Frequency-Aware Reconstruction of Fluid Simulations with Generative Networks

Simon Biland, Vinicius C. Azevedo, Byungsoo Kim, Barbara Solenthaler
ETH Zurich
Eurographics & Eurovis 2020 Short Paper, arXiv:1912.08776


Convolutional neural networks were recently employed to fully reconstruct fluid simulation data from a set of reduced parame-ters. However, since (de-)convolutions traditionally trained with supervised L1-loss functions do not discriminate between lowand high frequencies in the data, the error is not minimized efficiently for higher bands. This directly correlates with the qualityof the perceived results, since missing high frequency details are easily noticeable. In this paper, we analyze the reconstructionquality of generative networks and present a frequency-aware loss function that is able to focus on specific bands of the datasetduring training time. We show that our approach improves reconstruction quality of fluid simulation data in mid-frequencybands, yielding perceptually better results while requiring comparable training time.





	author = {Biland, Simon and C. Azevedo, Vinicius and Kim, Byungsoo and Solenthaler, Barbara},
	title = {{Frequency-Aware Reconstruction of Fluid Simulations with Generative Networks}},
	year = {2020},
	month = {may},
	booktitle = {EGEV 2020 - Short Papers},
	publisher = {The Eurographics Association},


This work was supported by the Swiss National Science Foundation (grant No. 200021_168997).