Deep Neural Network for Compositional Reservoir Phase Behavior Modeling

The DNN model is trained offline using conventional phase equilibrium calculations to learn the mapping between input parameters and phase behavior outputs. This preliminary training phase stores the knowledge of accurate phase calculations, enabling the model to make rapid predictions during reservoir simulation without performing full phase equilibrium calculations at each time step, thus resolving the contradiction between accuracy and computational speed.

The DNN model creates a computational copy of the conventional phase equilibrium calculation process. Instead of repeatedly executing the complex conventional calculations, the trained neural network model replicates their functionality with simplified computational operations, maintaining accuracy while dramatically reducing the computational time required for phase behavior determination during simulation.

The computationally intensive phase equilibrium calculations are performed in advance during the offline training phase to build the DNN model. This preliminary action transfers the computational burden from the online simulation phase to the offline training phase, enabling accurate phase behavior determination during simulation with minimal computing resource consumption at each time step.

The DNN model creates a lightweight computational copy of the complex phase equilibrium calculation process. The trained neural network replicates the functionality of conventional phase calculations using simple matrix operations and activation functions, maintaining accuracy while reducing computing resource consumption by several orders of magnitude during reservoir simulation execution.

The DNN model provides a computationally efficient copy of phase equilibrium calculations that can be executed rapidly at each grid cell and time step. This enables high-resolution simulations with increased spatial and temporal discretization without proportionally increasing computational costs, as the neural network model requires minimal computational resources compared to conventional phase calculations.

The DNN model changes the computational parameters from complex iterative phase equilibrium calculations to simple neural network forward propagation operations. This parameter transformation enables the model to handle high-resolution simulations efficiently, as the neural network operations scale linearly with the number of grid cells and time steps rather than requiring iterative convergence at each location.