Deep learning in turbulent convection networks.

Area of Science:

• Fluid Dynamics
• Turbulence Research
• Heat Transfer

Background:

• Turbulent Rayleigh-Bénard convection is crucial for understanding heat transport in various natural and industrial systems.
• Horizontally extended systems exhibit complex, slowly evolving turbulent superstructures.
• These superstructures, larger than the convection layer height, manifest as temporal patterns of fluid upwelling and downwelling.

Purpose of the Study:

• To investigate heat transport properties in turbulent Rayleigh-Bénard convection.
• To analyze the role of turbulent superstructures and their defects in heat transfer.
• To leverage deep learning for dimensionality reduction of complex fluid dynamics.

Main Methods:

• Application of deep-learning algorithms, specifically a U-shaped deep convolutional neural network (CNN).
• Reduction of complex 3D turbulent superstructures to a temporal planar network in the midplane.
• Achieving data compression exceeding five orders of magnitude at high Rayleigh numbers.

Main Results:

• Identification of a discrete transport network with dynamically varying defect points.
• Discovery of "hot spots" representing locally enhanced heat flux within the network.
• Quantification of heat transport contributions from superstructures and background turbulence.

Conclusions:

• The fraction of heat transport by superstructures decreases as the Rayleigh number increases.
• Despite decreasing contribution, individual superstructures can remain significant heat transporters.
• Small-scale background turbulence gains importance in heat transport at higher Rayleigh numbers.