Ensemble-Based Assisted History Match Using Machine Learning and Optimization Model

Abstract

Abstract History matching is a key task during reservoir modeling for making feasible operational decisions in field development. However, classic approaches for history matching are expensive in terms of time and computational requirements. The main objective of this project is to accelerate the HM process using an innovative approach that includes an additional step that allows the integration of machine learning with a numerical simulation to deliver efficient and superior-quality results for the history match. This solution comprises an ML model and an optimization process that uses as input the first simulation results coming from sampling with a Monte Carlo algorithm to train the model. This allows for finding the optimized distribution and ranges of the variables that help the most in mismatch reduction. The second stage runs a set of numerical simulations based on the Monte Carlo method with the recommended ranges and distribution; the new simulations results feed back into the ML algorithm, which returns more narrow ranges and distribution; this process continues until an acceptable solution is found. As an application example, three iterations and 475 simulation runs were needed to achieve a desirable solution. This result was compared with the traditional calibration technology using the same number of simulation runs, and the new approach showed better overall results for the mismatch error. Comparing the total time used in both cases, the solution with ML is more efficient, taking 2.5 hours over 10 hours with the classic approach (four times faster), and it is delivering better results than the traditional solution in terms of accuracy. The main output of this solution is an ensemble of matched models providing a robust description of subsurface uncertainty, which means a high degree of predictability in the forecast scenarios for quantifying the associated risk for the field development plan (FDP). This offers the capability to estimate the chance of meeting production volumes above the threshold for economic success after established years of forecasting and making a high-fidelity decision in future development plans. Another important characteristic is the efficiency in the number of runs needed to find a solution and the time invested, and finally, the API developed is created with a user-friendly interface with well-defined steps. The implementation of new methodologies and the way of integrating numerical simulations with digital solutions such as machine learning on a scalable compute environment demonstrate to have effectiveness due to the framework for continuous improvement reducing the error of the reservoir model when new data comes in. An ensemble-based approach for reservoir modeling and history matching helps to ensure success in making high-fidelity decisions in future development plans.