Abstract:<p>This study performs data‐driven modeling of mesoscale solids stress closures for filtered two‐fluid model (fTFM) in gas‐particle flows via an artificial neural network (ANN) based machine learning method. The data used for developing the ANN‐based predictive data‐driven modeling framework is systematically filtered from fine‐grid simulations. The loss function optimization result reveals that coupling two loss functions promotes more accurate predictions of the mesoscale solids stresses than using a single loss function. Further comprehensive assessments of closure markers demonstrate a systematic dependence of the mesoscale solids stresses on the filtered particle velocity and its gradient as additional anisotropic markers, instead of using the conventional isotropic filtered rate of solid phase deformation as a closure marker. An optimal three‐marker mesoscale closure is thus proposed. Comparative analysis of the conventional filtered model and present three‐marker model shows that the data‐driven model can substantially enhance the prediction accuracy.</p><p>This article is protected by copyright. All rights reserved.</p>

What problem does this paper attempt to address?

The paper attempts to address the issue of mesoscale solid-phase stress closure in gas-particle flows. Specifically, the authors utilize a data-driven approach based on artificial neural networks (ANN) to establish mesoscale solid-phase stress closure relationships in filtered two-phase flow models (fTFM). Traditional methods cannot accurately predict the impact of mesoscale structures in coarse-grid simulations, while fine-grid simulations, although capable of capturing these structures, are computationally expensive. Therefore, this study aims to improve the prediction accuracy of mesoscale solid-phase stress while ensuring computational efficiency through a data-driven approach. ### Main Issues 1. **Accurate Prediction of Mesoscale Solid-Phase Stress**: - Traditional uniform drag models and kinetic theory of granular flow (KTGF) cannot accurately predict the impact of mesoscale structures in coarse-grid simulations. - Fine-grid simulations can capture mesoscale structures but are too computationally expensive for industrial-scale reactors. 2. **Optimization of Mesoscale Stress Closure Relationships**: - Traditional closure relationships often use homogenized markers (such as filtered solid volume fraction, filtered strain rate, etc.), but these markers may not adequately describe the physical anisotropy of mesoscale stress. - More suitable markers need to be found to improve the prediction accuracy of mesoscale stress. ### Solutions 1. **Data-Driven Modeling**: - Use high-resolution data generated from fine-grid simulations to establish a data-driven mesoscale solid-phase stress prediction model through artificial neural networks (ANN). - Optimize the loss function by coupling multiple loss functions to improve prediction accuracy. 2. **Marker Optimization**: - Evaluate the impact of different markers (such as filtered particle velocity, filtered particle velocity gradient, etc.) on prediction performance. - Propose an optimal three-marker closure model, including filtered solid volume fraction, filtered particle velocity, and its gradient. 3. **Model Validation and Generalization Enhancement**: - Evaluate the model's prediction performance using data from training, validation, and test sets. - Increase the number of training cases to enhance the model's generalization capability. ### Significance of the Study - **Improved Prediction Accuracy**: The data-driven approach significantly improves the prediction accuracy of mesoscale solid-phase stress, which is crucial for understanding and optimizing the operation of gas-particle flow reactors. - **Reduced Computational Cost**: Compared to fine-grid simulations, the data-driven model can significantly reduce computational costs while maintaining high prediction accuracy, making it suitable for industrial-scale applications. In summary, this paper addresses the modeling issue of mesoscale solid-phase stress closure relationships in gas-particle flows through a data-driven approach, providing new insights for improving the control accuracy and efficiency of reactor operations.

Data‐driven modeling of mesoscale solids stress closures for filtered two‐fluid model in gas–particle flows

Machine learning approaches to close the filtered two-fluid model for gas-solid flows: Models for subgrid drag force and solid phase stress

Data-driven Nonlinear Constitutive Relations for Rarefied Flow Computations

Investigations of data-driven closure for subgrid-scale stress in large-eddy simulation

Development of a Filtered CFD-DEM Drag Model with Multiscale Markers Using an Artificial Neural Network and Nonlinear Regression

Conventional and data‐driven modeling of filtered drag, heat transfer, and reaction rate in gas–particle flows

Data-driven RANS closures for improving mean field calculation of separated flows

Construction and Analysis of the Mesoscale Drag Force Model Based on Machine Learning Methods

Machine Learning Surrogate Models for Landau Fluid Closure

Deconvolutional artificial-neural-network framework for subfilter-scale models of compressible turbulence

Deep learning interfacial momentum closures in coarse-mesh CFD two-phase flow simulation using validation data

Neural network models for the anisotropic Reynolds stress tensor in turbulent channel flow

Data-Driven Stochastic Closure Modeling via Conditional Diffusion Model and Neural Operator

An offline multi‐scale unsaturated poromechanics model enabled by self‐designed/self‐improved neural networks

Machine Learning based algorithms for modeling natural convection fluid flow and heat and mass transfer in rectangular cavities filled with non-Newtonian fluids

Deep learning models for assisted decision-making in performance optimization of thin film nanocomposite membranes

Multiscale modeling framework of a constrained fluid with complex boundaries using twin neural networks

Neural network-based closure models for large-eddy simulations with explicit filtering

Generalizable data-driven turbulence closure modeling on unstructured grids with differentiable physics

Machine-learnt closure models based on computational space training frameworks for large-eddy simulations in natural convection

Learning nonlocal constitutive models with neural networks