Geometry-Aware Physics-Informed Neural Network: A Framework for Solving the Reynolds Equation with Starvation for Lubricated Contacts

Abstract

Machine learning has gained increasing attention in fluid power applications, particularly in condition monitoring and control.A promising development in this field is physics-informed machine learning (PIML), and specifically physics-informed neural networks (PINNs), which integrate governing physical laws directly into the training process.Unlike traditional networks, PINNs do not require data collection and can significantly accelerate computationally demanding simulations.This capability is valuable for studying lubricated contacts, such as seals in pneumatic valves, which operate pneumatic systems by controlling flow rate and pressure, and where accurate yet efficient modeling is essential.Seals are critical for maintaining pressure and preventing leakage; their malfunction can lead to costly failures, downtime, and environmental risks.However, seal behavior remains difficult to characterize due to the coupled hydrodynamic (HD) and elastodynamic (ED) effects within the sealing interface.Conventional elastohydrodynamic lubrication (EHL) simulations based on fluid-structure interaction models provide accurate insights.However, they are computationally expensive, limiting their usability for real-time or large-scale investigations.To address these challenges, this work extends a validated PINN framework for modelling lubrication behavior inside a pneumatic seal by introducing an unsupervised learning autoencoder (AE) to compress complex gap geometries into low-dimensional latent representations.The framework, consisting of a PINN, a loss balancing, and a hyperparameter tuning algorithm, has already been validated for modelling hydrodynamic lubrication simulations with transient phenomena and cavitation.The addition of the AE to the framework yields a geometry-aware PINN (GAPINN) capable of fast, accurate, and geometry-adaptive analysis of sealing systems, thereby enhancing the applicability of PIML in tribological research.

Authors 5

1. Faras Brumand‐Poor Aachen Institute for Fluid Power Drives and Systems ifas

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University , Institute for Fluid Power Drives and Systems (ifas

2. Artem Konstantinov Aachen Institute for Fluid Power Drives and Systems ifas

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University , Institute for Fluid Power Drives and Systems (ifas

3. Darren Lam Ming Hui
4. Jakob Wirtz Aachen Institute for Fluid Power Drives and Systems ifas

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University , Institute for Fluid Power Drives and Systems (ifas

5. Prof. Katharina Schmitz Aachen Institute for Fluid Power Drives and Systems ifas

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University , Institute for Fluid Power Drives and Systems (ifas