Abstract:The use of deep learning methods for modeling fluid flow has drawn a lot of attention in the past few years. In situations where conventional numerical approaches can be computationally expensive, these techniques have shown promise in offering accurate, rapid, and practical solutions for modeling complex fluid flow problems. The success of deep learning is often due to its ability to extract hidden patterns and features from the data, enabling the creation of data-driven reduced models that can capture the underlying physics of the domain. We present a data-driven reduced model for predicting flow fields in a bed configuration of hot particles. The reduced model consists of a parametric 3D convolutional autoencoder. The first part resolves the spatial and temporal dependencies present in the input sequence, while the second part of the architecture is responsible for predicting the solution at the subsequent timestep based on the information gathered from the preceding part. We also propose the utilization of a post-processing non-trainable output layer following the decoding path to incorporate the physical knowledge, e.g., no-slip condition, into the prediction. The evaluation of the reduced model for a bed configuration with variable particle temperature showed accurate results at a fraction of the computational cost required by traditional numerical simulation methods.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to predict the flow field in hot particle bed configurations by constructing a data - driven reduced - order model (ROM) based on a parameterized 3D convolutional autoencoder (3D Convolutional Autoencoder, 3D CAE). Specifically, the paper attempts to solve the following key problems: 1. **High computational cost**: Traditional numerical simulation methods (such as the discrete element method (DEM), computational fluid dynamics (CFD), and lattice Boltzmann method (LBM)) require a large amount of computational resources when simulating large - scale hot particle bed configurations, resulting in high time and economic costs. The data - driven reduced - order model proposed in the paper can significantly reduce the computational cost while maintaining high prediction accuracy. 2. **Spatial - temporal dependence**: Flow field data has complex spatial - temporal dependence, and it is difficult for traditional methods to efficiently capture these dependencies. The 3D CAE model proposed in the paper can effectively handle the spatial - temporal dependence in the input sequence and predict the flow field solutions at subsequent time steps. 3. **Incorporation of physical knowledge**: In order to improve the prediction accuracy of the model, the paper proposes adding a non - trained post - processing output layer after the decoding path to embed physical knowledge (such as the no - slip condition). This helps to ensure that the prediction results of the model conform to physical laws. 4. **Parameterized prediction**: The paper extends the existing autoencoder framework so that it can handle new parameter values (such as particle temperature), rather than just reproducing known simulation results. This parameterized reduced - order model can be used for optimization design and sensitivity analysis to further improve the performance of hot particle bed configurations. ### Model architecture and methods The model architecture proposed in the paper mainly includes the following parts: - **Encoding path**: Consists of four 3D convolutional layers, each using the tanh activation function. The encoding path is responsible for capturing the spatial - temporal dependence in the input sequence. - **Pre - processing layer**: Before the encoding path, there is a non - trained layer for combining the input sequence and the corresponding simulation parameters (such as particle temperature). The parameters are given in the form of images for easy combination with the input sequence. - **Decoding path**: Consists of four 2D transpose convolutional layers, each also using the tanh activation function. The decoding path is responsible for reconstructing the original - sized flow field image and predicting the flow field solution at the next time step. - **Post - processing layer**: A non - trained post - processing layer that uses the physical information in the parameter images (such as particle position) to enforce the no - slip condition, thereby improving the prediction accuracy. ### Dataset and training process - **Input data**: High - dimensional simulation snapshots \( s(t_m) \) and the corresponding simulation parameters \( \mu \). The snapshot data is obtained from the numerical model, interpolated onto a 256 × 128 uniform Cartesian grid, and then normalized to the range between [- 1, 1]. - **Output data**: The flow field solution at the next time step \( s(t_{h + 1}) \). - **Training process**: Using the supervised learning method, the optimal training parameters \( \theta \) are found by minimizing the loss function \( L(s(t_m),\hat{s}(t_m)) \). The optimization problem can be solved by the gradient descent algorithm or its variants (such as the Adam optimizer). ### Conclusion By constructing a parameterized 3D convolutional autoencoder reduced - order model, the paper successfully solves the high computational cost problem of flow field prediction in hot particle bed configurations and significantly reduces the demand for computational resources while maintaining prediction accuracy. In addition, through the introduction of a post - processing layer with physical knowledge, the prediction accuracy of the model is further improved.

Parametric 3D Convolutional Autoencoder for the Prediction of Flow Fields in a Bed Configuration of Hot Particles

Deep-Learning Strategy Based on Convolutional Neural Network for Wall Heat Flux Prediction

Physics-Based Deep Learning for Flow Problems

Predicting Pressure Fields from Incomplete Velocity Fields Based on Deep Convolutional Neural Network

Deep Convolutional Recurrent Autoencoders for Flow Field Prediction

Investigation of physics-informed deep learning for the prediction of parametric, three-dimensional flow based on boundary data

Combined space-time reduced-order model with 3D deep convolution for extrapolating fluid dynamics

A Point-Cloud Deep Learning Framework for Prediction of Fluid Flow Fields on Irregular Geometries

An Accuracy-Enhanced Transonic Flow Prediction Method Fusing Deep Learning and a Reduced-Order Model

Prediction of Aerodynamic Flow Fields Using Convolutional Neural Networks

Heat Flux Prediction of Radiation Balance Wall by Deep Convolutional Neural Networks

Research on the Role of Hybrid Mesh Warm-up in Flow Prediction Based on Deep Learning

Predicting Transonic Flowfields in Non-Homogeneous Unstructured Grids Using Autoencoder Graph Convolutional Networks

Reconstruction and fast prediction of a 3D flow field based on a variational autoencoder

Fast Prediction of Flow Field Around Airfoils Based on Deep Convolutional Neural Network

Learning-Based High-Performance Algorithm for Long-Term Motion Prediction of Fluid Flows

DeepCFD: Efficient Steady-State Laminar Flow Approximation with Deep Convolutional Neural Networks

Unsteady Flow-Field Forecasting Leveraging a Hybrid Deep-Learning Architecture

A deep-learning model for predicting spatiotemporal evolution in reactive fluidized bed reactor

Combined space–time reduced-order model with three-dimensional deep convolution for extrapolating fluid dynamics

Physics-driven Learning of the Steady Navier-Stokes Equations using Deep Convolutional Neural Networks