[Paper Review] Acceleration of the NVT-flash calculation for multicomponent mixtures using deep neural network models

Yiteng Li, Tao Zhang+1 more|arXiv (Cornell University)|Jan 27, 2019

Reservoir Engineering and Simulation Methods51 references4 citations

TL;DR

This study proposes a deep neural network model to accelerate NVT-flash calculations for multicomponent hydrocarbon mixtures by learning from data generated via a dynamic model, achieving up to 244× speedup over conventional iterative methods while maintaining high accuracy and eliminating the need for separate stability testing.

ABSTRACT

Phase equilibrium calculation, also known as flash calculation, has been extensively applied in petroleum engineering, not only as a standalone application for separation process but also an integral component of compositional reservoir simulation. It is of vital importance to accelerate flash calculation without much compromise in accuracy and reliability, turning it into an active research topic in the last two decades. In this study, we establish a deep neural network model to approximate the iterative NVT-flash calculation. A dynamic model designed for NVT flash problems is iteratively solved to produce data for training the neural network. In order to test the model's capacity to handle complex fluid mixtures, three real reservoir fluids are investigated, including one Bakken oil and two Eagle Ford oils. Compared to previous studies that follow the conventional flash framework in which stability testing precedes phase splitting calculation, we incorporate stability test and phase split calculation together and accomplish both two steps by a single deep learning model. The trained model is able to identify the single vapor, single liquid and vapor-liquid state under the subcritical region of the investigated fluids. A number of examples are presented to show the accuracy and efficiency of the proposed deep neural network. It is found that the trained model makes predictions at most 244 times faster than the iterative flash calculation under the given cases. Even though training a multi-level network model does take a large amount of time that is comparable to the computational time of flash calculations, the one-time offline training process gives the deep learning model great potential to speed up compositional reservoir simulation.

Motivation & Objective

• Address the computational bottleneck of equation-of-state-based flash calculations in compositional reservoir simulation.
• Overcome the inefficiency of conventional flash frameworks that require separate stability testing and phase splitting.
• Develop a deep learning model capable of predicting phase behavior—including single-phase and vapor-liquid equilibrium—directly from P, T, and composition.
• Balance model accuracy and training efficiency by optimizing data size and network architecture.
• Enable reusable, fast predictions for large-scale compositional simulations through one-time offline training.

Proposed method

• Use a dynamic model based on mole and volume evolution equations to iteratively simulate NVT-flash processes and generate training data.
• Train a deep neural network to map input conditions (P, T, feed composition) to phase compositions and mole fractions.
• Incorporate a reformulated loss function and dropout regularization to reduce overfitting and improve generalization.
• Apply batch normalization to accelerate training convergence and improve optimization stability.
• Design a multi-layer network architecture informed by prior research on deep learning for flash approximation.
• Train the model on three real reservoir fluids: one Bakken oil and two Eagle Ford oils with 5, 8, and 14 components respectively.

Experimental results

Research questions

• RQ1 Can a deep neural network accurately approximate the phase behavior of complex multicomponent hydrocarbon mixtures in NVT-flash calculations?
• RQ2 To what extent can a deep learning model eliminate the need for separate stability testing in flash calculations?
• RQ3 How does the size of the training dataset affect the accuracy and efficiency of the deep learning model?
• RQ4 What is the trade-off between training cost and prediction speed in the proposed deep learning framework?
• RQ5 Can the trained model generalize across different reservoir fluid compositions and phase states (single-phase, vapor-liquid)?

Key findings

• The deep neural network model achieves up to 244 times faster prediction than conventional iterative flash calculations under the tested conditions.
• The model successfully identifies single-phase (vapor or liquid) and two-phase (vapor-liquid) states without requiring a separate stability test.
• With a training data size of 201×201, the model achieves good accuracy with only minor fluctuations compared to the 301×301 baseline, indicating diminishing returns beyond moderate data volumes.
• The model maintains high accuracy across diverse components, including C1, C2, C5–C6, C7+, and C13+ in Bakken and Eagle Ford oils, as validated by comparison with iterative flash results.
• Although training time is comparable to iterative flash computation, the one-time offline training enables repeated, ultra-fast inference, making it highly suitable for large-scale compositional simulation.
• The approach demonstrates strong potential for accelerating phase equilibrium calculations in complex systems, including future extensions to confined environments like nanopores.