October 2024 - Artificial Intelligence and Machine Learning in Tribology Update

For our October episode we are following up with instructors from the Artificial Intelligence and Machine Learning education course from our 2024 STLE Annual Meeting in Minneapolis.   Returning from that course is Max Marian of Leibniz Universität Hannover, Nick Garabedian of Karlsruhe Institute of Technology, and Omar Zouina of Karlsruhe Institute of Technology moderates as we discuss updates to this rapidly evolving field and it's advancements related to Tribology.  

If you want to hear more about how Artificial Intelligence and Machine Learning play a role in technology, consider signing up for our education course at our Annual Meeting in Atlanta on May 18, 2025.   AI-ML is just one of the many courses we will be offering this year in addition to over 100 technical sessions, panel discussions, and numerous networking opportunities.  Registration opens in early November, but for more information, click here: https://www.stle.org/AnnualMeeting/call .

Meet our speakers:
Omar Zouina, 2nd year PhD candidate at the Karlsruhe Institute of Technology, Institute of Applied Materials (IAM) and Microtribology Center, my research focuses on solid lubricants particularly graphite with an emphasis on the lubrication mechanisms under high mechanical loads for rolling bearings applications.

Max Marian is a professor specializing in tribology and machine design. His research focuses on improving energy efficiency and sustainability through surface modification techniques. He has published extensively, received numerous awards, and is involved in several professional organizations. 

Dr. Nick Garabedian works at the intersection of linked data engineering, experimental tribology and machine learning. He is currently the CEO of datin – a newly-founded software startup for research data management and sharing.

For more information on STLE, please visit https://www.stle.org/ If you have an idea for our podcast, or interested in being a guest, please Email STLE Director of Professional Development Robert Morowczynski at rmorowczynski@stle.org . Also, we love your feedback, please take a minute to provide us with your thoughts at Perfecting Motion Podcast Feedback.

Transcript

Unknown Speaker  0:00  
Music.

Speaker 1  0:08  
Welcome to perfecting motion, an STLE podcast series that talks with members and industry professionals about current issues and trends impacting the global tribology and lubricants community.

Speaker 2  0:22  
Hi and welcome to our October episode of the STLE Perfecting Motion podcast. This month's episode follows up on our first ever artificial intelligence and machine learning education course at the STLE 2024, annual meeting. Today we have with us two professionals who worked in that course and moderating this discussion will be Omar Zwina, a PhD student in Germany. Omar, I'm going to hand it over to you as you moderate this wonderful discussion.

Speaker 3  0:49  
Thank you, Robert for the kind introduction. So Hey everyone. Welcome to this month's episode of the STLE podcast, perfect in motion, where we discuss the trends, aspects and issues in the science of friction, wear and lubrication. As I was introduced, I'm Omar your host for today, I would like you to join me on the right where we explore how AI and machine learning can reshape the way we asteroids work with our systems and push the boundaries of innovation. Joining me today are two well known scle members, Professor, Professor Max Marian and Dr Nick Garabedian, who are driving the future of tribology through data and machine learning. Personally, I would like to introduce them as the pioneers of such a trend. They've both made significant contributions to how we implement structured data and AI potentials in the fascinating field of Tribology. Thanks a lot to you for taking the time. Great to have you here. And I would like to give the floor to our guests who introduced themselves, their background and how they got into tribology in the first place. Nick, you want to start

Speaker 4  1:54  
Sure? Yeah, so my name is Nick Garabidian. I'm a mechanical engineer by training. I did my PhD at the University of Delaware with Professor Dave Burris in nanotribe biology. But after I finished my PhD in 2019 I took a sharp pivot towards data science when I came for a postdoc at the Carroll Institute of Technology in Germany. And yeah, I've been doing data science in tribology for the past five years, and since the beginning of October 2024 I have been working on establishing a data science company. Yeah.

Speaker 5  2:30  
And also thanks from my side for Bob and Omar for the kind introduction and rolling us out the red carpet. That's a big introduction you gave us. I hope we can somehow fill some of it. Yeah, my name is Max Marion. I'm currently a professor at Leibniz University, Hanover, as well as at Pontificia university, at Catholic Chile. I'm also a industrial engineer, actually by training, and probably as many of us came into tribology by pure accident without knowing a lot what's behind the word during the studies? And yeah, stumbling into into a student assistant job, and then somehow never leaving again, to the topic of tribology, sticking to the field. And after completing my PhD and all my studies and PhD in Erlang, and I moved to chitlin. And yeah, nowadays I'm in Germany. Originally, I was working in simulations, also in experiments, and then encountered the world of machine learning or some optimization tasks I had to do that involved a lot of numerical efforts and simulations. So that's what brought me into this world.

Speaker 3  3:35  
Thanks guys for sharing your personal experiences. Those are very nice and true for both of you. Now I know that, and it goes without saying that for the past few years, AI has been really big and massive topic. So I would like to kick off this discussion with asking you both, which AI development would you guys say is your favorite, and also which one was the most impactful? I would say, Yeah,

Speaker 4  4:06  
I have to agree with Max the red carpet that you rolled out. Yeah, certainly something to feel. So we hope we do that. But yeah, so in terms of AI developments and favorite AI developments, I would actually go a little bit away from the r, d domain. And I will say that the general AI awareness in society and in the world has been very exciting to see. For example, it gives me a good feeling as a scientist to know that, for example, the World Economic Forum and Microsoft are ranking AI for science. It's one of the probably most long term beneficial fields for AI application, apart from that, regulatory bodies engaging with AI and trying to figure out how to put some boundaries in the usage of AIS for the benefit of society, is also refreshing to see and another. Other interesting thing that I saw recently was that Cloudflare is now introducing a system that will charge scraping bots, those that train the large language models, they will charge them when they take some content from anyone's page. And that sort of economy that involves regulation, that involves distributing payments and charging for the things that happen in the AI ecosystem is very, very encouraging, so I can't wait to see what will happen in research and development, and especially in trade biology. Max, yeah,

Speaker 5  5:34  
exactly as you're saying, like, I think there is a lot of going on in artificial intelligence nowadays, and there's some impactful AI developments, especially in the recent years. And, yeah, it's even more accelerating every day for me, I think there are a couple of trends or aspects that that stand out, also that go beyond our day to day scope of research. First of all, I guess we all know the new trend of generative models like chat, GDP or Delhi, that has changed the way how we create or able to create content, enabling the production of text, images and other things with a rather minimal input. That, of course, also brings challenges when we're thinking about educational terms, but also a lot of potential. For example, like with chatgpt or Delhi, we can generate images from from just a description. And of course, this also opens new avenues for US researchers in transmitting our ideas and concepts and also the results, and also from my personal experience. Ai, such as chat, GDP, can also be used to generate content surprisingly current and relevant. And I personally also often use it for brainstorming or maybe also sometimes overcoming some creative blocks. And also a second trend that also Nick already mentioned is natural language processing that has improved how we interact with machines talking about voice assistance, chat bots ever becoming more intuitive and responsive, and maybe it also names us in the future translation services more easily, so maybe it helps us break down language barriers. And I think that the natural language models have the potential also to some sort of democratized knowledge as it gets more accessible from all around the globe, which is a great trend. And one last aspect that stands out for me is that hasn't come that much yet into tripology, but I think it will in the next years. Is reinforcement learning. This already has to lead to significant advancements in gaming and in robotics and AI agents have mastered games like go or StarCraft, showcasing the ability to master complex strategies. For example, in AlphaGo defeated the current gold champion, which is, I don't know if everyone is familiar with it, but it's basically a a famous board game which is considered to be a little bit or even more complex than than chess in terms of of strategy. So this happened that the first AI defeated a human champion in that game in 2015 16. And interestingly, it is also somewhat coincident with the rapid increase of publications we can see in the area of tribology and many other areas. So it kind of took off from that year on, showing the the potential to not just keep up, but also overcome the human brain. And beforehand, were tasks that have been considered impossible. So I think we're now able to build AI that can solve problems, and not just in games, but also nowadays in real world applications.

Speaker 3  8:31  
I'm gonna pick up on this last point, actually. So you said that AI will have very good use cases, especially in science, but we know that tribology is more like a melting pot of multiple sciences brought together in one place, and it's multidisciplinary per se. Now, how do you think AI will influence such science like tribology, and this is more or less like, also like a personal question to you both is like, how did you get into AI as tribal as Max mentioned it a bit more in his introduction, but who would like more tea about this? It's a very good point. I

Speaker 5  9:12  
guess, as you mentioned, the interdisciplinary nature of tribology is the big challenge of introducing AI in that sector, but also the big potential to help us and solve the issues, because tribology itself has a lot of influencing factors from various physical domains, chemistry and so on. And if that wouldn't be enough, we're dealing with loss variables. So it's not like material properties, where we can replicate some tests. When we have done a test, it's done, and we cannot replicate it one by one 100% and that's certainly one of the challenges that we face in AI. So we're dealing with material loss, with energy losses and so on. And there's so many influencing factors that even the most well known and experienced topologists somehow get surprised by what factors actually can impact. Effect or experiments, or also what has to be considered in simulations, as mentioned previously, during my PhD, I worked in the optimization of micro textures for lubricated contacts. And also in these kind of contexts, we have various physical domains that we have to cover. We have the rheology of the fluids. We have the elastic material properties. We have thermal energy aspect and so on that we have to consider in the simulation. And if we want to optimize a complex geometry, that could be whatever we can imagine of a geometry, and if we want to find an optimum of how to design these textures, we, on the one hand, have an optimization problem of given parameters. Or on the other hand, maybe we have the aspect of design innovation, so maybe we can find an Optima that is beyond the mathematical description that we have thought about beforehand. And this task is exactly what brought me into AI in order to appreciate AI, in order to optimize text to geometry for a given load case for a given application. And so then what brought me into the topic, and when we analyze the literature that has been published on AI, this is also one of the most common aspects optimization of systems. On the other hand, we also have the mechanism analysis. So AI can help us in tribology, how lubricants, how friction materials work. So what is the mechanism behind in the topology processes? Then we also have the aspect of status or condition monitoring, where AI can help us identifying trends or features in signals and extractor the relevant information to predict the behavior, or maybe also to diagnose conditions. And then on the other end, that goes hand in hand with that, we have the behavior prediction of running systems in real time. So that's, I think, one of the major challenges, and it does not only cover experimental work, but also can rely or build upon existing simulation tasks. And I'm

Speaker 3  11:56  
going to hand over right now to Nick so like to hear from you?

Speaker 4  12:00  
Yeah, I guess there is quite a bit of overlap of my story with what Max said. I didn't do any simulations during my PhD, though, so it's a bit different. I broadly divide the reasons why I got into AI into two categories. One was purely scientific, and the other one was more high level of how my PhD gets applied in industry, and it's interesting, because the expected outcome of my purely experimental PhD is not to go on to found an AI company. You know, I wasn't trained to become a expert in data science, so it's an interesting pivot. And I've often thought about why that happened in my career. And the first reason, the scientific one, was because, at the end of my PhD, I was, you know, I was happy with the outcomes of my work. I was trying to conduct experiments in force and size scale that we would consider to be in between the nanotribology and the macro microtribology world, trying to connect those two worlds. And I was fascinated with how many gaps there are in Tribology. So I was trying to figure out what is one of the easiest ways to get to a so called tribology equation, meaning have a some sort of equation that you plug in. This is my system. These are the operating conditions. This is what friction I would like to have out of it, and it will give you the properties of the components and the materials and so on. So more like a brain that thinks tribology, and sort of, you know, a few years after I had that dream chat, GPT came out, and now I guess everybody can associate something in their life with what I'm saying. So that was the broad scientific curiosity that I had. And, you know, I realized, Hey, I could do that. And that was the first reason. The second reason was that, again, at the end of my PhD, I'm an engineer, and I was doing science, and I was impatient to see some of my developments go into practice in the industry. And I read somewhere, I don't know how much we can trust that, but it said something like Apple, the company, can bring things from R and D into practice within a year, and in academia, that takes at least 10 years, or something like that. So I was thinking about that, and I was getting a little concerned, or I was seeing the potential of all these developments that happen in R D to be applied in industry, and I was looking for a vehicle to translate that knowledge into practice. So imagine having the R D fundamental scientists working on a problem that a particular end user is experiencing in the field, for example, the driver of particular person should influence the R and D direction of a company producing lubricants, for example, or producing gears. And you know, currently, to make that connection between these two worlds, you need to talk to a bunch of people who will reinterpret to each other what the previous. Has said, and I was looking for this connection between these two worlds. And AI, again, emerged this melting pot for that. So these are generally two directions that pushed me towards AI. And last but not least, I need to say tribology, to me, is a very fertile ground for AI innovation, much more than most R and D domains, because we are so interdisciplinary, and it's almost a no brainer to start working in AI, because it's hard to think about all the aspects of any tribological contact at the same time and to rely only on ourselves to get there. And I think AI can really make our lives a lot better. And Karen is our human creativity provide us with the tools to use our human creativity and come up with solutions to the greatest challenges of Tribology. Yeah, so

Speaker 3  15:52  
this is a very interesting series of points that you both have brought up with the discussion I would like to just get on the last point there. And before we move to another topic, we'd like to build up on it like, and just give you like, a quite straightforward question is, how would you sell AI to any tribologist? Because you mentioned plenty of things, but how would you break it down to few benefits that any tribologist will be just hooked, and then they will just say, Okay, I'm sold for AI usage in Tribology. The

Speaker 4  16:24  
example I give to people is simple. If you're an experimentalist and you run tribology experiments, you'll usually follow some procedures which dictates what tests you run. Imagine having an AIS system that tells you what is the next set of conditions that you have to test to bring you to your desired optimization result that you want. For example, if you want to minimize friction with a particular system, what should the velocity and the load be to get there? And AI can help you get there faster instead of making a design of experiments, type of matrix of tests, maybe AI can help you actually do it with a lot fewer tests, or instead of doing controlled, randomized experiments, you can have aI prescribed to you what experiments to run. So just this can save a ton of time and a ton of resources for everybody that is running tribological experiments. That's an example from me, yeah, that's

Speaker 5  17:21  
definitely a good point Nick mentioned, and I think there's even some more to that. Nick already mentioned the experiments and planning experiments and getting to optimization or getting to the right experience in a faster time. And I think there's even more to that than speeding it up or optimizing the procedures. When talking about applying AI to experimental data, we have to think about having our data in a certain structured way so it is basically AI, readable and usable, and contains all the required information to extract the knowledge from it. And also this process, maybe of setting up these databases and thinking really well of how to structure your data, that alone is a very, very valuable aspect of considering what are the influencing factors. So it's not just letting aI think, but like the, let's say, AI giant waiting for us, and we have to somehow pathway also makes us maybe think about certain aspects. How do we have to structure the experiments? What is the relevant information, and so on? And this makes us maybe also train the human brain and structure our knowledge in a better way, not just for the AI, but for ourselves as well. And in addition, if AI lets us encounter new patterns in our data that we haven't seen before, maybe also identifying research gaps, or maybe also contrasting information to experiments with a similar set of parameters, but saying the complete opposite. So maybe we can also identify errors in our experiment or in our existing data knowledge. And on the other hand, of course, we have the acceleration of also of numerical methods, of simulation tools, which are partly time consuming nowadays, but AI might help us to speed it up. And in the end, I think we need two things, or we need some added value from Ai, which could be higher speed, so lower research cycles, so to say, and also identifying patterns or maybe maybe design innovations that haven't seen beforehand. Yeah,

Speaker 3  19:31  
let me get back to one on point parameter that you've mentioned, which is data. And it goes without saying that data is basically the backbone of any trained model of anything and anything that has to do with AI in principle. Now, how would you qualify a data that's coming from experimental work or something that comes from simulation in the first place? How would you even deal with data that's probably not proper? Least structured, per se, because we can see the multiple practices within the community. So not everyone is really aware of the importance of structuring data, especially with experimentalists, where we have, like, tons of parameters to take into account before conducting and while and after conducting the experiment, I'll hand it over to Nick Because personally, I know that he has done a lot of work concerning this point, in particular, the

Speaker 4  20:28  
way I explain this and when we usually talk about data, it's actually hard to understand what the problem with data in tribology is really and I usually have to say that with the advent of computers, we got a lot of benefits. You know, we were able to bring a lot of our knowledge in a structured form that is saved in a very resilient and easy to share manner. However, that way of saving information, usually for the most part, counts on natural text or spreadsheets, and the transformation from paper and paper based note taking to natural text and spreadsheets, so from paper to Word and Excel actually obscures quite a bit of the human Creativity. In doing that, we have lost some of our ability to make quick connections between our thoughts that we would jot down easily on paper before. So it turns out that we still don't have a protocol for efficiently and completely storing all the knowledge that we have as humans. So actually, what happens in companies, and I've seen it in academia, the carrier of knowledge turns out to be the human brains, and that I believe we could do better at, we can start to collect our data. Max already alluded to the fair acronym by saying findable, Accessible, Interoperable and reusable data. So that goes by fair. If we start to really describe what we do in tribology in that manner, I believe we can start to discover exactly the connections that Max was talking about right. Start to see new trends and new applications that we haven't even fathomed before. So yeah, data, we have to care about it. We have to really think about how we record it, how we store it, and make sure that data and knowledge is recorded in a way that makes it talk to other data and knowledge autonomously. And that might sound a little bit like very abstract and science fictiony, but I really believe that we have to go back to the roots of how the Internet was started with these independent agents that talk to each other and produce new knowledge. So we should do something like this for Tribology. Now, before

Speaker 3  22:48  
I hand it over to Max, I would like to ask a question, do you see that there's standards that are made from institutes, individuals or institutions, whatever, to just deal with their data in a better way, or it's still non standard, and everyone is trying to do it as properly as they can, or probably not giving it much importance. Yeah,

Speaker 4  23:11  
standards definitely have a place in this discussion. I'll actually let Max explain more about standards, but what I will say is that I try to stick to the word harmonization, not standardization. It's a bit different in the sense that standards often sound like something that you have to stick to by the book, and you don't have any flexibility. If that's true or No, that's a discussion. But in any case, standards seem to scare a lot of people that are in the AI ecosystem. Instead, when we talk about harmonization, it's agreeing on a format of or on a structure of how you express knowledge. And when you harmonize, you actually enable the connection between knowledge and data over space and over time, from the past to the future and from one place to another. So we have to just agree. I believe that we should do better. Now, specific examples, I have not seen examples of well recorded data to the level that I desire. I'm not going to mention the people in this call as good examples, but I think there is quite a bit of work in that direction to be done. Also, it's not just what I say. It has to be a discussion in the community how this should be done. It's not what one person or one group believes should be done. It's what everybody agrees on and Max maybe pick up on the standards.

Speaker 5  24:35  
Yeah, it's a good point, definitely. I mean, I also prefer the term of formalization rather than than standards. But I guess it applies to how to store the data and process the data and so on. But it already starts a little bit earlier in the process, when conducting experiments and using an equipment that can be either from a commercial provider or also from a self developed, test driven and even when we think about like one of the most simple and most. Use tribal meters, which would be ball or pin on disk, linear, reciprocating. There's different methods of how you you avoid your coefficient of friction during that so we have this friction cycle every time the ball or the pin goes back and forth. And then the question will be, Are we avoiding some average frictional value, the maximum value, the point at highest velocity during the cycle, and so on. And this will lead us to different results. And then it depends on not only what we want, but also what the technical equipment allows us to export. So what sort of data pre processing is already done by the sensor or by the controller in the tribal meter, what we then get out what is the measurement frequency and so on, and that will influence also our results. So when comparing and storing different data, we need a lot of information from that testing devices as well. That has to be included in order to make a fair comparison and also to derive conclusive knowledge from that without comparing apples and pears in the end. So I think that is one very important aspect that we also have to think and to discuss about. And of course, when storing the data, ideally, we can have it in the most raw format possible available. But still, as mentioned, there is some sort of pre processing on the controller, and we have to be aware of that impact, because that will also influence of how potential AI agents will will learn and extract information and potentially make decisions. So that's one of the issues. And As Nick said, probably is not the one right way of doing it. It is a discussion, and of course, when talking about standards, it's, on the one hand, nice and especially relevant for industry to have certain standards to comply, but on the other hand, we're talking about research, and different types of research topics, different types of material, lubricants and so on, might also require differently, nuanced evaluation procedures so they have the justification. Each of the different protocols might have the justification, and it's probably not about deciding on the one right aspect, but it's just being on transparent and having that information accessible also for further evaluation. So that's one of the topics that's, of course, associated, especially when we're talking about experimental data. So I think there is nowadays still a very large or strong bottleneck in that sense that we still don't have the common understanding of how to process, how to provide, how to store and publish, especially also the data. So that's the bottleneck we have on the experimental side. On the other hand, we have the aspect that I also like quite a lot, is when, which we haven't talked that much yet, but is simulation data. So basically, it's not just about experiments and feeding AI with that. It's also about simulation data that, by nature, somewhat comes a little cleaner, probably at least it's a more clear result that the same input to a simulation will give us the same output every time, which already kind of reduces the deviations we have in the data, which is another challenge in the experimental one. And of course, we can use simulation data which is basically built upon physical knowledge already existent, and also train machine learning or AI tools to just speed up and make it accessible for a broader range of people, and also maybe really bring it in larger level or higher level system simulation, for example. And I think one aspect that I didn't get to yet is the deviations on the data, which could also be be interesting aspect to study and also to process. So in order not to just come to absolute conclusions, but also really considerate, okay, what is the expected significance? What are potential deviations of predictions or optimizations? So what is the uncertainty in what AI learns when making the predictions, as we humans do as well, we might learn from data on again, our knowledge, and then when we try to extrapolate, we have a gut feeling, and I think we have to enable AI with that as well, which comes by including all the relevant information from the test trick to the test conditions and also the raw signals of the measurement data.

Speaker 3  29:10  
Those are very, very fair points that you mentioned here. Now I'd like to touch down on more of a positive note regarding the topic. So one thing that I like about the tribology community that we're relatively small compared to other sciences, and we're kind of cohesive, and you can see a lot of advancements in the right path. Now this question is more since the podcast is about AI and machine learning potentials. So do you see good practices within the community, and how would you qualify the involvement of organizations such as the st le GFT here in Germany, and what has already been done in the right path, in your opinion?

Speaker 5  29:55  
So I think the answer is yes, there is certainly a good. Practice, or at least a trend towards the right direction. What we can see popping up in the last couple of years, more and more, are some sort of technical tracks or sessions dedicated to machine learning and artificial intelligence at various conferences all around the globe, also mentioning the the SD le manual meeting, for example, since or in this already, for the third year, there is a session dedicated to AI, which is a good signal, and also the restaurants, so to say, from the response from the community in attending the sessions and submitting talks. And we can see a lot more talks on AI all around the globe, more papers being published and so on. And as also mentioned in the introduction, as Felina offers this teaching course, or one course that has been half a day at the last annual meeting, which will probably be extended a little bit, and you can see that there is a good response, both from industry and academia. So there's a lot of interest. And I think people are losing the fear or the distance of AI and then actually employing it and exploring it. And I'm sure not all what has been done and what is being done is leading in the right direction, but there is a lot we have to explore what is possible and and what does also make sense. And so I think that's a good thing, that there is now the awareness of the topic, and more and more things are done. On the other hand, I think that we should consider the way of how we are doing it also the conferences, because AI, I think, is not a topic of its own, and I think we should rather consider it as a method that we can employ to the different areas. And one thing that I thought about the other day was that actually we have this isolated artificial intelligence or databases sessions or machine learning sessions at the different symposia, colloquiums, conferences and so on. And then we have a lot of people from different areas that do AI and can exchange, which is perfectly fine, but at a certain point, I think we have to transmit it in the normal technical tracks of the conferences, for example, the Australian meeting, because this is where we want to employ it. And we're not doing it because we want to do AI. We want to do it for a purpose. And I think that aspect should not be forgotten.

For more information, visit www.STLE.org.