Scientific progress over the last decade has been significantly facilitated by the evolution of a new breed of intelligent solutions, characterized by their ability to learn without being explicitly programmed with the governing physics. In the oil and gas industry, machine learning alternatives are becoming increasingly popular, however most solutions within this discipline are still very raw in their conceptualization and application. In this work, three major areas in petroleum engineering are addressed and resolved using machine learning: well placement evaluation and optimization, time-series output prediction, and geological modeling. Simultaneous optimization of well placements and controls is a recurring problem in reservoir management and field development. Because of their high computational expense, reservoir simulators are limited in their applicability to joint optimization procedures requiring many evaluations. Data-driven proxies could provide inexpensive alternatives for approximating reservoir responses, however, geologic complexity of most reservoirs often makes it impossible to model or reproduce the response surface using well location data alone. We propose a machine learning approach in which the feature set is augmented by a connectivity network comprised of pairwise well-to-well connectivities for any potential well configuration. Connectivities are represented by 'diffusive times of flight' of the pressure front, computed using the Fast Marching Method (FMM). The Gradient Boosting Method is then used to build intelligent models for making reservoir-wide predictions such as net present value, given any set of well locations and control values. Accurate prediction of future reservoir performance and well production rates is important for optimizing oil recovery strategies. In the absence of geologic models, this could purely be considered as a time-series analysis problem. The premise of this class of problems is that relationships between input and output sequences can be learned from historical data and used to predict future output. However, because the state of the reservoir changes with time, the value of a future output variable such as production rate also depends on its own history. We introduce a novel scheme to predict reservoir output during recovery processes using Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks. The method was applied to two case studies wherein predictive models were built to forecast well production using historical rate data, yielding satisfactory results. A synthetic demonstration showed that the proposed method outperformed Capacitance Resistance Modeling (CRM) in terms of prediction accuracy. Spatial interpolation and geologic modeling of petrophysical properties are traditionally performed using conventional geostatistical algorithms. The most common techniques include the Sequential Gaussian Simulation (SGS) for continuous variable modeling, and multiple-point simulation (MPS) for facies or categorical variable modeling. These techniques produce adequate results but are prone to subjectivity and could rely heavily on the modeler's intuition. Machine learning techniques provide a more automated alternative for geologic modeling, and have the ability to more accurately predict petrophysical properties outside the data locations. We propose a new hybridized method in which Bayesian Neural Network (BNN) predictions are used as kriging covariates in conjunction with SGS. The hybridized models show improved prediction accuracy in comparison with kriging and SGS, while retaining geological realism and producing exact estimates