Prediction Of Flow Field Research Articles

Swirl-induced motion prediction with physics-guided machine learning utilizing spatiotemporal flow field structure

PurposeHigher energy conversion efficiency of internal combustion engine can be achieved with optimal control of unsteady in-cylinder flow fields inside a direct-injection (DI) engine. However, it remains a daunting task to predict the nonlinear and transient in-cylinder flow motion because they are highly complex which change both in space and time. Recently, machine learning methods have demonstrated great promises to infer relatively simple temporal flow field development. This paper aims to feature a physics-guided machine learning approach to realize high accuracy and generalization prediction for complex swirl-induced flow field motions.Design/methodology/approachTo achieve high-fidelity time-series prediction of unsteady engine flow fields, this work features an automated machine learning framework with the following objectives: (1) The spatiotemporal physical constraint of the flow field structure is transferred to machine learning structure. (2) The ML inputs and targets are efficiently designed that ensure high model convergence with limited sets of experiments. (3) The prediction results are optimized by ensemble learning mechanism within the automated machine learning framework.FindingsThe proposed data-driven framework is proven effective in different time periods and different extent of unsteadiness of the flow dynamics, and the predicted flow fields are highly similar to the target field under various complex flow patterns. Among the described framework designs, the utilization of spatial flow field structure is the featured improvement to the time-series flow field prediction process.Originality/valueThe proposed flow field prediction framework could be generalized to different crank angle periods, cycles and swirl ratio conditions, which could greatly promote real-time flow control and reduce experiments on in-cylinder flow field measurement and diagnostics.

Prediction of instantaneous flow characteristics of hydrocyclone with long short-term memory network based on computational fluid dynamics data

Accelerating the prediction time of separation performance and flow field characteristics in industrial hydrocyclones holds paramount importance for real-time control. Machine learning methods exhibit significant advantages in this particular aspect. This study presents a network model that integrates a long short-term memory (LSTM) layer with a fully connected layer to accurately forecast the flow field and separation performance under various operational conditions, leveraging CFD data. The study evaluates the effect of factors such as the number of LSTM layers and neurons per layer on the model's performance, aiming to identify the optimal network structure. The results demonstrate that the predicted flow field and performance of the hydrocyclones using LSTM closely align with the outcomes from CFD simulations. The average absolute percentage error is constrained to within 6%, which significantly enhances the prediction efficiency. The findings contribute to a more efficient and automated separation process in industrial environments.

Performance of a supercritical carbon dioxide compressor using a streamline curvature based throughflow method

Supercritical carbon dioxide (sCO2) has emerged as a promising working fluid for the Brayton power cycle, offering enhanced efficiency due to its high density and low viscosity near the critical point. A critical component of the sCO2 Brayton cycle is the turbomachinery, namely the compressor. The pronounced sensitivity of the thermophysical properties of sCO2 in the vicinity of the critical point where the compressor operates can significantly impact the performance characteristics of sCO2 compressors. This study presents a two-dimensional (2D) streamline curvature-based throughflow method, considering real gas thermophysical properties, to analyze a centrifugal compressor operating with sCO2 at near-critical inlet conditions. The method’s validity is established through comparisons of the predicted performance characteristics with those of the Sandia National Laboratories compressor and by comparing the predicted flow field with the pitch-averaged flow field obtained from computational fluid dynamic (CFD) analysis. Moreover, various compressor intake conditions including subcritical vapor, supercritical vapor, supercritical liquid, and near-critical conditions are investigated to explore their impact on compressor performance. The streamline curvature-based throughflow method demonstrates its effectiveness by achieving significant agreement between the flow field and performance characteristics, both at design and off-design conditions where efficiency predictions were within 5.84% near the best efficiency point compared to CFD simulations. At near critical conditions, a maximum of 3201% increase the computational run-time required for CFD based computation was observed compared to superheated conditions while only a maximum of 164% increase in run-time was observed in the throughflow based computations as a result of the numerical under-relaxation required.

Fast and high-precision compressible flowfield inference method of transonic airfoils based on attention UNet

Traditional numerical simulation methods for airfoil flowfields are complex and time-consuming, and deep learning-based inference methods for Reynolds-averaged Navier–Stokes equations (RANS) solutions of transonic airfoils have limitations in terms of their robustness and generalization. A novel data-driven inference method named as attention UNet (AU)-RANS is proposed for efficient and accurate prediction of flowfields around airfoils with strong compressibility and large-scale turbulent separation. First, to enhance the learning the boundary flow information and inference of the entire flowfield solution, an innovative data preprocessing method is proposed to convert the physical quantities and coordinate information of RANS solutions into neural network spatial information. Second, an attention mechanism is introduced in UNet to suppress feature responses in irrelevant background regions and enhance sensitivity to the geometrical features of the input airfoil and varying inflow conditions. The quantitative and qualitative analyses of AU-RANS inference results demonstrate that the well-trained model can effectively infer RANS solutions for airfoil flowfield and can accurately predict the shock waves and flow separation phenomena under high Mach number conditions with a large angle of attack.

Development and assessment of an actuator volume method in rotating frame for predicting the flow-field of horizontal-axis wind turbines

The prediction of the unsteady flow-field in horizontal-axis wind turbines using blade-resolved CFD simulations is challenging and computationally demanding. The simulations become more affordable when using a representation of the blades based on the generalized actuator disc concept; however, the projection of the computed forces in the computational domain introduces additional mesh parameters and requires codes that impact the computation time in unstructured solvers and parallel computing. In this work, a new approach to model the effects of the rotating blades using source terms and a sliding mesh technique is proposed. The developed framework also includes the tower and nacelle. The force distributions are obtained using an in-house code, which incorporates Buhl’s correction, three-dimensional effects, and tip and hub loss corrections. The validation was performed against blade-resolved simulations and experimental data at three tip-speed ratios. At design conditions, both models are in good agreement in terms of velocity deficit and angle of attack, being underestimated at off-design conditions. The local effects of the actuator volumes, the induction region, and the near-wake have been analysed in detail. This new approach is flexible, robust, and provides good accuracy, whereas the computational cost is considerably lower than high-fidelity simulations.

Moving least‐squares aided finite element method: A powerful means to predict flow fields in the presence of a solid part

AbstractWith the assistance of the moving least‐squares (MLS) interpolation functions, a two‐dimensional finite element code is developed to consider the effects of a stationary or moving solid body in a flow domain. At the same time, the mesh or grid is independent of the shape of the solid body. We achieve this goal in two steps. In the first step, we use MLS interpolants to enhance the pressure (P) and velocity (V) shape functions. By this means, we capture different discontinuities in a flow domain. In our previous publications, we have named this technique the PVMLS method (pressure and velocity shape functions enhanced by the MLS interpolants) and described it thoroughly. In the second step, we modify the PVMLS method (the M‐PVMLS method) to consider the effect of a solid part(s) in a flow domain. To evaluate the new method's performance, we compare the results of the M‐PVMLS method with a finite element code that uses boundary‐fitted meshes.

Reconstruction and prediction of the rotor wake flow field in hovering state

A clear understanding of the flow structure and evolution of the rotor wake flow field is one of the key problems to be solved in the development of rotor aerodynamics. The tip vortices are the primary structure of the rotor wake flow field. A theoretical model of the vortex ring was established to verify the feasibility of flow field reconstruction. Particle image velocimetry (PIV) is used to measure the two-dimensional velocity field in the rotor wake. Because of the wandering motion of tip vortices, a new conditional average algorithm is proposed to deal with the flow field. On this basis, it is proposed that the plane interval of the rotor wake flow field reconstructed by interpolation should be less than 20 deg. The vortex core radius is obtained by the analysis method based on the circulation. A Back Propagation (BP) neural network was established to predict the axial and radial positions of the tip vortices and the radius of the vortex core. The inputs are the vortex ages measured by the experiment, and the outputs are the position and radius of the vortex core under any vortex ages. The predicted values were in good agreement with the experimental results. The study provides further analytical methods and new insights for the characteristics and evolution of tip vortices.

Fast flow field prediction approach of supersonic inlet in wide operating range based on deep learning

The flow in a supersonic inlet at wide operation range is quite complicated and contains complex shock reflections and shock wave/boundary layer interactions (SWBLI). So it is difficult to achieve the operating state monitoring in real-time by the sensor signals only. The recently booming deep learning (DL) is promising to provide a new way for flow field reconstruction due to its excellent data mining capability. For these, a fast flow field prediction model based on deep learning is proposed for the supersonic inlet in a wide operation range. To establish the data set for model training, numerical simulations at various Mach number and flight altitude obtained by Latin Hypercube Sampling (LHS) method are carried out. The trained model successfully established the mapping relationship between the inlet flow field and the input features. Both the start and unstart flows can be reconstructed by the model with high accuracy and efficiency. Then, two different types of input features are compared to test the influence of different input features on the model performance. The results indicated that the pressure distribution inputs perform better than that of operating conditions, which is because the wall pressure distribution contains more information and characteristics of the flow. Furthermore, compared with the traditional Computational Fluid Dynamics (CFD) simulations and wind tunnel experiments, the model has higher efficiency in predicting the inlet flow field with high accuracy. Finally, interpolation and extrapolation tests are conducted and the results revealed that the model is of high robustness and generalization ability.

Accelerating the development of inertial microfluidic devices using numerical modelling and machine learning

Inertial microfluidics allows for passive, label-free manipulation of particles suspended in a fluid. Physical experiments can understand the underlying mechanisms to an extent whereby inertial microfluidic devices are used in real-world applications such as disease diagnostics. However, design processes are often iterative and device optimisation can be improved. Numerical modelling has complementary capabilities to physical experiments, with access to full flow field data and control of design parameters. Numerical modelling is used to uncover the fundamental mechanisms in inertial microfluidics and provides evidence for physical experiments. In recent years, numerical modelling has been increasingly coupled to machine learning algorithms to uncover additional physics and provide fast solutions. In this perspective, I discuss the role numerical modelling will play in future inertial microfluidic device research and the opportunities to combine numerical modelling with machine learning algorithms. Two key areas for future research applying machine learning are highlighted; fast predictions of flow fields and the optimisation of design parameters. Developments in these areas would significantly reduce the resources required in device design and have the potential to uncover new applications.

Self-supervised learning based on Transformer for flow reconstruction and prediction

Machine learning has great potential for efficient reconstruction and prediction of flow fields. However, existing datasets may have highly diversified labels for different flow scenarios, which are not applicable for training a model. To this end, we make a first attempt to apply the self-supervised learning (SSL) technique to fluid dynamics, which disregards data labels for pre-training the model. The SSL technique embraces a large amount of data (8000 snapshots) at Reynolds numbers of Re = 200, 300, 400, and 500 without discriminating between them, which improves the generalization of the model. The Transformer model is pre-trained via a specially designed pretext task, where it reconstructs the complete flow fields after randomly masking 20% data points in each snapshot. For the downstream task of flow reconstruction, the pre-trained model is fine-tuned separately with 256 snapshots for each Reynolds number. The fine-tuned models accurately reconstruct the complete flow fields based on less than 5% random data points within a limited window even for Re = 250 and 600, whose data were not seen in the pre-trained phase. For the other downstream task of flow prediction, the pre-training model is fine-tuned separately with 128 consecutive snapshot pairs for each corresponding Reynolds number. The fine-tuned models then correctly predict the evolution of the flow fields over many periods of cycles. We compare all results generated by models trained via SSL and models trained via supervised learning, where the former has unequivocally superior performance. We expect that the methodology presented here will have wider applications in fluid mechanics.

A novel method for predicting fluid–structure interaction with large deformation based on masked deep neural network

Traditional fluid–structure interaction (FSI) simulation is computationally demanding, especially for bi-directional FSI problems. To address this, a masked deep neural network (MDNN) is developed to quickly and accurately predict the unsteady flow field. By integrating the MDNN with a structural dynamic solver, an FSI system is proposed to perform simulation of a flexible vertical plate oscillation in fluid with large deformation. The results show that both the flow field prediction and structure response are consistent with the traditional FSI system. Furthermore, the masked method is highly effective in mitigating error accumulation during temporal flow field predictions, making it applicable to various deformation problems. Notably, the proposed model reduces the computational time to a millisecond scale for each step regarding the fluid part, resulting in an increase in nearly two orders of magnitude in computational speed, which greatly enhances the computational speed of the FSI system.

A fast three-dimensional flow field prediction around bluff bodies using deep learning

This study presents a deep learning approach for predicting the flow field in the incompressible turbulent three-dimensional (3D) external flow around right-rhombic prism-shaped bluff bodies. The approach involves treating the nodes of the unstructured grid in the computational fluid dynamics domain as a point cloud, which is used as an input for a neural network. The neural network is trained to map the spatial coordinates of the nodes to the corresponding velocity and pressure values in the domain. The PointNet, a reliable solution in 3D vision tasks, is selected as the neural network architecture. Implementing this architecture makes it feasible to use irregular positions of the nodes of an unstructured grid as an input without needing interpolation. A dataset, comprising 3511 cases, is generated for training and testing the network. This is achieved by changing the geometric parameters of a right rhombic prism and varying its angle to the flow stream. Then, the continuity and momentum equations for turbulent flow are solved using a solver. Given the need for a larger number of points to accurately represent a 3D flow, the architecture of PointNet is modified. This modification involves adding extra layers and adjusting the number of neurons inside the layers to overcome this challenge. Once the training is completed, given the unseen samples from the test dataset to the model, our model can predict the velocity and pressure of the flow field at a speed that exceeds our conventional solver by several orders of magnitude with a maximum relative error of 4.58%.

Flow field prediction of S-shaped shock vectoring nozzle with rear deck based on deep learning

The S-shaped shock wave vectoring nozzle with afterdeck can significantly improve the overall performance of the exhaust system, taking into account Thrust vectoring, infrared stealth and afterbody fusion. One of the technical difficulties in its design process lies in the complex flow field characteristics under different operating conditions. Currently, the mainstream method is to obtain nozzle flow field characteristics through CFD numerical simulation, but the CFD method is time-consuming and costly. Therefore, based on the depth learning principle, a depth Convolutional neural network based on U-NET framework is established to quickly predict the flow field of S-shaped shock wave vectoring nozzle with afterdeck. Using CFD data for training, the results show that the depth learning model has high prediction accuracy and can clearly predict the flow field characteristics inside the nozzle, especially the secondary flow and the complex wave structure near the afterdeck. The correction Coefficient of determination of the prediction model is greater than 0.97. And the time consumption is about 0.0689% of that of a conventional solver. It has good application prospects in quickly evaluating the flow field of S-shaped nozzles.

Sag-flownet: self-attention generative network for airfoil flow field prediction

Sag-flownet: self-attention generative network for airfoil flow field prediction

Estimation of sequential transient flow around cylinders using recurrent neural network coupled graph convolutional network

This paper develops a novel reduced-order model called SGCNN, which combines the Graph Convolutional Networks (GCN) with the Recurrent Neural Networks (RNN), to predict sequential transient flow around cylinders using unstructured flow field data. In flow field prediction, we use GCN's unique message passing mechanism to extract and aggregate the characteristics of the flow field grid in non-Euclidean space, eliminating the pixelization process of CNN in non-Euclidean space, and significantly improving the prediction accuracy. GRU is used as the unit of the RNN network to obtain the changeable physical characteristics through the hidden layer, thereby achieving continuous prediction of the flow field. According to our literature review, this is the first investigation to estimate sequential transient flow fields around cylinders. The proposed model is initially verified with the flow around single cylinder. Taking the numerical simulation as the reference, the relative errors of the predicted sequential velocity and the pressure fields are less than 3% and 2%, respectively. Subsequently, the complexity of the investigation is increased by varying the Reynolds number and introducing additional cylinders in the flow region. The predicted physical fields consistently align well with CFD simulations, and the computational time is remarkably reduced by more than 90%.

Fast flow field prediction of three-dimensional hypersonic vehicles using an improved Gaussian process regression algorithm

Conducting large-scale numerical computations to obtain flow field during the hypersonic vehicle engineering design phase can be excessively costly. Although deep learning algorithms enable rapid flow field prediction with high-precision, they require a significant investment in training samples, contradicting the motivation of reducing the cost of acquiring flow field. The combination of feature extraction algorithms and regression algorithms can also achieve high-precision prediction of flow fields, which is more suitable to tackle three-dimensional flow prediction with a small dataset. In this study, we propose a reduced-order model (ROM) for the three-dimensional hypersonic vehicle flow prediction utilizing proper orthogonal decomposition to extract representative features and Gaussian process regression with improved automatic kernel construction (AKC-GPR) to perform a nonlinear mapping of physical features for prediction. The selection of variables is based on sensitivity analysis and modal assurance criterion. The underlying relationship is unveiled between flow field variables and inflow conditions. The ROM exhibits high predictive accuracy, with mean absolute percentage error (MAPE) of total field less than 3.5%, when varying altitudes and Mach numbers. During angle of attack variations, the ROM only effectively reconstructs flow distribution by interpolation with a MAPE of 7.02%. The excellent small-sample fitting capability of our improved AKC-GPR algorithm is demonstrated by comparing with original AKC-GPRs with a maximum reduction in a MAPE of 35.28%. These promising findings suggest that the proposed ROM can serve as an effective approach for rapid and accurate vehicle flow predicting, enabling its application in engineering design analysis.

Predictions for unsteady flow fields with deep learning models of LSTM and GRU

Predictions for unsteady flow fields with deep learning models of LSTM and GRU

An Adapted ResNet-50 Architecture for Predicting Flow Fields of an Underwater Vehicle

An Adapted ResNet-50 Architecture for Predicting Flow Fields of an Underwater Vehicle

A Data-Driven Machine Learning Approach for Turbulent Flow Field Prediction Based on Direct Computational Fluid Dynamics Database

A novel approach is presented for predicting compressible turbulent flow fields using a neural network-based data-driven method. Accurate prediction in turbulent regions heavily relies on the resolution of available data. Traditional methods, employing image-based techniques by mapping scattered computational fluid dynamics (CFD) data onto Cartesian grids, encounter data scarcity in critical areas such as the boundary layer and wake. Recently, convolutional neural networks (CNN) have gained prominence as the most widely referenced technique in fluid dynamics, utilizing flow field images as datasets for flow field prediction. However, CNN requires datasets with a high pixel density to enhance training accuracy in crucial regions, thereby increasing the input data volume and machine training time. To address this challenge, our proposed method deviates from using flow field images and instead generates datasets directly from the flow field properties of CFD grid points. By employing this approach, several advantages are realized. Firstly, the network benefits from the favorable characteristics of unstructured grids, such as varying point spacing near the object surface and in the far field, which effectively reduces the amount of input data and consequently the machine training cost. Secondly, the construction of the training dataset eliminates the need for interpolation or extrapolation, thereby preserving the accuracy of CFD data. In this case, a simple multilayer perceptron can be trained using the proposed dataset. Various flow field properties, including static pressure, turbulent kinetic energy, and velocity components, can be predicted with high accuracy within a few seconds.

Shock wave prediction in transonic flow fields using domain-informed probabilistic deep learning

Transonic flow fields are marked by shock waves of varying strength and location and are crucial for the aerodynamic design and optimization of high-speed transport aircraft. While deep learning methods offer the potential for predicting these fields, their deterministic outputs often lack predictive uncertainty. Moreover, their accuracy, especially near critical shock regions, needs better quantification. In this paper, we introduce a domain-informed probabilistic (DIP) deep learning framework tailored for predicting transonic flow fields with shock waves called DIP-ShockNet. This methodology utilizes Monte Carlo dropout to estimate predictive uncertainty and enhances flow-field predictions near the wall region by employing the inverse wall distance function-based input representation of the aerodynamic flow field. The obtained results are benchmarked against the signed distance function and the geometric mask input representations. The proposed framework further improves prediction accuracy in shock wave areas using a domain-informed loss function. To quantify the accuracy of our shock wave predictions, we developed metrics to assess errors in shock wave strength and location, achieving errors of 6.4% and 1%, respectively. Assessing the generalizability of our method, we tested it on different training sample sizes and compared it against the proper orthogonal decomposition (POD)-based reduced-order model (ROM). Our results indicate that DIP-ShockNet outperforms POD-ROM by 60% in predicting the complete transonic flow field.