TriboAI: Making Language Models Read Friction Curves

Introduction

Artificial intelligence (AI) is becoming an increasingly valuable tool across engineering disciplines, and tribology is no exception. As researchers generate larger and more complex datasets from friction and wear experiments, interpreting these results consistently and efficiently has become a growing challenge. During the TriboNet webinar, “TriboAI: Making Language Models Read Friction Curves,” speakers Iñigo Llavori and Xavier Borras presented an innovative approach that combines physics-based simulation with large language models (LLMs) to help researchers better understand friction curves.

Rather than replacing tribologists, the goal of TriboAI is to support them by providing intelligent assistance during data analysis. By combining synthetic datasets, deterministic physics, and natural language processing, the project aims to improve the interpretation of friction signals while making tribology data more accessible and consistent.

Why AI Matters in Tribology

One of the key messages from the webinar was that tribology generates an enormous amount of valuable experimental data, but much of it is difficult to use for AI applications because it lacks proper annotation. Researchers often rely on expert knowledge to interpret friction curves, making the process both time-consuming and subjective.

To overcome this limitation, the presenters proposed generating synthetic friction data using validated physical models. Because every simulated friction curve is created using known parameters, each dataset automatically includes its ground truth. This makes it possible to train AI models without requiring experts to manually label thousands of experimental results.

This approach reflects the ideas discussed in TriboNet’s article, Need for AI in Tribology, which explains how artificial intelligence can help engineers manage increasingly complex tribological data while complementing human expertise rather than replacing it.

Simulating Friction Curves Through Physics

A significant part of the webinar focused on the physics-based simulator developed by the research team. Instead of attempting to solve every aspect of contact mechanics, the simulator is designed to reproduce the behaviour observed during reciprocating friction experiments. By incorporating established tribological principles into the model, researchers can recreate realistic friction curves without relying solely on experimental measurements.

The simulator captures several key tribological phenomena, including static-to-kinetic friction overshoot, contact compliance, abrasive wear, Hertzian contact behaviour, sensor noise, and changes in contact geometry during sliding. Because each of these effects can be adjusted independently, researchers can investigate how individual physical mechanisms influence the overall shape of a friction curve.

During the demonstration, the speakers showed how enabling or disabling certain parameters produced noticeably different friction responses. For example, removing abrasive wear kept the contact area nearly constant throughout the simulation, while increasing sensor noise affected the clarity of the measured signal. Such flexibility allows researchers to explore different experimental scenarios without performing time-consuming laboratory tests.

Although the current version primarily focuses on abrasive wear under quasi-static conditions, the presenters explained that the modular framework allows additional physical models to be incorporated in the future. Dynamic loading, temperature effects, lubrication behaviour, and other wear mechanisms could eventually become part of the simulator, making it an even more comprehensive tool for tribology research.

Synthetic Data as the Foundation for AI

One of the webinar’s central ideas was that physics-based simulation provides an effective solution to one of AI’s biggest challenges in tribology: the lack of labelled data.

Unlike image recognition or natural language processing, tribology datasets rarely come with detailed annotations describing exactly what each friction curve represents. Producing these labels manually requires extensive expertise and considerable time. By contrast, every synthetic curve generated by the simulator already includes its complete set of physical parameters, meaning the “correct answer” is known from the moment the data is created.

The presenters explained that this makes synthetic datasets particularly valuable for training machine learning models. Millions of friction curves can be generated under different operating conditions, providing AI systems with diverse examples before they are exposed to real experimental data. Experimental measurements can then be used to validate and refine the models, combining the strengths of simulation and laboratory testing.

Rather than replacing experiments, synthetic data accelerates AI development while reducing the dependence on costly manual annotation. This strategy provides a scalable pathway for developing intelligent tribology assistants capable of understanding increasingly complex friction behaviour.

Teaching Language Models to Read Friction Curves

One of the most exciting moments of the webinar was the live demonstration of the TriboAI assistant. Unlike conventional analysis software that simply calculates an average coefficient of friction, the assistant evaluates the entire friction signal and explains its observations using natural language.

Using a friction curve from the TriboNet Challenge, the assistant examined signal quality, sampling density, sensor utilization, friction loop morphology, and differences between forward and backward sliding. It identified features such as post-reversal overshoot, asymmetric friction plateaus, and variations between sliding directions that might otherwise be overlooked during routine analysis.

Instead of recommending a single coefficient of friction, the AI concluded that reporting only one value would oversimplify the experiment because the two sliding directions exhibited different friction behaviour. The assistant therefore suggested reporting separate values and explaining the observed asymmetry.

This demonstration highlighted how AI can move beyond simple numerical calculations to provide meaningful engineering insights. Rather than replacing expert interpretation, the system encourages researchers to examine their data more carefully and consider factors that may influence the validity of their conclusions.

AI as an Engineering Assistant

Throughout the presentation, the speakers emphasized that TriboAI is intended to support engineers rather than replace them. Instead of making decisions on behalf of researchers, the AI assistant acts as an intelligent reviewer that helps identify patterns, detect anomalies, and provide additional context for interpreting friction curves.

For example, if a friction curve exhibits strong asymmetry between the forward and backward sliding directions, the assistant can flag this behaviour and encourage researchers to investigate potential causes. These differences may arise from the tribological system itself, but they can also result from issues such as sensor calibration, specimen alignment, or inconsistencies in the experimental setup.

Similarly, the AI evaluates the quality of the recorded signal before drawing conclusions. Excessive noise, poor sensor utilization, or insufficient sampling density may indicate that the experiment should be repeated or interpreted with caution. Rather than simply reporting numerical values, the assistant provides explanations in natural language, helping researchers better understand why a particular dataset should—or should not—be trusted.

The broader importance of preparing tribology data for intelligent analysis is also explored in TriboNet’s webinar, Findable AI-Ready (FAIR) Data in Tribology, which discusses how well-organized and standardized datasets can support the next generation of AI tools in tribology.

Discussion and Future Development

The webinar concluded with an engaging discussion that highlighted both the current capabilities of TriboAI and its future potential. Participants raised several questions regarding the expansion of the simulator beyond abrasive wear, including the possibility of incorporating dynamic loading, temperature effects, acoustic emissions, and additional wear mechanisms.

The presenters explained that the current simulator focuses primarily on quasi-static reciprocating friction experiments because they provide a well-defined starting point for generating synthetic datasets. However, the modular architecture of the system allows new physical models to be added over time as research progresses.

Several attendees also suggested integrating TriboAI directly into tribometers. Instead of analysing data only after a test has finished, the assistant could monitor friction signals in real time, alerting researchers to abnormal behaviour, unexpected changes in friction, or potential experimental errors. Such capabilities could reduce testing time, improve data quality, and prevent researchers from spending valuable resources on invalid experiments.

The speakers also emphasized that synthetic data should complement, rather than replace, experimental measurements. Laboratory testing remains essential for validating simulation models and ensuring that AI systems continue to improve as new data becomes available.

Beyond Friction Curve Analysis

Although TriboAI currently focuses on interpreting friction curves, the concepts presented have much broader implications for tribology. Artificial intelligence has the potential to assist researchers throughout the engineering workflow, from selecting materials and designing lubricants to planning experiments and analysing test results.

TriboNet has previously explored these broader applications in Can We Select the Desirable Material Using AI?, which discusses how artificial intelligence can support material selection, and Can We Formulate Lubricants Using AI?, which examines AI-assisted lubricant formulation and optimization. Together with TriboAI, these developments illustrate how AI is gradually becoming an integral part of modern tribological research.

Conclusion

The TriboNet webinar “TriboAI: Making Language Models Read Friction Curves” demonstrated how the combination of physics-based simulation and artificial intelligence can reshape the way friction data is analysed. By generating synthetic datasets with known ground truth and training large language models to interpret friction signals, TriboAI offers a practical solution to one of the biggest challenges facing AI in tribology: the shortage of high-quality labelled data.

Perhaps more importantly, the webinar showed that AI should be viewed as a collaborative tool rather than a replacement for engineering expertise. By helping researchers assess signal quality, identify meaningful friction behaviours, and interpret experimental results more consistently, AI assistants like TriboAI can improve both the efficiency and reliability of tribological analysis.

While the project is still evolving, its early results demonstrate significant promise. As additional physical models, experimental datasets, and machine learning techniques are incorporated, TriboAI has the potential to become an indispensable companion for researchers and engineers working to better understand friction, wear, and surface interactions. The webinar provided an exciting glimpse into a future where human expertise and artificial intelligence work together to advance the science of tribology.