrophysical properties. Blind-tested across fields.
Multi-property AI inference from open-hole logs.
Multi-gas AI prediction from laboratory sorption data.
Performance inference from rock + completion parameters.
Candidate screening, pilot, and field-wide deployment.
Integrated subsurface, drilling, and facility scope.
Map regeneration to localize undrained zones.
RFour Energy is forged by petroleum engineers who cut their teeth in the field—mastering the daily grind of subsurface analysis, complex development planning, and the relentless pressure of production. For us, the physical reservoir always came first; the algorithms were built to serve it.
We deploy Machine Learning at the exact point where classical physics gets noisy and conventional spreadsheets reach their limit. Our focus is precision: gas well surveillance, automated diagnostics, deep reservoir inference, and real-time well health classification. The tools showcased here are the refined outputs of our private internal workflows.
RFour Energy serves as the vault for what we architect, code, and occasionally release. We operate without the noise of traditional marketing. This isn't a brand built for the masses; it is a seal of technical integrity for the work we deliver.
Daily production data for gas and oil wells, turned into well-level diagnosis and alerts. Built end-to-end on Cloudflare Workers + D1 with a frontier LLM whose tool calls are grounded in the surveillance data itself — not a chat wrapper over a document.
DCA, material balance, gas well diagnostics, nodal analysis, rate-transient analysis (RTA), rock typing intelligence, and EOR screening — open to anyone who finds them useful.
Technical notes on reservoir engineering, production surveillance, decline analysis, and applied machine learning in upstream oil & gas — ordered by theme, from subsurface foundations through the reservoir and the wells to reserves and data science.
A mature field needs a mature organization — a light matrix of three value streams: base management, development & well delivery, and modeling, planning & asset development, with one integrated forecast and a clear governance rhythm.
The reservoir & petroleum engineer’s role across the exploration well lifecycle — planning and data acquisition, drilling execution and well testing, evaluation, reserves, and the development plan.
Reading a basin's story in plain language — the petroleum system, basin types, stratigraphy, structural traps, and how to read a cross-section. No degree required.
An ultrasound of the earth. Echoes, acoustic impedance, two-way time versus depth, stacking and migration — and how to read horizons, faults, and bright spots on a section.
The 1D mechanical earth model and its mud weight window, the 2D section and stress polygon, the 3D volume with fault slip tendency, and 4D coupling where depletion moves the stress path — with a plain test for which rung a decision actually needs.
Each method assumes a mechanism, and picking the wrong one fails in a predictable direction: compaction trends and Eaton's exponent, the Bowers loading and unloading curves, the centroid in dipping reservoirs, and what actually counts as calibration evidence.
Two different failures with one symptom. Pay can hide because the log lacks contrast — laminated sands, fresh water, conductive minerals — or because the rock cannot flow at all. The six mechanisms, why Archie fails, and the test that tells LRLC apart from a low-quality reservoir: pore-throat classification, capillarity, and whether correcting saturation changes anything.
A Python, agentic AI, Excel, and dashboard framework for modern reservoir characterization. Six methods, one platform — Winland R35, FZI/HFU, PGS, Lorenz, J-function, and ML on logs.
OGIP and drive diagnosis from pressure decline. P/Z, Cole, Roach plots with three worked numerical examples.
OOIP and drive diagnosis from pressure decline. Havlena–Odeh straight line, drive indices, and three worked examples whose true answer is known.
From the conservation equations to forecasting practice: discretisation and grid error, the Peaceman well model, where uncertainty concentrates, history matching as an ill-posed problem where a good match can still forecast badly, and the different job simulation does in green versus mature fields.
Thermal, gas/miscible, and chemical EOR on one map — steam, SAGD, CO₂, polymer, surfactant, ASP. Mechanisms, screening criteria, and when each method actually pays. Companion screening tool at /eor.
Minimum miscibility pressure and what CO₂ does to oil, then the two architectures: a pattern flood limited by sweep rather than displacement, and cyclic single-well huff and puff where soak time and blowdown do the work — with a decision table and what actually kills projects.
Made from air on location, yet the reservoir least wants it: the MMP ladder against the fracture limit, gravity-stable crestal displacement where buoyancy becomes the mechanism, and cyclic N₂ driven by repressurization rather than dissolution.
Polymer fixes sweep, surfactant fixes trapped oil, alkali makes surfactant affordable, ASP does all three — with the capillary desaturation curve, optimal salinity and the Huh relation, and the cyclic chemical treatment that works by imbibition where no flood can sweep.
Practitioner's guide to nodal analysis. Where IPR meets VLP, why the operating point is never decided at any single component, and how to translate seventy years of correlations into something that runs on your laptop.
Producing stacked pay zones singly or together, compared from both sides: differential depletion, crossflow and thief zones on the reservoir side; composite IPR, artificial lift and allocation on the production side — with a decision table and the intelligent-completion middle ground.
Gas wells don't fail because the reservoir runs dry — they fail because liquids win the upward race. Physics, equation, interpretation.
When to use exponential, hyperbolic, or harmonic. The b-factor as reservoir physics, not a fitting knob.
When Arps over-extrapolates, plot 1/q vs Nₙ/q, fit a line, take the inverse slope. A rigorous DCA alternative.
Mature waterfloods need a different forecasting tool than rate-time decline. Practical guide with worked example.
Coning, channeling, or near-wellbore problem? Read WOR signatures correctly and the workover plan writes itself.
Problems ranked by how treatable they are, why coning and normal sweep are not candidates, mechanical versus chemical options, disproportionate permeability reduction, and why placement geometry decides the outcome more than the choice of chemical.
A physics-first, multidisciplinary workflow from production gap to risked-value ranking — G&G, reservoir, and production engineering. Nodal analysis, a diagnosis-to-intervention map, plus artificial-lift, sand-control, and water shut-off screening, with an end-to-end workflow chart.
Candidate screening on permeability and damage, width models and the conductivity optimum that depends on what is held fixed, fluids and proppant, the Nolte-Smith diagnostic while pumping, monitoring, and the gap between created, propped and effective length — with published field applications.
The complete framework under SPE-PRMS 2018 — the two-axis matrix, Reserves/Contingent/Prospective, 1P/2P/3P, P90/P50/P10, project maturity, aggregation, and PRMS vs SEC vs UNFC.
Why P50 isn't enough — and how Monte Carlo Simulation transforms volumetric reserves estimation. Distributions, multi-zone aggregation, and tornado analysis.
A systematic tour across the value chain — seismic CNNs, facies classification, surrogate models, production forecasting, ESP failure prediction, physics-informed ML, and the subsurface pitfalls that sink naive models.
From observed field behavior back to mechanism — inverse problems and history matching, experimental design, ML surrogates, and the tool-assisted agentic AI loop, with a step-by-step Python stack and an exception-based surveillance dashboard, all under physics-first governance.
RCAL and SCAL end to end — the bias that makes a plug dataset optimistic, the corrections that are often skipped, the cleaning step that destroys wettability, and the laboratory artifact that manufactures residual oil.
Fluid classification, sampling and representativity, the four laboratory studies and what each measures, the differential-to-field-basis conversion nobody checks, and the consistency tests that catch a bad report before it reaches the model.
A decline forecast that cannot be re-run is not a forecast. What has to be true of the code that replaces the workbook — judgements as declared parameters, a bounded Arps fit that raises rather than guesses, and a validation harness reporting blind-test error instead of a single R².
Four ways a machine-learning result on production data can be wrong while every validation check reports success — an unasked baseline, a model that structurally cannot extrapolate, a score that measured the choosing, and a feature containing its own answer. Each measured on a five-well history.
When the product is heat, not fluid. How petroleum reservoir engineering transfers to geothermal — energy-in-place, material balance, reinjection, and thermal breakthrough.
Reservoir question, ML approach, dataset you'd like a second pair of eyes on, or just a topic worth discussing — drop a note. Usually replies within a business day.
