New Horizons in Machine Learning Applications for Tribology

Tribology, the science of friction, wear, and lubrication, remains fundamental to modern engineering systems. From transportation and energy conversion to manufacturing and biomedical devices, tribological processes strongly influence efficiency, reliability, and sustainability. However, the underlying phenomena are inherently complex. Tribological behavior emerges from nonlinear interactions between materials, interfaces, surface topography, lubricants, operating conditions, and environmental factors across multiple spatial and temporal scales. Consequently, tribology has historically relied heavily on experimental investigations and empirical models.

In recent years, machine learning (ML) and artificial intelligence (AI) have begun to transform this landscape [1]. By enabling the extraction of patterns from large and multidimensional datasets, ML methods offer new opportunities to complement traditional physics-based approaches and to address long-standing challenges in tribological research. The first two Special Issues in Lubricants on ML in tribology demonstrated the breadth of possible applications and highlighted the growing interest in this interdisciplinary field [2,3]. The present third edition, entitled “New Horizons in Machine Learning Applications for Tribology”, further reflects the rapid expansion of research at the interface between tribology and data science. The contributions collected in this Special Issue illustrate how ML approaches are increasingly being used not only for predictive modeling, but also for condition monitoring, automated data analysis, and the integration of physics-based knowledge into data-driven frameworks. Together, these works provide insight into the evolving role of ML within tribology and its potential to accelerate discovery and improve engineering applications.

A recurring theme in several contributions is the prediction of tribological behavior using data-driven models. Predictive modeling is particularly attractive in tribology because friction and wear are emergent properties that depend on numerous interacting variables. Traditional analytical models often struggle to capture such complex dependencies, while ML methods can learn nonlinear relationships directly from experimental or simulation data. For example, Granja and Higgs [4] developed a deep learning-based model capable of predicting the coefficient of friction across the full Stribeck curve for lubricated tribo-pairs. By training a neural network on extensive ball-on-disk experimental datasets covering multiple materials, lubricants, loads, and temperatures, the model successfully captured transitions between boundary, mixed, and hydrodynamic lubrication regimes, demonstrating the ability of data-driven approaches to generalize across diverse operating conditions. In a related study [5], the same authors proposed a data-driven framework for simultaneously predicting friction and wear evolution over time in lubricated systems tested in a four-ball tribometer. Their results highlight the potential of deep learning to uncover correlations between lubricant formulation, operating parameters, and tribological performance that are difficult to capture using conventional empirical models. Complementary work by Li and Guo [6] applied a Bayesian-optimized neural network to predict lubrication performance parameters such as film thickness and load-carrying capacity in hyaluronic-acid-based aqueous lubricants. By integrating operating parameters and fluid properties into a machine learning framework, the model provides a computationally efficient tool for exploring lubricant formulations and operating conditions in complex lubrication systems. Similarly, Kolev et al. [7] combined tribological experiments with a convolutional neural network to predict friction behavior in chromium-coated SnSb11Cu6 alloys. Their work demonstrates how ML can complement experimental materials research by identifying the relative importance of microstructural parameters—such as coating hardness—for tribological performance.

Beyond the predictive modeling of friction and lubrication phenomena, another major focus of this Special Issue is the application of machine learning for condition monitoring and fault diagnosis in tribological components. Bearings, for instance, are among the most critical elements in rotating machinery, and their failure can lead to severe economic losses or safety risks. Several contributions address this challenge using advanced deep learning architectures. Sun et al. [8] proposed a hybrid model combining variational mode decomposition with convolutional neural networks (CNNs) and bidirectional long short-term memory networks (LSTMs) for rolling bearing fault diagnosis. By incorporating an improved optimization algorithm, the proposed framework achieved very high diagnostic accuracy on benchmark datasets. Shen et al. [9] further advanced this area by developing a CNN–LSTM-based approach optimized through a fruit fly optimization algorithm to predict the remaining useful life of rolling bearings. Their results demonstrate improved accuracy and robustness compared with existing approaches, highlighting the potential of hybrid deep learning models for predictive maintenance applications. Addressing the common challenge of limited labeled data and varying operating conditions, Li et al. [10] introduced a transfer learning-based method that integrates simulation-generated vibration signals with real-world measurement data for bearing fault diagnosis. Their multi-adversarial domain transfer framework significantly improves diagnostic performance across different operating conditions, illustrating how simulation and data-driven techniques can be combined effectively.

In addition to monitoring machinery components, ML also shows strong potential for accelerating tribological data analysis and interpretation. Tribological experiments often generate large volumes of time-series data and surface characterization images that require extensive manual evaluation. Zhao et al. [11] demonstrated how LSTM networks can automatically identify stationary phases in tribological experiments. This capability is particularly valuable because tribological measurements frequently include running-in periods that may distort the calculation of characteristic friction and wear parameters if not properly identified. In another study, Zhao and Lin [12] presented a deep learning-based image segmentation approach using a U-Net architecture for the automated analysis of worn surfaces in polymer-based composites. By enabling pixel-level classification of wear features, this method offers a powerful tool for objective and high-throughput characterization of tribo-stressed surfaces.

Finally, several contributions highlight the increasing interest in hybrid approaches that combine machine learning with physical models—an important step toward improving interpretability and robustness. Saleh et al. [13] presented a physics-informed neural network (PINN) framework capable of predicting pressure and film thickness distributions in plain bearings in real time. By embedding physical constraints derived from lubrication theory into the learning process, PINNs offer improved generalization and reduced dependence on large training datasets compared with purely data-driven models. Similarly, Pfitzer et al. [14] demonstrated how machine learning can support parameter estimation in dynamic friction models for ball joints used in vehicle suspension systems. Their work shows that combining phenomenological friction models with ML-based parameter identification can enable efficient and real-time capable simulations for engineering applications.

Taken together, the contributions to this Special Issue clearly illustrate the expanding role of machine learning in tribology. The presented studies range from purely data-driven prediction models to hybrid frameworks that integrate physical knowledge, as well as applications in predictive maintenance and automated data analysis. This diversity reflects the growing maturity of the field and the increasing availability of tribological datasets suitable for machine learning.

Looking ahead, several challenges remain. The development of standardized datasets, improved data sharing, and the implementation of FAIR data practices will be essential to fully unlock the potential of machine learning in tribology. In this context, norm-setting institutions will play a crucial role in establishing frameworks that ensure data quality, interoperability, and long-term accessibility, while also recognizing prior scientific contributions and ensuring appropriate attribution. Moreover, integrating physical knowledge into machine learning frameworks will remain a key research direction to ensure model interpretability, reliability, and extrapolation capability. As these challenges are addressed, machine learning is likely to become an integral component of tribological research and engineering design.

The Guest Editors would like to express their sincere gratitude to all authors who contributed to this Special Issue and to the reviewers for their valuable comments and constructive feedback. We also thank the editorial staff of Lubricants for their continuous support throughout the publication process.