🔹Title: Nascent fractal scaling during transition to turbulence: discovery using machine learning, virtual sensors, and a public database system


🔹Speaker: Charles Meneveau (Johns Hopkins University)


🔹 Abstract: When a boundary layer developing on a smooth surface transitions from a laminar to a turbulent state, theprocess is often accompanied by the appearance of spatially localized patches (spots) of turbulence thatgrow and merge downstream to become the fully turbulent boundary layer. A long-standing question hasbeen whether these incipient spots already contain properties of high-Reynolds-number, developedturbulence. Here, we pose this question for geometric scaling properties of the interface separating theturbulence within the spots from the outer flow. For high-Reynolds-number turbulence, such interfaces areknown to display fractal scaling laws with a dimension near D = 7/3, where the 1/3 excess exponent above2 (smooth surfaces) follows from Kolmogorov scaling of velocity fluctuation's spatial increments. The dataused in this study to examine geometric scaling properties are from a direct numerical simulation (DNS).The data are archived in an open database system (the Johns Hopkins Turbulence Database), where dataaccess is facilitated by a user-friendly "virtual sensors" approach. At present JHTDB contains over 1/2Petabyte of DNS data from various turbulent flow simulations, and it has been used in over 250 peer-reviewed journal publications on turbulence from authors world-wide. Based on these data, we show thatthe spot boundaries (interfaces) can be determined by using an unsupervised machine-learning methodthat identifies such interfaces without the need to choose arbitrary thresholds. Scaling properties of theinterface are studied and links to fractal properties of turbulent non-turbulent interfaces in high Reynoldsnumber flows are established. This work has been performed with Drs. Zhao Wu and Tamer Zaki, whilethe database (supported by the NSF) has resulted from a long-term collaboration with the JHTDB team.