Development of machine learning predictive models for history matching tight gas carbonate reservoir production profiles

History matching (HM) is an important task that is performed once a reliable reservoir model is constructed during numerical reservoir simulation studies. This HM process requires modification of uncertain reservoir parameters in order to match historical production data. As a result of these modifications, HM has been a long-standing industrial challenge in terms of computational cost and time consumption and furthermore, it requires much experience from the modeler. The tight gas production profiles of a damaged formation used in this study have been difficult to match due to the complex flow mechanism, computational expensiveness and minimal interaction between production wells. In addition, incorporating complex fracture networks due to stress sensitivity into the field’s reservoir model for prediction is also a major challenge. Therefore, this paper proposes supervised machine learning predictive data analytics techniques of multivariate adaptive regression splines (MARS), stochastic gradient boosting (SGB) and a generalized regression neural network (GRNN) to history match the target field’s well productivity profiles. The obtained results from the machine learning simulation techniques indicate that, unlike the full-physics numerical simulator HM approaches, the practical turnaround CPU time required for obtaining the target field’s history match trained models were less than a minute. Comparatively, in terms of the developed predictive models’ statistical performance measures using root mean squared error (RMSE) and coefficient of determination (R2), the testing of the MARS and SGB history match models exhibit superiority over the GRNN model. The trained MARS and SGB models being the best from the testing phase were validated with a surrounding production well profile for blind HM which gave impressive quality match results for the MARS and SGB models. The obtained validated results were then compared to the random forest (RF) ensemble method with MARS (RMSE = 0.0063, R2 = 0.9998) and SGB (RMSE = 0.0478, R2 = 0.9931) demonstrating the advantages of our proposed models as compared to the RF method (RMSE = 0.0502, R2 = 0.9928). Also, the models developed in this study do not require re-training when updating with newly available datasets for HM. In all, the MARS, SGB and RF representative trained models will serve as robust alternative reservoir management and planning tools to supplement traditional numerical simulator HM in order to minimize substantial risks and uncertainties of the target field.