Methodology of surrogate models (metamodels) and their prospects for solving petroleum engineering challenges

Abstract

Traditional mathematical modeling tends to set limits to the amount of computational experiments to be conducted with mathematical models in order to optimize solutions. Over recent years, metamodeling – a new, alternative approach – has been rapidly developing with a view to facilitate the solution of optimization challenges. Such models are built upon the ideas of machine learning, when the models’ training set is derived from numerous prototypes of input and output data (results of field and computational runs with varied data types). The models are built to imitate (replace) the physics-based mathematical models and reduce computational time and decision-making period. In this paper, we offer approaches for metamodeling (data models) of various levels of complexity depending on actual tasks. We show that application of metamodels significantly reduces the amount of time and computational resources required for solving a wide range of petroleum engineering challenges, e.g. selection, monitoring and optimization of field development. Metamodels returns results without any loss of quality compared to traditional «physical» models.