AI & Deep Learning for CFD Flow Prediction

In a series of publications Prof. Thi-Thu-Huong Le, Hoyeun Kang, and
colleagues from Pusan National University (PNU) in Korea, have
successfully developed a new deep learning (DL) CFD methodology for
flow prediction using FEATool Multiphysics with AI and machine learning.

Although traditional Computational Fluid Dynamics (CFD) solvers are
becoming increasingly common tool in aerospace, automotive, and other
engineering fields, while effective, face limitations, primarily
related to high computational cost. Especially when high spatial and
temporal resolution is needed for accurate simulation, traditional CFD
solvers become computationally expensive, and can sometimes require
days or even weeks to perform accurate analysis.

Deep learning have potential to offer an alternative approach that may
significantly reduce computational cost, creating an AI CFD model on
an accurate fluid dynamics dataset. The autors propose such a new CFD
model, named CFDformer, which combines a Vision Transformer (ViT)
and a U-shaped Convolutional Neural Network (U-Net) in an
encoder-decoder architecture to predict fluid flow on 2D geometries.

A DL-CFD model agent can offer a number of advantages compared to
traditional CFD simulations

• Act as Surrogate Models - Deep learning models, trained on data
  from CFD simulations, can act as surrogate models, effectively
  approximating the solutions to the Navier-Stokes equations. This
  approach bypasses the need for computationally intensive iterative
  calculations.

• Speed Up Simulations - Deep learning models can significantly
  reduce simulation times. For example, CFDformer, a hybrid model
  combining a Vision Transformer and a U-Net, was shown to decrease
  analysis time by up to 99.94% compared to standard CFD solvers.

• Complex Flow Scenarios - Deep learning models can handle
  complex flow scenarios with relatively high accuracy, and are
  particularly well-suited for modeling flows around obstacles, such
  as found in aerodynamics applications.

• Adapt to New Conditions - Some deep learning models can generalize
  well to new, unseen conditions not encountered during training. This
  allows them to make reasonable predictions for conditions within a
  specific range, even without explicit training data for those
  conditions.

• Enhancing Feature Extraction - Deep learning models can be
  designed to extract both local and global features from input data,
  improving the accuracy of flow approximations. For instance,
  CFDformer uses convolutional layers to capture local spatial
  features and a Vision Transformer to analyze global flow features.

The FEATool Multiphysics toolbox was used to generate datasets of 2D
incompressible and laminar flows around various obstacles, including
cylinders, triangles, rectangles, and pentagons. In particular, as
FEATool supports multiple CFD solvers,
such as OpenFOAM, using multiple solvers is ideal way to generate
accurate and trustworthy, benchmark and validation CFD studies and
datasets. These datasets served as the ground truth for training and
evaluating deep learning models.

FEATool's easy to use GUI, and integration with a MATLAB programming
and scripting API allowed for easy manipulation and analysis of the
simulation data, streamlining the process of preparing the data for
deep learning. Once FEATool completed the simulations, the researchers
extracted the velocity and pressure data at each grid point. This data
was then preprocessed and formatted as input for training deep
learning models.

Please visit https://www.featool.com for more information and toolbox download.

References

• Kang H., et al. A new fluid flow approximation method using a vision transformer and a U-shaped convolutional neural network, AIP Advances, Volume 13, Issue 2, 2023, doi: 10.1063/5.0138515.

• Prihatno A.T., et al. 2D Fluid Flows Prediction Based on U-Net Architecture, Proceedings of International Conference on Artificial Intelligence in Information and Communication (ICAIIC), 2023, doi: 10.1109/ICAIIC57133.2023.10066980.

• Le T.T.H, Kang H. et al. CFD Prediction of Indoor Airflow using Deep Learning, Conference paper, 2022.

• Le T.T.H., Kang H., Kim H. Towards Incompressible Laminar Flow Estimation Based on Interpolated Feature Generation and Deep Learning, Sustainability, 14, 11996, 2022, doi: 10.3390/su141911996.

• Kang H. CFDformer: Novel Fluid Flow Approximation based on ViT and U-Net, GitHub repository for CFDformer MATLAB dataset generation, doi: 10.5281/zenodo.7527624, 2023.