In this review, we discuss recent progress in tribological research driven by data science. First, we examine the advancements and challenges of industrial AI technologies, such as edge computing and smart factories. We then focus on discussions at international tribology conferences regarding data processing and sharing. Next, we report on an international joint research project on detecting surface roughness and addressing the challenges associated with data sharing. Finally, we highlight the potential of applying generative AI and federated learning techniques to tribology.

View full abstract

The application of AI (Artificial Intelligence) and ML (Machine Learning) in tribology is challenging due to the complexity of friction phenomena, which involve complex physical and chemical processes, making data handling and interpretation difficult. Nevertheless, numerous studies, including those using advanced AI techniques, have been reported for individual phenomena. While reports in Japan are increasing, they remain limited. This article introduces review papers from overseas to provide readers with a comprehensive overview of utilization methods, presents case studies in hydrodynamic lubrication analysis, and discusses future prospects.

View full abstract

This paper introduces fundamental concepts and representative applications of artificial intelligence (AI) as an emerging tool in the field of tribology, with a particular focus on maintenance tribology. Emphasizing the “offensive” approach of the author’s research, four AI-based methods used in the laboratory are presented as part of the “Tribology × AI” framework. These methods contribute to the development of new perspectives and problem-setting strategies in tribological studies. Specifically, rule learning is applied to the abnormality detection and condition monitoring of sliding surfaces, while classification, clustering, and regression are utilized for that of lubricants. Among these, rule learning and clustering represent unsupervised learning techniques, whereas classification and regression fall under supervised learning. By enabling the rapid processing of complex, high-dimensional data and the accurate identification of patterns, AI significantly reduces the effort required for data analysis, organization, and interpretation. This accelerates the overall pace of tribological research and opens new avenues for exploration.

View full abstract

With the recent advancements in machine learning, there have been increasing reports on “tribo-informatics,” a field that predicts friction and wear properties by applying machine learning to vast amounts of historical friction test data. The utilization of tribo-informatics is expected to remarkable enhance the efficiency of developing sliding materials. This review article focuses on the development of sliding materials and introduces research trends in tribo-informatics related to this area. Furthermore, the article highlights the authors’ efforts to predict the average friction coefficient and adhesion height for metallic materials under high-temperature conditions, presenting past results, current challenges, and prospects.

View full abstract

Attempts are underway to evaluate the lubrication performance and friction and wear mechanisms of lubricant additives using deep learning and machine learning. Feature importance by decision tree models, Grad-CAM in CNN, and dimensionality reduction are methods to visualize the learning process of machine learning models, and are expected to be useful for evaluating the correlation between chemical and geometric information on the friction surface and the friction coefficient. In addition, by combining deep neural network and LIME, one of the explainable AI, it is possible to analyze the correlation between the molecular structure of antiwear additives and antiwear performance.

View full abstract

Recent advances in artificial intelligence (AI) and machine learning (ML) have rapidly accelerated their application in tribology. While initial studies primarily focused on predicting friction and wear from limited parameters using simple algorithms, recent research has explored more complex data-driven approaches. Additionally, unique applications have emerged, such as AI-controlled morphing surfaces that actively avoid damaged areas in contact interfaces. Despite these achievements, challenges remain, especially in model interpretability and bridging the gap between AI specialists and tribologists. The field is now moving toward multimodal AI frameworks that combine time-series sensor data, images, acoustic signals, and chemical analyses. Future trends also include hierarchical integration using advanced algorithms such as Transformers and graph neural networks, as well as the adoption of explainable AI (XAI) and causal inference. Such innovations are expected to enhance our understanding of the physical and chemical mechanisms underlying friction and wear, thereby enabling the development of advanced materials and active tribological systems.

View full abstract

In recent years, deep learning has rapidly gained attention in the field of tribology due to its potential. In this study, we apply convolutional neural networks (CNN) to estimate friction forces directly from high-resolution SEM and optical images of polymer friction interfaces. Using Grad-CAM, we visualize the image regions contributing most significantly to the CNNʼs predictions, revealing that CNNs tend to focus on surrounding structures rather than the contact center. Furthermore, we explore friction interface image generation under unobserved conditions using continuous conditional Generative Adversarial Network (CCGAN) and Stable Diffusion. These generative models allow us to produce interface images for unseen frictional forces, some beyond the experimental range, and confirm their plausibility using independently trained CNNs.

View full abstract

The advancement of artificial intelligence (AI) is dramatically transforming molecular simulation technologies, paving the way for significant breakthroughs in tribology. Traditional molecular simulations, such as quantum and classical molecular dynamics, face limitations in calculation speed and force-fields description when applied to complex tribological phenomena. This paper introduces neural network potentials, particularly graph neural network-based Preferred Potential (PFP), which greatly enhance simulation capabilities by learning huge amount of the first-principles calculation results. These AI-driven models allow for rapid computation, effectively bypassing the cumbersome construction of complex potentials and significantly speeding up calculations - by more than ten thousand times faster than conventional first-principles calculations. The paper discusses successful applications of PFP in simulating the adsorption and reaction dynamics of zinc dialkyldithiophosphate (ZnDTP) adsorbed on iron surface, and the interactions of refrigerants on metal surface. Furthermore, it highlights the use of PFP-generated data in materials informatics for screening base oils, illustrating AI’s potential to generate high-accuracy datasets for innovative material discovery. The paper also addresses the need for scalability in simulations through LightPFP technology, which allows handling of much larger atomistic models by distilling PFP-generated data into lightweight machine learning models. This technological leap enables the analysis of dynamic behaviors in complex systems with hundreds of thousands of atoms. Overall, AI model-based molecular simulations not only enhance efficiency and versatility in existing studies but also open new avenues for research across various material domains, setting the stage for broader industrial applications.

View full abstract

In electric railways, current collecting materials such as contact wires and contact strips are subject to significant wear due to Joule heat. Previously, the authors proposed a wear mode map that quantitatively shows the transition condition of four wear modes on the basis of steady state heat conduction analysis of electrical potential and temperature distribution around a contact. However, the map used the φ‐θ theory that can only be applied in the steady state, so its scope of application is limited to stationary contacts or sliding contacts at very low speeds. In this paper, the authors develop an unsteady heat conduction analysis model to adapt the wear mode map to high-speed conditions, and propose non-dimensional parameters that generalize the transient characteristics of temperature rise. Then the authors propose an unsteady wear mode map and identify the parameters that govern the wear mode transition phenomena in high speed conditions. Finally, the validity of the unsteady wear mode map is confirmed by explaining the measured contact voltage range of molten bridges generated in an experiment using the map and transfer materials on the wear surface.

View full abstract

It has been reported that combining phosphorus-based (P-based) and sulfur-based (S-based) additives enhances anti-wear performance compared to using each additive individually. In this study, we proposed a comprehensive method for optimizing additive concentrations through friction tests using an additive supplementation system. This approach enables efficient determination of the optimal combination of P- and S-based additives. Additionally, we conducted in-situ damage evaluation by monitoring acoustic emission (AE) signals. Based on surface observations and AE measurements, we investigated the effects of additive concentration on friction and wear behavior, as well as the underlying mechanisms. AE energy in the 0.131–0.141 MHz range correlated with the friction coefficient, while energy in the 0.209–0.219 MHz range was associated with wear, suggesting that surface roughness changes may influence AE responses. For the sample with S added before P (S + P), the initial surface roughening caused by the S-based additive promoted wear, which continued until a sufficient amount of P-based additive was introduced. The P-based additive then formed a protective film, suppressing further wear. In contrast, for the sample with P added before S (P + S), a smooth and durable P-based film formed initially and remained stable even after the addition of the S-based additive.

View full abstract

[in Japanese]

View full abstract