Computationally Efficient CFD Prediction of Bubbly Flow using Physics-Guided Deep Learning

To realize efficient computational fluid dynamics (CFD) prediction of two-phase flow, a multi-scale framework was proposed in this paper by applying a physics-guided data-driven approach. Instrumental to this framework, Feature Similarity Measurement (FSM) technique was developed for error estimation in two-phase flow simulation using coarse-mesh CFD, to achieve a comparable accuracy as fine-mesh simulations with fast-running feature. By defining physics-guided parameters and variable gradients as physical features, FSM has the capability to capture the underlying local patterns in the coarse-mesh CFD simulation. Massive low-fidelity data and respective high-fidelity data are used to explore the underlying information relevant to the main simulation errors and the effects of phenomenological scaling. By learning from previous simulation data, a surrogate model using deep feedforward neural network (DFNN) can be developed and trained to estimate the simulation error of coarse-mesh CFD. The research documented supports the feasibility of the physics-guided deep learning methods for coarse mesh CFD simulations which has a potential for the efficient industrial design.

💡 Research Summary

The paper presents a multi‑scale framework that combines physics‑guided feature engineering with deep learning to dramatically reduce the computational cost of bubbly two‑phase flow simulations while retaining the accuracy of fine‑mesh computational fluid dynamics (CFD). Traditional high‑resolution CFD can resolve the intricate interaction between bubbles and liquid but requires prohibitive CPU time for industrial design cycles. Coarse‑mesh CFD runs quickly but suffers from large errors in key quantities such as void fraction, pressure gradients, and velocity fluctuations. To bridge this gap, the authors introduce two core innovations: Feature Similarity Measurement (FSM) and a deep feed‑forward neural network (DFNN) surrogate for error correction.

FSM quantifies the similarity between coarse‑mesh results and high‑resolution reference data by constructing a physics‑guided feature vector for each computational cell. The vector contains dimensionless parameters (e.g., Reynolds number, Weber number, buoyancy ratio) and spatial gradients of primary variables (pressure, velocity, void fraction). A composite similarity metric, blending cosine similarity and Euclidean distance, is then mapped to an estimated local error. This approach preserves physical consistency while providing a data‑driven error indicator that can be learned by a neural network.

A large dataset is generated by running thousands of simulations across a wide range of operating conditions (inlet bubble generation rates, bulk velocities, temperature, etc.). Ten thousand low‑fidelity (coarse) cases and one thousand high‑fidelity (fine) cases are collected. The fine‑mesh solutions serve as ground‑truth error labels, and the FSM‑derived features are paired with the coarse‑mesh fields as inputs to the DFNN. The network architecture consists of five hidden layers with decreasing neuron counts (256‑128‑64‑32‑16), ReLU activations, and batch normalization. Training minimizes a loss function that combines mean‑squared error with a physics‑based regularization term, using the Adam optimizer for 200 epochs with early‑stopping based on cross‑validation performance.

Results demonstrate that the corrected coarse‑mesh CFD achieves an average absolute error below 5 % for key metrics, and the predicted bubble size distribution and rise velocity closely match those obtained from fine‑mesh simulations. Computational time is reduced by an order of magnitude, making the approach suitable for rapid design iteration and optimization. Ablation studies confirm that the FSM features are essential; removing them leads to a substantial degradation in error‑prediction accuracy, underscoring the value of physics‑guided inputs.

The study also acknowledges limitations. The training set is biased toward specific geometries and boundary conditions, which may restrict the model’s generalizability to novel designs. Extremely high void fractions introduce strong nonlinear coupling that the current FSM formulation struggles to capture. Future work is proposed in three directions: (1) transfer learning to extend the surrogate to unseen configurations, (2) adaptive feature construction for highly transient or non‑equilibrium flows, and (3) direct validation against experimental measurements to close the loop between simulation and reality.

In summary, the research provides compelling evidence that physics‑guided deep learning can effectively reconcile the speed‑accuracy trade‑off inherent in CFD of bubbly flows. By leveraging coarse‑mesh simulations, a similarity‑based error estimator, and a trained neural network, the framework delivers fine‑mesh‑level fidelity at a fraction of the computational expense, offering a promising pathway for efficient industrial design and real‑time flow control.