SPE Reserv. Eval. Eng.

A machine-learning methodology using domain-knowledge constraints for well-data integration and well-production prediction

View publication


In this paper, we propose a machine-learning methodology using domain-knowledge constraints for well-data integration, prior/expertknowledge incorporation, and sweet-spot identification. Such methodology enables the analysis of the effects of the main variables involved in production prediction and the evaluation of what-if scenarios of production prediction within geological zones. This methodology will allow streamlining the process of data integration, analytics and machine learning for better decisions, saving time, and helping geologists and reservoir and completion engineers in the task of sweet-spot identification and completion design. We tested the proposed methodology with production, completions, and petrophysical data from a field within a geological target zone. For instance, using local Kriging, we estimated gamma ray features from gamma ray measurements from vertical and horizontal wells, and we integrated those features into the production- and completion-well data, generating an integrated data set for machine-learning modeling. Besides the usual black-box machine-learning models, we used generalized additive models (GAMs) and shape-constraint additive models (SCAMs) for predictive modeling. Those models permit the incorporation of prior/expert knowledge in terms of interaction terms and mathematical constraints on the shape of the effect of the covariates, such as petrophysical and completion parameters, resulting in greater accuracy and interpretability of the predicted production vs. classical black-box machine-learning modeling. We also defined hypothetical what-if scenarios of oil production, such as by estimating the empirical distributions of production estimates using hypothetical settings for completions within the region of interest.