Abstract
In the past two decades, researchers have made remarkable progress in accelerating flash calculation, which is very useful in a variety of engineering processes. In this paper, general phase splitting problem statements and flash calculation procedures using the Successive Substitution Method are reviewed, while the main shortages are pointed out. Two acceleration methods, Newton's method and the Sparse Grids Method are presented afterwards as a comparison with the deep learning model proposed in this paper. A detailed introduction from artificial neural networks to deep learning methods is provided here with the authors' own remarks. Factors in the deep learning model are investigated to show their effect on the final result. A selected model based on that has been used in a flash calculation predictor with comparison with other methods mentioned above. It is shown that results from the optimized deep learning model meet the experimental data well with the shortest CPU time. More comparison with experimental data has been conducted to show the robustness of our model.
Introduction
In industries such as petroleum engineering, carbon sequestration, and chemical manufacturing, simulating the behavior of complex mixtures under varying temperatures and pressures is critical. This process, known as Vapor-Liquid Equilibrium (VLE) or "flash calculations," traditionally relies on Equations of State (EOS) and iterative algorithms like the Successive Substitution Method (SSM) or Newton’s method.
However, these traditional methods face significant challenges. First, they are computationally expensive; in large-scale reservoir simulations, these calculations occur millions of times, consuming the majority of CPU resources. Second, they often suffer from stability issues, failing to converge or producing unphysical solutions near critical points or within complex mixtures.
Methodology
To address these limitations, this research proposes a Deep Neural Network (DNN) approach to act as a direct predictor for flash calculations, bypassing the need for iterative loops.
Model Architecture
The researchers developed a fully connected network consisting of five hidden layers with 100 nodes per layer. The model is designed to take eight physical parameters as inputs: the critical pressure ($P_c$), critical temperature ($T_c$), and acentric factor ($\omega$) for each component in a binary mixture, along with the system's overall temperature ($T$) and pressure ($P$). The outputs are the mole fractions of the liquid ($X$) and vapor ($Y$) phases.
Each node within the network processes inputs through weights and biases, followed by an activation function to produce an output.
Training Techniques
To optimize the model for a relatively small and noisy dataset, several modern deep learning techniques were employed:
- Xavier Initialization: Used to maintain signal stability during training.
- Regularization: Dropout and Weight Decay (L2) were implemented to prevent overfitting.
- Batch Normalization: This technique accelerated convergence by approximately 10x.
- Activation Functions: After testing 12 different functions, ReLU and ELU were found to offer the best balance of accuracy and stability.
Data and Training
The model was trained and validated using 1,332 data points from the Korea Thermophysical Properties Data Bank (KDB), focusing on binary mixtures of alkanes ($C_1$ to $C_7$). The experimental data contains an inherent mean relative error of about 20%, which defines the baseline for the model's accuracy.
Results and Discussion
The deep learning approach demonstrated significant advantages over traditional benchmarks, including SSM, Newton’s method, and the Sparse Grids Method.
- Massive Acceleration: The DL model achieved a staggering 2051.6x speedup over the traditional SSM and was roughly 1000 times faster than Newton’s method.
- Stability: Unlike iterative solvers that require precise initial guesses, the neural network provides instantaneous results for any input within its trained range, effectively smoothing out thermodynamic instabilities.
- Accuracy: The model's predictions aligned well with experimental data, often maintaining relative errors below 5%. It outperformed SSM in complex scenarios where traditional methods typically fail.
- Complexity Insights: Interestingly, increasing network depth beyond five layers led to diminishing returns and overfitting, suggesting that for physical regressions like VLE, moderate complexity is optimal.
Conclusion
While the current model is limited by data scarcity and is specific to binary alkane mixtures, it proves that deep learning can drastically accelerate flash calculations. Future work will focus on generalizing the model to more complex mixtures and exploring hybrid approaches that use DL to provide "first guesses" for rigorous EOS solvers, combining AI's speed with classical thermodynamic consistency.
Summary updated