Enhanced Uncertainty Quantification of Reservoir Simulation Models Through Synergistic Integration of Particle Swarm Optimisation with the Ensemble Kalman Smoother

Abstract The objective of this paper is to put into perspective the use of direct search algorithms like Particle Swarm Optimisation (PSO) versus data assimilation techniques like the Ensemble Kalman Smoother (EnKS) versus proxy modelling in reservoir simulation and in which context they should be used. Their performance on a benchmark reservoir simulation model (PUNQ-S3 model) is compared, and then a hybridised approach is taken to capitalise on their strengths to improve the final history matches. The procedure followed here is to first run the particle swarm optimiser from scratch on the benchmark problem, thereafter to run the EnKS on the same problem, and finally to create a proxy model from a set of designed experiments which would then be optimised on. The PSO and EnKS experimental tests use the same number of total reservoir simulation runs, whereas the proxy modelling method has only a quarter of the aforementioned runs. The performance of each one is then compared against the other, and their strengths and weaknesses with respect to global optimisation and local optimisation judged. Finally, a hybrid approach of first using the PSO to do a global search, and then the EnKS to do a local search to improve the best match runs of the PSO is conducted. The PSO performs much better performance at history matching the case study reservoir model and shows its global optimisation ability to scan through the hyperparameter space, whereas the EnKS struggles to locate the global optima in the search space. Using the EnKS on the best PSO matches showed a further refinement of the particle swarm history matches leading to slightly improved overall matches. This is the first study where both particle swarm optimisation and EnKSs are used synergistically. Particle swarm optimisation is best used at the outset of a history matching project to allow for a wide probing of the search space, and this coupled with using the EnKS to do local optimisation at the end of the particle swarm optimisation runs can lead to better overall matches.