IMPROVING PRODUCTION STRATEGIES IN UNCONVENTIONAL OIL AND GAS RESERVOIRS THROUGH MACHINE LEARNING

This research involves the application of supervised, unsupervised, and deep learning ML modeling approaches using empirically-derived well completion, production, and geologic datasets from prominent unconventional O&G plays in the U.S. The anticipated outcome of this work is to provide substantial contribution to the knowledgebase pertinent to O&G field development and reservoir management approaches (transferable to other subsurface applications) founded in data-driven strategies. ML-based models built through this work complete a multitude of tasks, including: 1) Evaluating potential well production response to various hydraulic fracturing completion designs using a gradient boosting ML algorithm; 2) hierarchical ranking of well design and geologic reservoir quality parameters and their associated interactions on production response by assessing parametric importance and partial dependence; 3) deriving well design strategies that maximize production given well placement through optimization; 4) development of time series-based predictive forecasting capability using long-short term memory neural networks that can generalize temporal or sequence-based tendencies in water and associated gas production trends; and, 5) to enable rapid identification of stratigraphic units within a basin using multiclass classification given total vertical depth and spatial positioning. The findings from this work show that ML provides fast, accurate, and cost-effective analytical approaches to a variety of O&G-related functions. These strategies can be used to analyze disparate datasets in innovative ways, provide utility in generating new insights, and may be used in ways to identify improvements over industry benchmarks. They offer robust approaches that can supplement existing reservoir management best-practices and improve the return on investment from field data acquisition