Integration of Deep-Learning-Based Flash Calculation Model to Reservoir Simulator

Abstract

Flash calculation is an essential step in compositional reservoir simulation. However, it consumes a significant part of the simulation process, leading to long runtimes that may jeopardize on-time decisions. This is especially obvious in large reservoirs with many wells. In this paper we describe the use of a machine-learning- (ML) based flash-calculation model as a novel approach for novel thermodynamics via this ML framework to potentially accelerate compositional reservoir simulation. The hybrid compositional simulation protocol uses an artificial-intelligence- (AI) based flash model as an alternative to a thermodynamic-based phase behavior of hydrocarbon fluid, while fluid-flow equations in the porous medium are handled using a conventional approach. The ML model capable of performing accurate flash calculations is integrated into a reservoir simulator. Because flash calculations are time consuming, this can lead to instability issues; using the ML algorithm to replace this step results in a faster runtime and enhanced stability. The initial stage in training ML models consists of creating a synthetic flash data set with a wide range of composition and pressure. An automated workflow is developed to build a large flash data set that mimics the fluid behavior and pressure depletion in the reservoir using one or more fluid samples in a large pressure-volume-temperature (PVT) database. For each sample, a customized equation of state (EOS) is built based on which constant volume depletion (CVD) or differential liberation (DL) is modeled with prescribed pressure steps. For each pressure step, a constant composition expansion (CCE) is modeled for the hydrocarbon liquid with, in turn, prescribed pressure steps. For each of the CVD and multiple CCEs steps, flash calculation is performed and stored to build the synthetic database. Using the automatically generated flash data set, ML models were trained to predict the flash outputs using feed composition and pressure. The trained ML models are then integrated with the reservoir simulator to replace the conventional flash calculations by the ML-flash calculation model, which results in a faster runtime and enhanced stability. We applied the proposed algorithms on an extensive corporate-wide database. Flash results were predicted using the ML algorithm while preceded by a stability check that is performed using another ML model tapping into the exceptionally large PVT database. Several ML models were tested, and results were analyzed to select the most optimal one leading to the least error. We present the ML-based stability check and flash results together with results illustrating the performance of the reservoir simulator with integrated AI-based flash, as well as a comparison to conventional flash calculation. We are presenting a comprehensive AI-based stability check and flash calculation module as a fully reliable alternative to thermodynamic-based phase behavior modeling of hydrocarbon fluids and, consequently, a full integration to an industry-standard reservoir simulator.