01 / EXPLORE
Your dataset, with provenance.
Upload CSV observations or explore a synthetic demonstration. Files and models stay in this browser tab; this app does not upload their contents or save them across reloads.
A user declaration does not validate the data. Changing provenance resets trained models. Use anonymized well IDs and a consistent liquid-volume basis.
CSV schema, units and supported ranges
Required for every row: well_id,method,reservoir_mpa,depth_m,density,whp_mpa,pi,friction,liquid_rate. Method must be gas, esp or rod. Add the relevant equipment columns below. Blank irrelevant equipment fields are allowed. These ranges define this educational app’s supported domain, not equipment specifications.
Gas requires injection. ESP requires frequency and stages. Rod requires diameter, stroke, spm and fillage. liquid_rate is the supervised target in m³/day. Missing or invalid required values cause row rejection. Exact duplicate observations within a well are removed.
Exploratory statistics
Summary covers accepted rows for the selected method. Pearson correlation is an association with liquid rate, not causation.
02 / LEARN
Train an explainable regression model.
One ridge model per lift method. A fixed seed assigns about 20% of well IDs to the test set. No well appears in both sets. Scaling uses training data only. Minimum: 30 records and 5 wells per method.
Fixed regularization α = 1. The test set is for evaluation, not tuning. Temporal forecasting requires a separate time-based validation design. Negative rate predictions are clipped to zero, and clipping is reported.
03 / INVESTIGATE
Lift Advisor
Assess the selected method using the current well and infrastructure inputs. Findings explain the reason and next check. A rate comparison alone cannot identify the best lift method.
Ask the advisor
Guided questions use the selected method and current form values. This is a rule-based conversation, with no external AI service.
Model card and boundaries
The ML target is liquid production rate, not “best lift method.” Training on synthetic data learns an approximation to LiftLab’s teaching equations. Measured-data models still need domain review and independent validation. Coefficients show standardized associations and can be unstable when inputs are correlated. No uncertainty intervals, causal effects, economics or automatic field control are provided.
The advisor flags user-reported infrastructure constraints and limits in the teaching model. It does not diagnose gas lock, tubing failure or other field faults from rate alone. Unknown inputs stay unknown. Proposed changes are investigation steps, not safe operating setpoints.
References: Ridge regression, data leakage, grouped validation, artificial-lift principles. The browser implementation is custom JavaScript; it does not run scikit-learn.