Application of deep reinforcement learning in parameter optimization and refinement of turbulence models

Area of Science:

• Computational fluid dynamics (CFD)
• Aerodynamics
• Machine Learning

Background:

• Traditional wind tunnel tests and on-site measurements for aerodynamic simulations are time-consuming and costly.
• Accurate turbulence models are essential for reliable CFD simulations, particularly for complex wind fields around buildings.
• Existing optimization methods like Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) have limitations in efficiency and accuracy.

Purpose of the Study:

• To develop and validate a novel parameter optimization method for turbulence models using Deep Deterministic Policy Gradient (DDPG).
• To enhance the accuracy of computational fluid dynamics (CFD) simulations for building wind fields by optimizing turbulence model parameters.
• To reduce the reliance on expensive and time-consuming experimental methods.

Main Methods:

• Utilized the Shear Stress Transport (SST) k-ω turbulence model as the base for optimization.
• Employed OpenFOAM for numerical simulations of complex building wind fields.
• Implemented Deep Deterministic Policy Gradient (DDPG) for turbulence model parameter optimization, using Gaussian Process Regression (GPR) as a surrogate model for initial CFD data.
• Conducted sensitivity analysis to identify key parameters affecting wind pressure coefficient (WPC) simulations.

Main Results:

• The DDPG optimized turbulence model parameters significantly improved the accuracy of wind pressure coefficient (WPC) simulations.
• Optimized WPC values (average, RMS, maximum, minimum) showed closer agreement with actual WPC data in single wind direction angles.
• In the 0°-50° wind direction angle range, DDPG optimization led to simulated WPC values more closely matching actual data.
• The DDPG method demonstrated a significant reduction in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) compared to GA and PSO.

Conclusions:

• The DDPG-based parameter optimization method offers a more accurate and efficient approach for turbulence modeling in CFD simulations.
• This method effectively reduces simulation errors and improves the prediction of wind pressure coefficients around buildings.
• DDPG optimization presents a superior alternative to traditional methods and other evolutionary algorithms for enhancing CFD accuracy.