In a project funded by the U.S. Department of Energy, researchers are using machine learning to teach algorithms to create fast and accurate models of key features in underground environments.
Simple, simulated oil and gas reservoirs were used initially to teach the software to sift through borehole pressure and flowrate data. The algorithms had to make decisions and predictions of the porosity and permeability of subsurface material. Correct answers, judged against known results, were rewarded. Researchers found that the algorithms quickly learned to predict accurately by maximizing their accrued rewards.
This reinforcement-based success should soon allow these algorithms to handle more complex simulations, potentially leading to the automated prediction of oil and gas reserves and even groundwater systems or seismic hazards.