Data-driven synthesis of smoke flows with CNN-based feature descriptors

Abstract

Deformation-limiting advection Descriptor learning CNN CNN Fig. 1. We enable volumetric fluid synthesis with high resolutions and non-dissipative small scale details using CNNs and a fluid flow repository.We present a novel data-driven algorithm to synthesize high resolution ow simulations with reusable repositories of space-time ow data. In our work, we employ a descriptor learning approach to encode the similarity between uid regions with di erences in resolution and numerical viscosity. We use convolutional neural networks to generate the descriptors from uid data such as smoke density and ow velocity. At the same time, we present a deformation limiting patch advection method which allows us to robustly track deformable uid regions. With the help of this patch advection, we generate stable space-time data sets from detailed uids for our repositories. We can then use our learned descriptors to quickly localize a suitable data set when running a new simulation. This makes our approach very e cient, and resolution independent. We will demonstrate with several examples that our method yields volumes with very high e ective resolutions, and non-dissipative small scale details that naturally integrate into the motions of the underlying ow.

Data‐driven pedestrian re‐identification based on hierarchical semantic representation

“…42,43 Chu and Thuerey, Girshick et al, and Krizhevsky et al showed that CNN is a good feature generator and selector, which is a novel data-driven method of unsupervised learning. [44][45][46] Deep learning methods are great means of getting features, which were adopted by Sermanet et al to localize, detect, and recognize humans. 47,48 Zeiler and Fergus made the convolutional network visualizing and to be easily understood for recognition.…”

“…() Chu and Thuerey, Girshick et al, and Krizhevsky et al showed that CNN is a good feature generator and selector, which is a novel data‐driven method of unsupervised learning. ()…”

Latent‐space Dynamics for Reduced Deformable Simulation

“…In this paper we turn to machine learning in an attempt to build a fast reduced model simulation algorithm for high resolution elastic bodies. Machine learning algorithms for physics simulation are relatively new, but there have been recent efforts in fluid simulation to learn pressure projection operations [TSSP16], velocity updates [LJS*15], one‐way coupled turbulence models [CT17] and interpolate between precomputed simulations [BPT17]. Until the recent appearance of DeepWarp [LSW*18] and [WKD*18], there had been no such work for deformable body simulation.…”

Data‐driven retrieval of spray details with random forest‐based distance

“…Yang et al utilized an artificial NN to avoid a time‐consuming projection step in the Eulerian method; a similar method is also proposed in the work of Tompson et al with convolutional NN. Furthermore, machine learning technology is successfully applied to patch‐based detail enhancement, superresolution smoke, deformation‐aware fluid, and a Lattice Boltzmann method (LBM) . Um et al approximated detailed subgrid scale sprays with NNs for FLIP liquid simulations.…”