Intelligent Performance Assistant
Run-time performance of a reservoir simulator is significantly impacted by the selection of the linear solver preconditioner, iterative method, and their adjustable parameters. The choice of the best solver algorithm and its optimal parameters is a difficult problem that even experienced simulator users cannot adequately solve by themselves. The typical user action is to use the default solver settings or a small perturbation of them that are frequently far from optimal and consequently the performance may deteriorate.
Working with our customer, ExxonMobil Upstream Research Company, we developed an adaptive control, on-line system that optimizes simulator performance by dynamically adjusting the solver parameters during simulation. The systems starts with a large set of parameters and quickly choose the best combinations. These parameters are continuously adapted during the simulation using the solver’s runtime performance measurements (e.g. solver CPU time) to guide the search.
This machine learning-based software system, called the Intelligent Performance Assistant (IPA), works with SparSol, the sparse linear solver also developed by NeurOK Software for ExxonMobil. The pair have been integrated into ExxonMobil’s proprietary reservoir simulator and deployed worldwide.
The system can handle a large number of combinations of solver parameters, currently in the order of 108, and consistently improves run time performance of real simulation models, frequently by 30% or more, compared to the performance with the default solver settings.
IPA also includes a persistent memory of solver performance statistics. The runtime statistics from these individual runs is gathered, processed using data mining techniques and integrated back into the IPA system, thus allowing for continuous improvement.