In this episode, we discuss a GPT style AI model being developed by a multidisciplinary team led by the University of Michigan to tackle the battery development bottleneck preventing wide scale adoption of electric vehicles.
This podcast is sponsored by SAE International.
This episode was brought to you by SAE International! They're THE go-to spot for automotive lovers and engineers—keeping you ahead of the game and hyped about the future of mobility!
Click HERE to learn more about the technical resource that we shared this episode from the SAE team regarding battery chemistry requirements for construction EVs!
When can we finally get the $20,000 EV? Well, folks, that's what the people at the University of Michigan want to answer because they are building a GPT-style AI that will tell us exactly how to build our batteries to make them more affordable. If that's got you hype, then buckle up for your future affordable EV.
I'm Daniel, and I'm Farbod. And this is the NextByte Podcast. Every week, we explore interesting and impactful tech and engineering content from Wevolver.com and deliver it to you in bite sized episodes that are easy to understand, regardless of your background.
Farbod: All right, folks, as you heard, we're talking about batteries. It's been a very hot topic for us on this podcast. Energy storage, EVs, all the good stuff. But before we get into today's episode, I want to quickly talk about a technical resource from today's new sponsor, the Society of Automotive Engineers. Now, this is a special sponsor for Daniel and I. Fun fact, we actually co-founded the Formula SAE team at George Mason University together. So, it's got a special place in our hearts. What you gotta know about SAE is that, as the name indicates, they are where you need to be if you wanna be an automotive engineer. If you're interested in automotive engineering, they are the spot to be, because they know some really cool stuff about what's going on in the automotive industry. Now, this resource that we wanna talk about, it's specifically about how battery chemistries are gonna be a strong indicator of how construction EVs, like the cars being used for construction purposes, are gonna start becoming deployed at a wide scale, right? They talk about the European market and the Chinese market and how both of these markets are using different chemistries and how the chemistries they use are going to lead to the actual cost of these vehicles and how the cost of the vehicles is going to really drive who's going to buy them and actually use it for construction. I thought it was pretty interesting. We've talked about EVs a lot on this podcast and the fact that they get deep into the chemistry of like nickel manganese versus the lithium ion phosphate and the potential of like the sodium ion and the drawbacks and what makes sense and what doesn't. That was really cool to me.
Daniel: Well, yeah, that's what I was going to say is you said if you want to be in automotive engineering you need to be plugged in with SAE. I'm going to say I think this exemplifies even more so that if you want to be involved in the future of automotives, SAE is definitely the place to be, right? They've got this really in-depth technical resource, not just about EVs as a trend or even EVs as a trend in construction. They're going all the way to the extent of understanding the different battery chemistries and different price points that will allow for high adoption levels of EVs in the construction industry. That's the level of detail they have. This is the level of knowledge they have on a ton of different topics.
Daniel: We were just happy to select this one because it's really interesting to us, and it collaborates really well with the topic we want to talk about today. But I think, again, case in point, it shows just how plugged in SAE is. We're going to plug this article on the show notes. You should definitely check it out and see all the other cool stuff that SAE is working on as well.
Farbod: Definitely. And I'll extend one more thing. What we try to do on this podcast is like deliver information that we think is really important in a digestible manner. Right. This article took me maybe less than five minutes to read. It had some technical notes in there about like considering power density versus energy density and why these chemistries matter. I got all of that and I haven't touched the chemistry book in over four years.
Daniel: Yeah, I agree. For as much in the technical weeds as this topic is, I say it is a really, really good job of boiling it down into a way that's really easy for someone like you and I who aren't super plugged into battery chemistries to understand. And it was an awesome primer to talk about the topic that we're going to be talking about today, which is building a chemical GPT of some sorts, to help design key battery components, which, you know, again, same ballpark here. And you can understand the context of if you're able to achieve the correct chemistry, achieve the correct power density, achieve the correct costs, then a ton of EV adoption will follow. That's what SAE tells us. And now we can jump into the technical weeds of, is it University of Michigan?
Farbod: Yep. It's University of Michigan.
Daniel: The University of Michigan is building an equivalent to chat GPT, like a large language model. But they're doing it for chemistry, trying to create a GPT chemist to help design key battery components.
Farbod: And when, I don't know, it's like these two hot topics right now in the world of technology, right? You got AI on one side. And then on the hardware front, you got electric vehicles. When I saw this, I was like, this is a pretty genius approach. One of the researchers was like, chat GPT kind of showed us what is, what has been missing from a solution like this, like an AI driven chemist. And that's just, you need a lot of data. And they, the example they use was if you have a Shakespeare bot and you only give it a handful of books or whatever, it's just going to spit out mediocre Shakespeare esque content. But if you give it all of Shakespeare's works, then it starts to really spit out the things that make sense for you. Like what sounds like new creative Shakespeare style content. So that's the approach that they've taken here, right? They're like, we want something that can tell us how we should be approaching chemical design. So, we really want this thing to be a chemist from the ground up. Now how have they gone about that? Or actually, you know what, before we get into that, why are they doing this to begin with?
Daniel: Well, that's what I was going to say. Some context here, existing methods for designing molecules specifically for battery components. The battery components that they're most interested in here are electrode and electrolyte pairings. It's the same conversation we were talking about at the beginning of the episode around like nickel manganese, cobalt versus lithium iron phosphate. Those are battery chemistries that involve the electrolyte and the electrode combinations, which are how energy is stored and then released from a battery. That's the core of the chemistry at the heart of it. That's how and why we get to certain power densities. That's how and why we get to certain costs, certain energy storage efficiencies, et cetera. So, when we're talking about battery electrolytes and electrode pairings, the existing methods for designing new molecules, new potential candidates for us to use as the electrolyte, they're limited, they're not efficient enough. It involves a lot of human intervention in terms of one, designing and coming up with new ideas and then on the tail end also a lot of human intervention and testing a bunch of those ideas to see which ones are feasible to see which ones might make it out into the real world. All that to say if we're going to find better materials for battery electrolytes to massively improve energy storage capabilities to get to the power densities that we know we need from a theoretical perspective to achieve certain levels of adoption in industries like construction or for widespread adoption of EVs say, you know, a really pertinent example for people that hits home for a lot of people is how can we design an affordable electric vehicle that doesn't weigh too much, that is relatively efficient and gets five to 600 miles of range, right? That's something that really hits home for a lot of people. I think if you could tell people that they could reliably get 600 miles of range out of their EV every single time they use it, regardless of whether they had the car loaded up or not and it would charge relatively efficiently, right? All these are really big pain points for EV users today, that comes down to the battery chemistry.
Farbod: For sure.
Daniel: And there aren't really efficient ways of coming up with new battery chemistry. So that's where they try to involve AI here is to help one, help us get to new potential battery chemistry, specifically the electrolyte electrode combinations that could hopefully potentially unlock tons and tons of applications in the future.
Farbod: Yeah. And like, what I didn't know is that when it comes to the battery chemistry research side of things, we have a lot of electro combinations which are able to perform far better than the current state of the art. And what has been the bottleneck has been that electrolyte, the medium by which the ionization is happening. So, it's like we have, it's like within the reach, we can almost grasp it, we can almost feel it, but this thing has been the bottleneck. And that's where we can get a lot of benefit from an AI that can just kind of go through every permutation, every combination and give us the feedback of is this gonna work or not? And I'm gonna refer back to the SAE article real quick. They were talking about how this report is very fixated on cost, because cost matters for commercialization and all that. With cost, a big driver of it is demand, supply and demand being the foundation of economics. Because we're limited to the battery chemistries, the typical lithium ion that most manufacturers for consumer level electronic vehicles are using, everyone's going after the exact same rare metals, the precious materials. If we can really have a way of exploring different electrolytes, then you can have different battery chemistries for different EV purposes, which is gonna bring the cost down and make everyone not go after the same rare materials and therefore kind of open it up a little bit more.
Daniel: Yeah, I agree, right? Even if you're looking in something that's in the similar ballpark of efficiency, similar ballpark of power density, et cetera, but you've provided an alternative in terms of which minerals are being used. That unlocks a whole additional aspect of supply, which allows us to potentially harvest these more sustainably, do it in a way that's maybe more ethical, and then maybe do it in a way that, maybe these minerals are more recyclable at the end of the day.
Farbod: And life, absolutely.
Daniel: All that being said, even if they were exactly as rare, exactly as challenging to mine, and exactly as unrecyclable as we've seen with nickel and cobalt, even still, if you were able to provide an alternative that duplicates, let's say doubles the amount of supply, that would reduce the cost to the consumer, reduce the cost to the manufacturer and then to the consumer considerably.
Farbod: Yeah. No, I agree with you. So, like the “Why” of this, I think is pretty straightforward. If they're able to crack this solution, if we're able to have some sort of AI that we can go to and say, hey, we have this electrode pairing that we're interested in come up with the right electrolyte, that could revolutionize the entire battery game. So how did they do this? How did they go about solving this problem? You brought up chat GPT. I mentioned that they were really inspired by the way it worked and by the way it was trained. The lead researcher, I believe, mentioned that with chat GPT, they shoved in a lot of text without any of the annotation. They're not giving it any sort of what is the meaning behind this. By brute force, they wanted to understand the relationship between the words itself. So, they took that pretty much same approach where they gave this model that they built the atomic structures of the molecules of interest, I think mostly hydrogen, carbon, oxygen, and nitrogen, and all these different combinations of the molecules and they didn't tell it any of the chemical properties. So, they didn't say, hey, with this molecule, you know, this one's flammable, this one's corrosive or any of that. They just gave it all that information and the training process went something like, well, now that you have this info, can you predict if an atom is missing for this molecule or not? If so, let's move on to the next step where now we give you a little bit of information. We say, hey, molecule A, this compound has this property based on that.
Daniel: These physical properties, these chemical properties start to paint in the picture.
Farbod: And then can you start. Hello, we have a little guest on the podcast, for those of you that are audio only, it's my cat Kishmish asking for some pets. She's also really into battery chemistry. Yeah, she loves battery chemistry. It's really resonating with her.
Daniel: Can't complain about that.
Farbod: Definitely not, but coming back, if you have the information about compound A, can this model now infer or predict what the properties of compound B would be like. That's the real meaty, juicy sauce behind this model that they've developed.
Daniel: Yeah, and I will say the challenge traditionally with someone trying to achieve this level of computation, let's say, would hinge on them having tons and tons of computing power to be able to, like this almost sounds like as brute force as AI can get. Right? You feed it a large data set, thousands and thousands of these small organic molecules relevant to energy storage and you force it to start to think through every single potential combination to understand, like you said, which atom is missing from this molecule? This sounds super computationally intensive. In my mind, I was like, how in the world are they achieving this? Well, the answer here is they're using a 34-petaflop supercomputer called Polaris that I guess they have access to. 34 petaflop means it can do 34 quadrillion calculations every single second. This is as super as super computers can get, and they're utilizing 200,000 node hours, right? Of training this AI model on this 34-petaflop, super Polaris supercomputer. As computationally intensive as this model and training the model might've been required, they had the supercomputer firepower to get it done.
Farbod: And now it's a good time to talk about like the partnership they have. It's actually the Department of Energy that is allowing them to do this research. It's giving them 200,000 node hours of access to the supercomputer to train this model because they are invested in us having better access to a tool that can streamline this entire battery chemistry problem, right? But in addition to that, they're also working with Deep Forest Sciences, which is a company that's been doing medical research using a very similar approach, but instead of going for the best battery chemistry, they're going after drugs. They've been trying to do drug development in this manner and they've had a lot of success.
Daniel: Well, and I think we've mentioned on the podcast a couple of times different applications of AI for drug discovery or for drug testing. This is, I think, one of the first times we've mentioned it in terms of battery chemistry. So, it's awesome to me to see them cross pollinate these discoveries that have been done, not with the same exact technology or maybe not with the same exact parameters, but an analogous version of this in the medical space and to watch that trickle into material science. And as we know in the hardware realm, I think it's, it's usually a material science development that leads to a huge cascade of hardware that follow.
Farbod: We love food. We've talked about it before, but it feels like the material scientists, they're like the farmers of the world. They come up with those ingredients and then it's everybody else that's taking those ingredients and making the dishes that we all love without those ingredients. Yeah, you can't cook anything interesting. You can't make a good secret sauce. You can't make the sauce.
Daniel: But you know, we talk a lot on this podcast as well about first principles thinking. The first principles in hardware always, it usually comes down to physics. And the part of physics that's limiting any certain application from happening, it's usually the material science. And so, it was really interesting to see them using a lot of computational power to try and help them discover new materials, discover new potential electrolytes that unlock more feasible electrodes. Like you were saying, these electrode pairings are already known and we're looking for electrolytes that can help make those feasible. Yeah. To me, this helps that you, we unlock a lot of potential in the future. We also learn a lot about which combinations don't work. That's just as important to make sure that we're not spending testing resources on potential combinations that won't make it out in the real world as well.
Farbod: And what a perfect segue into the so what. I love this one quote from the researcher. They said that when we learn chemistry, we learn each rule and then we learn about a dozen exceptions. So, one of the things they mention is that by giving all this foundational knowledge to this model and having them truly become a chemist, we're hoping it can tell us what are some new design rules, some design rules that we've never even considered, what those could be for us to follow down the road. So, we're hoping for it to teach us, right? And it's very well aligned with what you just said, which is not only can it tell us which ones to explore to get the best value out of, but which paths do we want to completely ignore because it's just a dead end.
Daniel: Well, and I will say, right, we talked about the AI bit, the machine learning bit, which is adapted from these large language models that were originally meant for trying to understand and predict language and now they're being used to understand and predict complex chemical structures. They also mentioned the potential to use robotics to help automate the testing on the back end there to get some empirical data, get some testing data from the real world to correlate and back, reinforce back into this model and help make it stronger, help make it more accurate. It's really interesting to me to see, like we're gonna use AI for generating a bunch of potential feasible combinations. We can potentially use AI robotics to help assist with the physical testing part, conducting experiments to see if it works. That will help reinforce it. That will help generate, like you're saying, some more foundational principles of laws of chemistry could be defined and discovered, defined, tested, and understood completely autonomously, which is really, really exciting to me because again, this is something that historically for the last couple of thousands of years has always involved humans being involved to postulate, to make the thesis, humans being involved to test it, humans being involved to recognize what that pattern is and then publish it to the rest of the world. And then more humans try and replicate that experiment to make sure that it's real. The first five of those six parts, they're talking about automating, using computational power, using next-gen robotics, et cetera, automating the first five out of those six steps. It allows humans to still be involved in the really important parts, right? Setting the targets for what this machine's gonna do, understanding what those important laws are and communicating them to the rest of the world. That's the stuff that scientists love to do, the stuff that they have research assistants doing, cleaning test tubes, doing a lot of reading, figuring out which pairings they should try and test in the lab and then going and testing them and recording the results and reporting back. That's where the AI is gonna come in and help accelerate a lot of the boring part of chemistry. And I imagine a lot of these chemists get to do what they love most, which is postulating and then understanding the results without having to get out in the weeds and discover 99 ways not to do the thing that they want to do along the way.
Farbod: That's where they can have the most value add to, right? That's their expertise. And by the way, in case you weren't impressed enough by this team that we're talking about today from University of Michigan, I'll note that this robotics side of the solution was also developed or co-developed by the lead researcher on this project. So, this is really an all-star team that…
Daniel: Talk about range, man.
Farbod: I know robotics and…
Daniel: AI and chemistry and EVs, right?
Farbod: We're all over the place here.
Daniel: It's like you stuck your hand into a grab bag of technology buzzwords and pulled them out, except you're not just trying to be buzzy.
Farbod: It's not just buzzwords.
Daniel: This is a team that's experts in this field, and kudos to their partnership also with the DOE, giving them the computational power, they need to achieve it. I imagine there are a lot of people with similar skill sets who would love to do the same thing and weren't able to leverage that literally, supercomputer to help achieve their goal.
Farbod: Yeah, I'm with you. What do you think? Let's do a quick wrap up.
Daniel: Yeah, wrap it up.
Farbod: All right. Folks, have you been wondering where can I finally get my hands on a $20,000 EV? Well, these folks at the University of Michigan are trying to help you out. You see, battery chemistry really impacts the cost and we've kind of cracked the code, let's say 50% of the way, because the electrodes that make up your battery, we have better versions of them. But the thing that the ions go in between the electrolytes, that's been the bottleneck. Well, these guys are trying to use a GPT style AI to tell us what kind of electrolyte to use. The setting that they are proposing is we tell the AI, we wanna use these two cool electrodes, it tells us what electrolyte to use, and then all of that goes to a robotic solution that will experiment in a lab and tell us if it actually worked or not. These folks are being funded by the Department of Energy and their work could definitely revolutionized the entire EV space.
Daniel: I dig it, man.
Farbod: I try, I do it for the fans. What can I say?
Daniel: Yeah, well, let's say before we wrap up, we have to thank our community in Spain. We trended in the top 200 podcasts in Spain, correct?
Farbod: 160 something.
Daniel: Let's go.
Farbod: It's amazing, man.
Daniel: So, we'll have to hold up our end of the deal and eat some Spanish food.
Farbod: Which is great, cause I love jamon.
Daniel: Yeah, he does.
Farbod: I do love jamon.
Daniel: For people that don't know, for both went how many years without eating ham?
Farbod: 20 something, 24?
Daniel: The first ham you ever had was Iberian ham.
Farbod: It was amazing. It was delicious. And I will still stand. Jamon is here and bacon is somewhere down here.
Daniel: The American blood coursing through my veins is boiling a little bit, but I'm not sure if it's that or if it's all the cholesterol and fat from the bacon.
Farbod: It's the cholesterol and fat. But yeah, thank you guys so much for listening. And as always, I guess we'll catch you in the next one.
As always, you can find these and other interesting & impactful engineering articles on Wevolver.com.
To learn more about this show, please visit our shows page. By following the page, you will get automatic updates by email when a new show is published. Be sure to give us a follow and review on Apple podcasts, Spotify, and most of your favorite podcast platforms!
The Next Byte: We're two engineers on a mission to simplify complex science & technology, making it easy to understand. In each episode of our show, we dive into world-changing tech (such as AI, robotics, 3D printing, IoT, & much more), all while keeping it entertaining & engaging along the way.