CNN-based flow field prediction for bus aerodynamics analysis.

Area of Science:

• Computational Fluid Dynamics (CFD)
• Artificial Intelligence (AI)
• Aerodynamics

Background:

• Predicting aerodynamic fields for heavy vehicles is crucial for design optimization.
• Traditional CFD simulations are computationally expensive and time-consuming.
• AI, specifically CNNs, shows potential for accelerating these predictions.

Purpose of the Study:

• To evaluate a CNN's ability to predict velocity and pressure aerodynamic fields in heavy vehicles.
• To assess the accuracy and computational efficiency of the CNN compared to CFD.
• To investigate the impact of discretization strategies on CNN performance.

Main Methods:

• Developed and trained a CNN using CFD simulations of three vehicle geometries (SC7, SC5, Ahmed body).
• Employed the RANS-based k-ω SST turbulent model for CFD simulations.
• Discretized field representations based on velocity gradients to enhance CNN accuracy.

Main Results:

• The CNN accurately predicted velocity and pressure fields with low errors, showing good agreement with numerical results.
• Achieved a substantial reduction in computational time, by four orders of magnitude.
• Demonstrated the effectiveness of gradient-based discretization for improving CNN prediction accuracy.

Conclusions:

• CNNs are highly capable of predicting aerodynamic fields for heavy vehicles.
• The developed CNN offers significant computational time savings compared to traditional CFD.
• This AI-driven approach presents a promising advancement for aerodynamic analysis and vehicle design.