Optimizing CO₂-Enhanced Oil Recovery Through Machine Learning

Producing more oil from an underground reservoir is more than a matter of increasing injection rates. Changes in permeability, pressure, and production conditions can significantly all have an impact on how the reservoir responds to CO and water injection. This project uses machine learning and data science to look at these relationships and to better understand the impact of CO-EOR decisions on both oil production and economic profits. I built a physics-based digital twin based on the SPE10 Model 2 reservoir benchmark and used it to generate 1,000 simulated reservoir scenarios. In each of the scenarios, I varied the permeability, CO injection rate, water injection rate, production rate, and initial pressure. I then looked at the resulting change in oil production, reservoir pressure, and calculated profit, then trained a Random Forest model to see which conditions were most strongly associated with these outcomes. The simulations show how computational modeling can be used to test a number of possible reservoir conditions without being bound to a single operating scenario. More generally, the project explores how machine learning can add onto reservoir modeling by taking the myriad of interactions between the variables and turning them into patterns. Future work will be used to include even a more heterogenous version and testing the model on even more datasets, other than just the SPE10 benchmark.