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

Conventional prediction of the lubricated contact is conducted by solving governing equations, e.g., Reynolds equation. Here, inspired by modern generative deep learning method, we propose to predict the lubrication scenario by deep neural networks.

In specific, we design and train a deep deconvolutional neural network to achieve immediate lubrication pressure and film thickness prediction by the trained deconvolutional neuraln nework with high predicting precision. Compared to classical methods that takes a few seconds for lubrication simulation, our data-driven method is able to ouput the lubrication scenartion within 0.1 s on an ordinary personal computer. In addition, to look into how the neural network generates the desired lubrication pressure and film thickness distributions, we extract the latent high-level features the neural network constructs from the input working conditions, which show a checkerboard-like behaviour at the output of each layer. Then, integrating all these checkboard-like patterns for the processing of the neural network's activation function transforms the latent features into the desired presssure distributions in hydrodynamic lubrication. We also show that the selection of non-linear activation function, e.g., sigmoid, helps to complete the non-linear transformation from the last layers' multiple oututs to the final lubrication distribution. Moreover, a slight finetune of the trained neural netowrk could easily expand the neural network's applicability range into cavitated lubrication contact and other scenarios, which endorse the versatility of our data-driven method for lubrication prediction.

In conclusion, we levarage modern generative deep learning models to achieve immediate and accurate hydrodynamic lubrication pressure and film thickness distribution prediction.