Fast Computation of Lubricated Contacts: A Physics-Informed Deep Learning Approach

The frictional behavior of pneumatic seals determines functionality in various fluid power switching applications.Understanding the complex relationship between friction and component properties is challenging.Experimental descriptions are often infeasible due to time and cost constraints.Therefore, an elastohydrodynamic lubrication (EHL) simulation model, the ifas-DDS, was implemented to calculate the friction in translational pneumatic seals [1].While the EHL simulation provides an accurate solution to the underlying partial differential equations, it is computationally expensive.An approach to obtain an accelerated solution is the use of neural networks (NN).However, their main disadvantage is that they do not necessarily embed the physical mechanism underlying a particular dataset.In recent years, a new form of NN has emerged, the physics-informed neural network (PINN), which is imposed with physical laws, increasing the algorithm's accuracy.This approach offers several possibilities, e.g., extrapolation and more robust training.In this paper, a variation of the Reynolds equation, implemented in the ifas-DDS, is solved with a PINN.The complete hydrodynamic PINN (HD-PINN) framework is presented and compared to the ifas-DDS afterward.The results show the possibility of the HD-PINN framework for modeling the EHL with high speed, less to no accuracy loss, and minimal tuning effort.