The logical framework of artificial intelligence was outlined by Alan Turing, the cracker of the German Enigma code in World War II, in a 1950 paper entitled “Computing Machinery and Intelligence,” in which he discussed how to build thinking machines and how to test their intelligence. It took almost 75 years to develop programs such as ChatGPT that are able to pass the “Turing test” of exhibiting intelligent behavior equivalent to, or indistinguishable from, that of a human.

Part of the growth in machine learning has sprung from the vast amount of information stored on the world wide web, which provides deep-learning models with data in an easily digested form.

Arguably tribology-related industries are ideally placed to benefit from machine-learning approaches, because they depend on a large array of products that have to be tailored to a variety of applications, from industrial machines and engines of all types to high-technology devices that have to function in all environments. Very often the tribological solution to a particular problem—a new coating, a new lubricant, a new protocol—is selected based on experience from an array of existing solutions.

Machine-learning approaches offer the promise of a completely new way of solving tribological problems and designing new products by analyzing databases of tribological properties to discover and predict new materials for a particular application.

In this column, we have already discussed ways in which machine-learning can be used to design new, lubricious 2D materials,¹ to propose novel viscosity-index improvers² or to speed up calculations of complex tribological interfaces.³ New STLE technical tracks on artificial intelligence and machine learning were inaugurated at the 2024 STLE Annual Meeting in Minneapolis, Minn.

To realize this potential transformation in the way in which we might discover new lubricants, new coatings, new antiwear additives (and on and on…), we will have to rely on access to data that can be read by machines. For programs like ChatGPT, such data were made readily available by the growth of the internet, where the information is stored in an already readable form. However, this aspect of readability represents a significant challenge for tribology-related machine learning.

This very problem has been addressed by STLE member Nick Garabedian and coworkers of the Institute for Applied Materials of the Karlsruhe Institute of Technology in Germany, who have described a method of storing data in a “FAIR” way so that it is Findable, Accessible, Interoperable and Reusable so that it can be used For Artificial Intelligence Research. 

A second problem is harder; accessing tribological data. Although machine learning can make predictions using small sets of data, it learns best the more data it has and the more varied those data are. While universities and national laboratories do generate tribological data, probably the largest repository of high-quality data is sequestered in industrial laboratories around the world, who may, of course, be loath to share it—although an unscientific poll at the 2022 STLE Annual Meeting in Orlando, Fla. suggested that some companies may be willing to do so. Some larger companies, on the other hand, may feel that they have enough data to go it alone. 

The growth of science in the 17th century was spurred on by the creation of nonprofit scientific societies as venues for sharing and disseminating new results and ideas; societies such as STLE could provide a similar 21st century role as a repository for tribological data, but how can companies be persuaded to store their hard-won data there? As a way of demonstrating the use of such shared data repositories to the community (both industry and academia), they could start by sharing low-value and easy-to-reproduce data. Additional data could be generated, for example, by round-robin competitions. Each organization that agrees to deposit their FAIR results into the STLE data bank would have access to all the data available there. Perhaps further access options might be negotiated for a fee, which could then be used as a credit for anyone that publishes machine-learning models providing new insights or leading to the development of new products. The goal is for the combined efforts of all participants to grow over time, resulting in the creation of new shared data and models, thereby creating confidence in the data repository and a growing realization of the importance and benefit of machine learning to the tribology community.

There are very likely to be other ways of developing such data banks, but however they arise, they will be of significant benefit to the companies and organizations that participate, as well as their national economies.

FOR FURTHER READING
1. Tysoe, W. T. and Spencer, N. D. (2020), “Designing lubricants by artificial intelligence,” TLT, 76 (6), p. 78. Available here.
2. Tysoe, W. T. and Spencer, N. D. (2020), “Designing lubricants by artificial intelligence—the sequel,” TLT, 76 (8), pp. 68-70. Available here.
3. Tysoe, W. T. and Spencer, N. D. (2023), “Turning simulations into formulae by machine learning,” TLT, 79 (2), p. 86. Available here.
4. Garabedian, N. T, et al. (2022), “Generating FAIR research data in experimental tribology,” Scientific Data, 9, 315.