Generative deep learning for hydrodynamic lubrication prediction by a deconvolutional neural network

Classical hydrodynamic lubrication simulation solves the Reynolds equation to unveil the lubrication pressure and film thickness distributions, yet it overlooks the latent high-level representations buried beneath the lubrication data. While deep learning has illustrated that the mining of high-level representations helps to generate desired outputs from input prompts, lubrication research has not fully exploited generative deep learning in lubrication prediction and generation. Here, we propose adopting a deconvolutional neural network to learn the latent representations in hydrodynamic lubrication data and directly generate a two-dimensional (2D) lubrication scenario from the given working conditions without solving any governing equation. Compared with the classical method, our approach can output the distribution of lubrication pressure and film thickness in less than 0.1 s on a personal computer and be extended to more complicated scenarios, including cavitation.