Geosteering geologists manually interpret well logs and directional survey information to update their estimates of the subsurface. The manual interpretation may be heuristic, time consuming, and bear a high cognitive load. Drilling is so fast today, the well may have already drilled through the section in question before a problem is identified and a solution found.
Automating the geosteering process may create more precisely located wellbores, and more accurate wellbores may yield higher production. Drilling automation tools optimize the driller's path to get back to a well plan, but they offer no assistance in updating that well plan to adapt to unexpected geology. We have designed a probabilistic model over the well path and geology, and used a Bayesian network to update the earth model continuously. The result is a posterior probability distribution for the well path and the formation structure, yielding more accurate wellbore position, better target structural information, and the ability to inform geoscientists about geologic changes in real time.
In this work, we used the Bayesian network to interpret three legacy wells. Two of the wells had a straightforward interpretation and were well placed in the target formation. The third well wavered along the target formation and could have pierced it. Using the Bayesian network, we predict trajectories of these three wells along with the geologic target geometry. These estimates were obtained autonomously, without intervention, and were the equivalent of geosteering interpretation.
A small, even uncorrelated, error in inclination measurement at each survey can integrate up to hundreds of feet of well position error at the end of a long lateral. Firm control over well position is important to an accurate estimate of the geology. Additionally, the well spacing in many of the fields allows a more rigorous "correlation" of the shale as it is being drilled through rather than an "eyeball" estimation of the well within the formation.