Toward accelerated data-driven Rayleigh-Bénard convection simulations.

Area of Science:

• Computational fluid dynamics
• Machine learning applications in physics
• Thermal convective flows

Background:

• Accurate simulation of thermal convective flows is crucial in many engineering applications.
• Traditional numerical methods face challenges with accuracy and computational cost, especially in near-wall regions.
• Subgrid models are often used but can introduce additional errors.

Purpose of the Study:

• To introduce a hybrid data-driven/finite volume method for 2D and 3D thermal convective flows.
• To improve the accuracy of heat flux prediction and pointwise accuracy in coarse simulations.
• To leverage machine learning for enhanced performance in fluid dynamics simulations.

Main Methods:

• A single-step loss, convolutional neural network (CNN) was developed.
• The CNN is activated exclusively in the near-wall region of the flow.
• The training procedure incorporated temporal flow development and distributional bias.

Main Results:

• The hybrid method significantly reduces errors in long-time heat flux prediction.
• Pointwise accuracy is increased in coarse simulations compared to traditional methods.
• The machine learning model's success is attributed to the specific training procedure.

Conclusions:

• The hybrid data-driven/finite volume method offers a promising approach for accurate and efficient simulation of thermal convective flows.
• Machine learning, particularly CNNs in near-wall regions, can enhance traditional numerical methods.
• Careful consideration of the training procedure is key to the success of data-driven models in scientific computing.