Real-time reservoir management, optimization and uncertainty assessment requires an effective assimilation of all available data into multiple reservoir models to be able to generate accurate forecasts and ultimately, fast and reliable decisions. As more data becomes available, more ambitious reservoir simulation problems can be addressed and innovative opportunities for understanding the reservoir behavior are now at our disposal. In particular, the extraction of knowledge and physical insights from vast amounts of data is providing powerful mechanisms to tailor the generation of non-intrusive reduced order models and specialized analytics to capture solution trends that may not be otherwise amenable to assimilate into current complex simulation systems. This presentation will provide a panoramic view of a few promising advances that have been taking place in data mining and machine learning to enable the development of non-intrusive data-driven surrogate models. Provided that there is a representative set of training reservoir scenarios, input parameters, and outputs can be projected into a much lower dimension that retains the main features contained in the simulation system. Once the complexity of the simulation is reduced, predictions can be efficiently carried out with the aid of neural or kernel models. Cases will be shown to illustrate the enormous potentials that this class of data-driven surrogate models offer to tackle current and forthcoming reservoir engineering challenges within the Oil& Gas business.