TY - JOUR
T1 - Deep learning in turbulent convection networks
AU - Fonda, Enrico
AU - Pandey, Ambrish
AU - Schumacher, Jörg
AU - Sreenivasan, Katepalli R.
N1 - Publisher Copyright:
© 2019 National Academy of Sciences. All rights reserved.
PY - 2019/4/30
Y1 - 2019/4/30
N2 - We explore heat transport properties of turbulent Rayleigh–Bénard convection in horizontally extended systems by using deep-learning algorithms that greatly reduce the number of degrees of freedom. Particular attention is paid to the slowly evolving turbulent superstructures—so called because they are larger in extent than the height of the convection layer—which appear as temporal patterns of ridges of hot upwelling and cold downwelling fluid, including defects where the ridges merge or end. The machine-learning algorithm trains a deep convolutional neural network (CNN) with U-shaped architecture, consisting of a contraction and a subsequent expansion branch, to reduce the complex 3D turbulent superstructure to a temporal planar network in the midplane of the layer. This results in a data compression by more than five orders of magnitude at the highest Rayleigh number, and its application yields a discrete transport network with dynamically varying defect points, including points of locally enhanced heat flux or “hot spots.” One conclusion is that the fraction of heat transport by the superstructure decreases as the Rayleigh number increases (although they might remain individually strong), correspondingly implying the increased importance of small-scale background turbulence.
AB - We explore heat transport properties of turbulent Rayleigh–Bénard convection in horizontally extended systems by using deep-learning algorithms that greatly reduce the number of degrees of freedom. Particular attention is paid to the slowly evolving turbulent superstructures—so called because they are larger in extent than the height of the convection layer—which appear as temporal patterns of ridges of hot upwelling and cold downwelling fluid, including defects where the ridges merge or end. The machine-learning algorithm trains a deep convolutional neural network (CNN) with U-shaped architecture, consisting of a contraction and a subsequent expansion branch, to reduce the complex 3D turbulent superstructure to a temporal planar network in the midplane of the layer. This results in a data compression by more than five orders of magnitude at the highest Rayleigh number, and its application yields a discrete transport network with dynamically varying defect points, including points of locally enhanced heat flux or “hot spots.” One conclusion is that the fraction of heat transport by the superstructure decreases as the Rayleigh number increases (although they might remain individually strong), correspondingly implying the increased importance of small-scale background turbulence.
KW - Machine learning
KW - Temporal networks
KW - Turbulent convection
UR - http://www.scopus.com/inward/record.url?scp=85065509912&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85065509912&partnerID=8YFLogxK
U2 - 10.1073/pnas.1900358116
DO - 10.1073/pnas.1900358116
M3 - Article
C2 - 30988195
AN - SCOPUS:85065509912
SN - 0027-8424
VL - 116
SP - 8667
EP - 8672
JO - Proceedings of the National Academy of Sciences of the United States of America
JF - Proceedings of the National Academy of Sciences of the United States of America
IS - 18
ER -