Gas-liquid two-phase flow is seen in various engineering disciplines and understanding of its interfacial structure is a great importance for proper model development. Flow regime maps have been developed by various researchers in the past. However, identification of flow regime involves subjectivity and technical difficulty as the nature of flow regime transition mechanism and criteria are still yet unknown. In the present article, various two-phase flow regime identification methods utilizing machine learning (ML) approach will be reviewed. Two-phase flow features such as two-phase mixture impedance, dynamic force signal, and high-resolution images from high-speed camera were selected for the ML model training and testing. For the machine learning models, artificial neural network (ANN), convolutional neural network (CNN), and convolutional long short-term memory (ConvLSTM) were tested, depending on the feature types. As a result, the ML approach can identify two-phase flow regime with high accuracy. There still remains issues related to ML usage, including hyperparameter selection, and ability to explain its decisions. However, adaptation of ML tools for multiphase flow research fields may provide utmost benefits for efficient signal and image classification approach.
We have recently developed a machine learning method capable of inferring the constitutive equation for the stress of a fluid with memory (e.g., polymeric fluid) from microscopic simulations. For this, we used a Gaussian Process regression scheme to learn the constitutive relation, given as the time-derivative of the stress, as a function of the local stress and velocity strain tensors. Crucially, no assumptions are made regarding the functional form of this relation. We have applied our method to the Hookean dumbbell model, for which the exact analytical constitutive relation (Maxwell equation) is known, in order to validate the approach. Our results are in excellent agreement with the analytical solution, showing that we are able to capture the history dependence of the flow, as well as the elastic effects in the fluid. Compared to full multi-scale simulations, in which the micro and macro degrees of freedom are directly coupled, our method provides a similar degree of accuracy, at a small fraction of the cost. In addition, the method can be easily generalized to more complex and realistic polymer models. (As part of this article, we include a Japanese summary of our main results, published in , CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).)
Building energy simulation (BES) is commonly conducted in architecture design process to evaluate building energy performance. The coupling between computational fluid dynamics (CFD) and BES not only improves energy simulation accuracy but also makes it possible to simultaneously consider energy consumption and indoor environment. Nevertheless, to conduct high fidelity CFD simulation is generally time-consumed and thus it is almost impractical to carry out a long-term coupled simulation that requires multiple CFD executions. A fast and accurate prediction method is therefore required to serve as a surrogate for CFD in the coupled simulation. Inspired by successful applications of deep learning neural networks (NNs) in various fields due to the high computation speed and prediction accuracy, the authors proposed a deep learning-based prediction method to achieve fast and accurate prediction of indoor air distributions. This paper provides a general introduction to the proposed prediction method with regard to its principle, implementation, performance, and application. Predictions of two-dimensional isothermal flow and three-dimensional non-isothermal flow are demonstrated. The results confirmed the feasibility of NN models for fast and accurate indoor airflow prediction. Meanwhile, as an example of practical application, the NN model is coupled with a BES tool to implement a coupled simulation framework for fast energy simulation considering non-uniform indoor environment. The coupling scheme is introduced and validation of the crucial functionality of the framework is presented.
A method for evaluating the plasma motion on the solar surface is developed by a combination of numerical simulation and machine learning. On the solar surface, we can observe the thermal convection of plasma fluid. Since thermal convection relates to several phenomena such as magnetic field generation and coronal heating, the estimation of convection velocity is essential. While we can evaluate the line-of-sight velocity with the Doppler effect, the horizontal velocity field cannot be directly obtained. On the other hand, magnetohydrodynamic numerical simulations have been developed and can reproduce solar convection. We can get precise quantities such as velocities and magnetic fields on the solar surface in the simulations. The horizontal velocity, which is challenging to observe, is also obtained in the numerical simulations. Here, we construct a neural network to evaluate the horizontal velocity field on the solar surface with the observable quantities such as radiation intensity and the corresponding horizontal velocity field obtained from the simulation and applying them to the observation. The correlation coefficient between the horizontal velocities obtained by the neural network and the simulated data is 0.84 only with radiative intensity and 0.90 with radiative intensity, the line-of-sight velocity, and the line-of-sight-magnetic field. Even when we apply the network to the observation data, the correlation coefficient of 0.6 is kept.
In our previous study, we measured the radial and axial temperature distributions of steam-air mixture in a vertical circular pipe (diameter, 49.5 mm; cooling height, 610 mm). The measured temperatures were used to evaluate the condensation heat flux qc and heat transfer coefficient hc, where the bulk quantities were defined at the center of the circular pipe. It is sometimes difficult to define the bulk quantities in a three-dimensional computation for a complex geometry or complex flow field. In this study, therefore, we evaluated a prediction method for hc without the definition of the bulk quantities by using the measured temperature distributions. We applied existing hc correlations to an arbitrary radial location y from the condensation surface, and obtained the distribution of local hc. As a result, we found that hc was radially almost constant in the turbulent boundary layer (dimensionless distance of y+ > 20). This means that existing hc correlations are applicable to computational cells of y+ > 40 in a three-dimensional computation.
In order to evaluate effects of water levels in the upper tank hut on flow characteristics in a vertical pipe under flooding at its top end, counter-current flow limitation (CCFL characteristics), void fraction and pressure gradient were measured for counter-current flows of air and water in a vertical pipe with the diameter of 40 mm at hut = 200 and 300 mm in addition to data at hut = 100 mm measured in our previous study, and the wall and interfacial friction factors were obtained from the data. As a result, the falling water flow rate decreased with increasing hut, and a Wallis-type CCFL correlation with the Kutateladze parameters was derived as a function of hut. With increasing hut, the thickness of the falling water film decreased due to decrease in the falling water flow rate, and the wall and interfacial friction factors decreased due to decrease in the water film thickness. The annular flow model with correlations for CCFL, and the wall and interfacial friction factors could predict the effects of hut on the liquid volume fraction αL and the transition from αL decrease regime to αL increase regime with increasing the gas volumetric flux.