Forecasting Water Breakthrough with a Reservoir Intelligence Platform

    Our reservoir intelligence platform forecasts when a gas well will start producing water, and how much, years before it happens. Built entirely on data the operator already collects, it combines survival analysis, three-month early warning surveillance, and per-well volume forecasting, validated against Mari Energies own historical breakthrough records.

    image

    38%

    production uplift

    75%

    breakthrough events caught years before

    3050%

    reduction in facilities planning uncertainity

    100%

    audit trail coverage with ESG reporting

    Business Overview and Strategic Direction

    EXECUTIVE SUMMARY

    Mari Energies, a major gas exploration and production operator, needed a better way to predict water breakthrough in a carbonate reservoir before it disrupted production and increased handling, treatment, disposal, and intervention costs. Conventional reservoir studies were too slow and infrequent for timely action.

    Stixor developed a reservoir intelligence platform that uses existing production, completion, and petrophysical data to forecast breakthrough timing and water volumes. Validated against historical well data, the platform gives engineering teams months to years of lead time to plan workovers, prepare water handling capacity, and manage disposal requirements.

    Executive Summary Image

    Operational Challenges in Gas Well Water Management

    Forecasting water breakthrough on a carbonate reservoir demands long-horizon accuracy, honest uncertainty, and engineering trust. Key challenges:

    PROBLEM STATEMENT

    Late Discovery

    Breakthrough was confirmed only after water volumes rose, leaving no lead time to plan a workover or size facilities.

    card-image

    Slow Simulation Cycles

    Full reservoir simulation studies are expensive and rerun infrequently, so the operating picture ages every month.

    card-image

    No New Data Available

    Any solution had to work from records already collected. New instrumentation or acquisition programmes were out of scope.

    card-image

    Engineering Trust

    A forecast that fires false alarms, hides its misses, or implies false precision does not survive contact with a reservoir engineering team.

    card-image

    Three-Module Reservoir Intelligence Platform

    One platform connects long term breakthrough forecasting, near term risk detection, and water volume planning, giving operators a clearer view of what is coming and time to act.

    Our Solution

    1.

    Long-Horizon Breakthrough Timing

    Cox survival analysis, benchmarked against multiple statistical and ML models, forecasts breakthrough timing from geological and production data. P10/P50/P90 ranges give engineers a practical view of expected timing and uncertainty.

    2.

    Three-Month Early Warning

    The platform combines WGR trends, threshold momentum, next quarter WGR forecasts, and pressure decline through gradient boosting to identify wells approaching breakthrough with actionable three month lead time.

    3.

    Water Volume Forecasting

    Per well time series forecasting uses gas rate and wellhead pressure to estimate future water volumes. Monthly updates help operators plan water handling, treatment, disposal, and potential interventions.

    01

    01 Data Assessment and Definition

    • Audit of existing production, completion, petrophysical, and water chemistry records
    • Breakthrough threshold and labelling rules agreed with the operator's reservoir engineers
    • Feature engineering reviewed against reservoir physics, not selected on statistics alone

    02

    02 Model Development and Selection

    • Approximately 100 modelling configurations evaluated across feature sets, model families, and label definitions
    • Spud and live model modes built to serve day-one planning and mature-well surveillance
    • Retraining on the enhanced dataset brought average timing error to 1.4 years

    03

    03 Validation and Handover

    • Leave-one-well-out cross-validation: every prediction came from a model that never saw that well
    • Blind test on four wells with water data withheld until forecasts were locked
    • Platform handed over with documented methodology, ranked drivers, and a production roadmap

    Validated Outcomes Across the Full Well Cohort

    IMPACT & RESULTS

    Prediction Performance

    11 of 13

    Event Wells Within P10–P90 Range

    75%

    Breakthroughs Detected 3 Months Early

    67%

    Breakthroughs Detected 1 Month Early

    42%

    Breakthroughs Detected 12 Months Early

    Validation Rigor

    ~100

    Model Configurations Evaluated

    100%

    Wells Reported, Including Poor Performers

    4

    Wells Used for Blind Testing

    1

    Model Trained Without Test-Well Data

    Technology Stack

    TOOLS USED

    Lang Chain

    Lang Chain

    Apache Kafka

    Apache Kafka

    PostgreSQL

    PostgreSQL

    Terraform

    Terraform

    TensorFlow

    TensorFlow

    Python / PyTorch

    Python / PyTorch

    PyTorch

    PyTorch

    Scikit-learn

    Scikit-learn

    icon
    icon

    Discuss Your Enterprise Use Case

    From small to large scale enterprises, we deliver next-gen AI, data engineering, and actionable insights.