Search for a command to run...
Abstract This paper presents a machine learning (ML)-based workflow for early-life Estimated Ultimate Recovery (EUR) forecasting, integrated with production trend diagnostics to support development decision-making in a mature offshore oil field. The methodology targets newly drilled wells and short production histories where classical decline curve analysis (DCA) remains unreliable due to limited data, reservoir heterogeneity, and evolving operating conditions. The approach was developed and validated using production and reservoir performance data from more than 50 wells across two geologically distinct sectors of the field, characterized by contrasting net pay thickness, permeability, and fluid properties. Input features were derived from early-time production behavior, including average oil rate, water cut, gas-oil ratio, and decline-related metrics evaluated over multiple time windows ranging from the first three months to three years of production. Supervised learning algorithms—including linear regression, ridge regression, and linear support vector machines—were trained using K-fold cross-validation to ensure robustness and generalizability under data-constrained conditions. Separate models were constructed for each reservoir sector to explicitly capture geological and performance variability. Among the tested algorithms, linear-based models demonstrated the most consistent and reliable performance. EUR prediction uncertainty was approximately ±45% when using only the first three months of production history and improved to approximately ±17% once two years of production data became available. These uncertainty ranges are consistent with early-life reserve estimation expectations commonly encountered in technical evaluations. Model performance was stable across both reservoir sectors, indicating that the workflow is transferable within heterogeneous field settings when appropriate zonal segmentation is applied. Beyond standalone EUR forecasting, the proposed workflow provides a structured mechanism for asset teams to continuously evaluate drilling outcomes at the area and reservoir-sector level and to integrate post-drilling well performance into the evolving geological and reservoir understanding of the field. By systematically comparing planned well productivity and pre-drill EUR expectations against ML-predicted EUR and observed early-life production behavior, deviations can be identified at a very early stage of well life. These deviations often reflect underlying reservoir heterogeneity, local variations in connectivity, fluid distribution, or completion effectiveness that may not have been fully captured during pre-drill subsurface characterization. The proposed methodology offers a pragmatic and interpretable alternative to complex black-box ML solutions, with the primary objective of de-risking development decisions and enhancing field planning in areas lacking established decline trends. The workflow deliberately relies on lightweight, statistically grounded linear models.