Page 402 - Intelligent Digital Oil And Gas Fields

The Future Digital Oil Field 339

The fourth component of the closed loop is technology to efficiently
optimize the decision variables using the updated models that relate these
decision variables to multiple objective functions. The techniques used here
are similar to those used for model updating, such as gradient-based and sto-
chastic optimization methods. Recent approaches, such as mimetic methods
that combine the benefits of both gradient-based and stochastic approaches,
can help reach solutions. Further, combined with scalable parallel cloud
computing and reduced-order models, the optimization problem has
become practically solvable in a closed-loop framework.
Sarma emphasizes that the final and probably most neglected component
of the closed-loop system is a system to enable users—from managers to
engineers to operators—to easily analyze, collaborate, and implement deci-
sions recommended by the closed-loop system. Oil producers and service
companies are notorious for delivering software with very poor and non-
intuitive user interfaces, and most such software are static in that they are
not connected to live data streams, and are certainly not collaborative, that
is to enable multiple disciplines to interact effectively. A software application
enabling closed-loop optimization has to be designed from the ground up
integrating the best design practices and tools from the software industry
to bring this archaic part of the closed-loop system to modern times.
Sarma elaborates that physics-based reservoir simulation and data-driven
machine learning offer complementary strengths. An ideal predictive model
would combine the speed and flexibility of machine learning with the pre-
dictive accuracy of reservoir simulation, at the right scale, so that operators
can integrate data in real time to quantitatively optimize reservoir manage-
ment decisions continuously. To this end, data physics models merge mod-
ern data science and the physics of reservoir simulation seamlessly. Data
physics models, like machine learning models, require only days to set up
and can be run in real time. Additionally, since they include all the same
physics as a reservoir simulation, they offer excellent long-term predictive
capability even when historical data are sparse, missing, or noisy.
Data physics models are optimized for speed and run orders of magnitude
faster. This speed permits repeated comparison and tuning to the underlying
history of production and consequently permits statistically quantifiable
comparative prediction performance between many alternative scenarios
in service of quantitative optimization.
Given this definition of data physics, while this approach is fundamen-
tally different from either traditional machine learning or simulation, on its
own and considering only the science behind the technology, it is not