Interactive petroleum-engineering portfolio

Test a simulated coreflood treatment scenario.

This case study turns a petroleum-engineering research question into an end-to-end ML product. Explore how thermal conditions, wax-prone crude and injection design shape predicted oil recovery and cooling-damage risk.

Understand the project Go to the interactive demo ↓

2frozen ML pipelines

21validated inputs

Why this project exists

From an engineering question to a deployable ML system.

Cooling wax-prone crude can precipitate solids, restrict pore flow and change the response to a chemical treatment. This project investigates whether a physics-informed synthetic coreflood dataset can support repeatable estimates when experimental records are limited.

01 · Engineering problem

Cooling can alter flow.

Wax onset, viscosity, wettability and injection conditions interact nonlinearly, making a single hand calculation insufficient.

02 · ML framing

Two related predictions.

XGBoost estimates oil recovery as a continuous value. Multinomial Logistic Regression estimates low, moderate or high damage.

03 · Portfolio evidence

More than a calculator.

The interface is the final layer of a documented workflow covering data design, quality, leakage control, model comparison and deployment.

Skills demonstrated

Physics-informed data design
Blind data-quality assessment
Leakage-safe preprocessing
Group-aware validation
Regression & classification
Production inference

Demonstration boundary

The data are synthetic and physics-informed, created for portfolio demonstration. Predictions illustrate ML practice and do not replace laboratory validation or field engineering decisions.

How the project was built

Eight controlled steps from study concept to live inference.

The result shown below is produced only after the data, preprocessing and candidate models have passed separate quality and validation gates.

01
Frame the problem
Translate the cooling-damage study into recovery regression and damage-class classification tasks.
02
Design canonical data
Specify one simulated coreflood per row, with units, ranges, correlations and physical constraints.
03
Emulate raw records
Add controlled missingness, duplicates, label inconsistencies, unit errors, outliers and batch effects.
04
Assess quality blindly
Use EDA to identify suspected issues without consulting the hidden clean data or issue map.
05
Clean & split safely
Correct approved issues and separate train, validation and test data by core family to prevent leakage.
06
Compare candidates
Evaluate baselines, linear models, Random Forest and XGBoost with fold-specific preprocessing.
07
Freeze final models
Select on validation, evaluate once on sealed test data, then approve XGBoost and Logistic Regression.
08
Serve & demonstrate
Package the frozen pipelines behind a validated API and expose them through this interactive experience.

Quick start

Choose an example scenario

01 / Rock systemCore & initial stateGeometry and pre-flood petrophysical conditions.+

Core diameter

Core plug diameter.Range 2.42–2.5

Core length

Core plug length.Range 4.8–5.1

Initial porosity

Initial pore-volume fraction.Range 14–22

Initial permeability

Initial absolute permeability.Range 10–40

Initial oil saturation

% PV

Oil-filled share of pore volume.Range 55–85

Confining pressure

MPa

Applied pressure around the core.Range 7–9

02 / Crude characterOil & interfaceWax tendency, fluid properties and wettability.+

Oil density at 20°C

g/cm³

Reference crude-oil density.Range 0.84–0.92

Wax content

wt%

Crude-oil wax fraction by mass.Range 15–35

DSC wax onset

°C

Wax precipitation onset measured by DSC.Range 35–65

Rheology anomaly

°C

Temperature where rheology departs from warm behavior.Range 20–50

Oil viscosity at 50°C

mPa·s

Reference viscosity used to derive test-temperature viscosity.Range 15–80

Initial contact angle

degrees

Oil-drop contact angle before treatment.Range 35–80

Formation-water salinity

mg/L

Dissolved mineral concentration.Range 120000–175000

03 / Experiment designTemperature & injectionCoreflood operating condition and injected treatment.+

Test temperature

°C

Coreflood temperature.Range 20–95

Temperature band

Position in the paired study design, not a universal threshold.

Injection fluid

Treatment fluid entering the core.

Nanoemulsion concentration

wt%

Must be zero for formation water.Range 0–0.1

Injection rate

mL/min

Volumetric injection rate.Range 0.05–0.2

Injection volume

Treatment volume expressed as pore volumes.

Ready for inferenceModels remain sealed during every request.

Behind the result

One request, two model decisions.

Every scenario passes through the same guarded service boundary. The web layer displays results; it never fits, tunes or alters a model.

01
Validate
Ranges, categorical values and cross-field physics checks.
02
Transform
Each pipeline applies only its frozen train-fitted preprocessing.
03
Infer
XGBoost predicts recovery; Logistic Regression predicts damage class.
04
Explain
Probabilities, derived values, provenance and warnings are returned together.