Machine Learning-Based Approach for Fast Prediction of Transient Tribo-Dynamic Behaviors in Lubricated Systems

Abstract Simulating transient tribo-dynamic problems remains challenging for real-time predictions of transient tribo-dynamic behaviors in lubricated systems due to high computational demands. This article presents an approach based on the machine learning framework to address this issue by developing a time-dimension involved neural network that decouples time-accumulated transient effects from steady-state characteristics using steady-state friction datasets. The proposed approach was applied to two classic tribo-dynamic problems, including the ring–liner conjunction case and the journal bearing case under engine-like condition. It appears that the proposed approach achieves below 0.1% average absolute percentage error in predicting results. The proposed approach is over 70–200 times faster than existing numerical modeling techniques, enabling real-time performance assessment of tribological systems. This approach offers a practical solution for rapid engineering evaluations where traditional methods prove computationally prohibitive.