In Situ Data-Driven Analysis and Learning of Turbulence Closures at Scale
TimeMonday, 15 November 20212:40pm - 3pm CST
DescriptionLarge eddy simulation, which requires modeling of the sub-grid stress (SGS) tensor, offers a compromise between accuracy and efficiency of computations of turbulence flows. Data-driven approaches, such as neural networks (NN), have recently emerged and present encouraging results for improved predictive capacity over traditional state of the art models. However, these NN models must be trained on instantaneous high-fidelity turbulent data. While this is feasible for smaller computations, learning from complex flows requires multi-terabyte databases.
The work presented offers a solution to this limitation by performing in situ learning, wherein the SGS NN model is trained concurrently with the flow simulation producing the data.
The proposed infrastructure combines the flow solver PHASTA with a data parallel machine learning algorithm through SmartSim, which deploys a distributed in-memory cluster (i.e., a database) and provides clients that can connect to the database from both applications. Implementation on the Theta system shows promising results.