Physics + AI Key to Production Optimization

Ferdinand Hingerl, PhD, Lead Data Scientist

We think that machine learning/artificial intelligence (ML/AI) skills will be part of each production engineer’s toolbox in the future, but differ on the clear distinction between AI vs. physics and the validity of hybrid approaches as presented in a recent article.

Several recent studies have demonstrated clearly that a combined AI+physics approach can merge the upsides of each world, which allows us to tackle engineering and science problems more efficiently with increasingly better accuracy and predictability. The article mentions reservoir simulation as an example to support his view, but we believe reservoir simulation is actually a case that will benefit substantially from a hybrid approach.

Hybrid models have become a traditional tool in the engineer’s toolbox. However, the article seems to disparage them as an unintentional consequence of engineers’ over-simplified understanding of ML. We anticipate that data-driven models will eventually learn the laws of physics from data alone and be able to transfer that knowledge from one problem to another. However, in the current practical solution of engineering-related problems, a balance between physics-based and data-driven components working together towards an improved solution is much more feasible. Hybrid models are traditional, intentional engineering approaches to model development that go beyond the two classes of ML applications in the article.

There have been efforts to solve well production optimization problems in mature waterfloods for years by purely using AI, without success. If one does not have a full geologic representation and understanding of basic flow relationships and flow paths in the subsurface, a purely datadriven solution will most likely not be able to provide solutions to even the most basic engineering problems, such as determining the optimal location for the next infill well or predicting remaining recoverable oil in place.

These previous efforts only broke into the market once experienced reservoir simulation engineers proceeded hybridizing their physics knowledge with the existing data driven approaches. A full CFD-based reservoir simulation is subject to a large number of simplifying assumptions (as outlined by the article), is computationally very costly, extremely demanding to set up (typically around two years for a full reservoir including history matching, with several geologists and reservoir engineers involved), and requires a huge amount of data that will be re-fitted in the end to match observed production values (e.g., relative permeability curves, transmissibilities, etc.). Therefore, while this technology is tested and proven, it also is tremendously challenging to use, provides only approximate results, is expensive, and likely will not be the way forward on its own, in our opinion.

We believe the way forward in reservoir engineering/simulation should be based on our system understanding (i.e., physics), which then should be merged with our new access to modern AI algorithms to account for noise and relationships, which are based on physics but too complex to be captured by equations. This amalgamation can happen in many different ways, such as physics-based engineered features for AI, physics-informed (non-linear) constraints and bounds on AI parameters (input and output), etc, but in the end, we will need to come up with smart ideas to merge the two worlds. Only then can we extrapolate with confidence, trust our predictions, and reduce the amount of data required for AI, while at the same time churning out predictions at much higher speed, increasing confidence, and reducing computational overhead as a purely physics-driven approach would allow us to provide.