Traditionally, history-matching has been a trial-and-error process, where various model parameters are repeatedly adjusted over successive simulation runs so that the difference, or mismatch, between simulated and observed values is minimized. In general, the relationship between such mismatch and the values of the parameters is quite complex. A great deal of expertise is required to figure out patterns that capture such relationship intuitively. Moreover, the number of parameters is so immense that the engineer has to restrict the attention to a small set of parameters, possibly missing influential ones.
In this work, we rigorously compute the derivatives of the mismatch with respect to each parameter, taking advantage of all pieces of simulator outputs, as well as our detailed knowledge of the fluid flow equations implemented within the reservoir fluid-flow simulator. With this new approach, a single simulation run followed by a derivatives calculation session is sufficient to detect how each parameter affects the mismatch, and hence, to decide how (or whether) to change each parameter to improve the match.
This approach has been successfully applied to history-match a three-phase case of a reservoir in North America with 45 years of production history and more than 40 wells. We started from a case that has already been matched through conventional means. We show that, based on our technique, very significant improvements can be achieved beyond the point where conventional means have been exhausted using only a small number of simulation runs.