Episode Transcript
[00:00:00] Speaker A: Welcome to the future of energy, where we'll meet the people who are working with the products of tomorrow. In each episode, we'll take you a step further into the world of electrification. From the roads we travel on to the vessels that navigate our oceans and even the skies above, we'll unravel the threads of innovation, weaving together the clean energy future. So fasten your seatbelts as we navigate the future of energy, one conversation at a time.
[00:00:24] Speaker B: You.
Good afternoon. I'm your host, Annali, and in today's episode of the Future of Energy, I am joined by Siddharth Khilar. Welcome to the studio, Sid.
[00:00:39] Speaker C: Thanks for having me.
[00:00:40] Speaker B: So, today you will be talking to us about AI, machine learning, but mostly about machine learning. And I think the best place to start is what is AI and machine learning?
[00:00:50] Speaker C: So, in simplest terms, it's essentially a machine's ability to attempt and complete complex tasks at a level of performance that's at least as good as a human, or sometimes even better. That's the academic definition, and colloquial definition depends on who you ask. They'll give you different answers.
[00:01:13] Speaker B: Because 2023 was an interesting year for AI, machine learning got a lot of attention, and a lot of people started questioning what's going to happen to the world and us as humans. Do you have any comments on that? To settle people's nerves?
[00:01:25] Speaker C: I think we'll be fine. I think we've always had the early paranoia set in. In fact, 2023 was the year when became mainstream. The inflection point. There were a few inflection points. One was in 2012, when the first big machine learning model beat the benchmark of all computer vision algorithms ever. So the first deep learning model actually was trained on a GPU and so on. That was, I think, 2012 called Alexnet. And then the next one came in 2017, which is the famous papers around called self attention is all you need, was the title of the paper, because it was based on this new neural network architecture called the transformers, or self attention, which has now turned into what you call llms, what we come to see as large language models or large multimodal models.
[00:02:20] Speaker B: But it's like the typical reaction, right? Like, when computers and the Internet became a sensation for everyone, and it was open to everyone, everyone started freaking out. But it already been around for a while, especially, like, governments have been using it.
[00:02:34] Speaker C: Yeah, I think we are going to be fine. I'm actually very hopeful, not just because I need a job, but also because it's exciting to be part of something that's transformational and be able to use it, apply it, shape in some way or form. In our industry, at our company in general, we as humans, there is enough ingenuity to figure out how to put this, any technology that comes our way, or we develop ourselves to put this to good use. So overall, it will be a net positive. Obviously, there will be bad actors. We will deal with bad actors as a collective, benevolent civilization. Now, I'm just using big words, but.
[00:03:24] Speaker B: Getting deep and philosophical.
[00:03:25] Speaker C: Yeah, it's my optimism. I'm an optimistic person in general when it comes to technology, and I think there are enough powerful companies and powerful people in the world that will make the right decision, even though sometimes they will be largely economics driven, but I believe they will be positive for the people.
[00:03:48] Speaker B: Well, it's a tool for of. Is that how you would see why it's necessary to.
[00:03:55] Speaker C: So, you know, one of my very close friends, Nanda Northwold, here in Stockholm, though he says this sentence, which is quite, it means a lot, and it has a deep meaning, is AI will not replace people.
It will likely be the augmentation. So it will replace the people who do not use AI, and it will replace them with people who use AI. That's a very key differentiation, is AI will not replace the people. It will likely be the people who do not use AI with the people who use AI. And in this case, it could be the same people who just start adopting it. So I think it's an augmentation, not an alternative.
[00:04:38] Speaker B: So how would you say AI and machine learning are transforming the battery manufacturing.
[00:04:43] Speaker C: Process we see as a company? And there are many teams using and playing in the sandbox of machine learning. It's not just my team.
At least I can count four different teams that are using data and machine learning and algorithms in some way. I think it's important to make a distinction around some of the hard science problems versus some of the day to day data number crunching problems, where you have immense amount of data that you'll be unable to process using traditional classical machine learning. So you need deep learning, which is a whole different field of how you build models and draw inferences. In terms of science, I think we are still bouncing between the idea of how much physics do you need to encode in these AI models that we develop, versus how much of it can you just be.
In typical terms, it's called end to end learning, where you just show it the raw sensor or the raw data, where a domain expert will likely convert the raw data into some features, like it's essentially translation of one language to another. In this case, it's raw sensor data to something that is more understandable for the domain experts like discharge curves and so on. There are many different features. So now we are going towards how much physics do we need to encode? And we are testing that is, how much physics do we need to know? What can we actually just assume that the model can learn? So that's a very exciting space that brings two things forward. One is we can start working with more and more unstructured data. So, so far, one of the biggest challenges is to bring unstructured data into structured form. And the other exciting area is around material science. So for our industry, material science is a huge pillar. And one thing that we can take a chapter from the drug discovery playbook is how they have applied AI to discover new drugs and utilize genetic sequencing and other imaging techniques to inform how to formulate drugs. So we've been thinking about in that space and playing with that. There is the data team in digitalization under Marcus Ulmaforsch, and they've done some incredible work on bringing a ton of data from the sensors into a large data lake, which enables a whole bunch of analytics and dashboards and business intelligence in the factories. But also now we've been working together to how to pull that data out and start making certain decisions around population of cells, around asking questions to unstructured data using large language models, thinking about how we bring together the knowledge around process failure modes, around control plans, around quality data, and combine it with the factory telemetry data and make better decisions. So we've been playing around in that space. We don't really have anything concrete to share. I think the team has made some very exciting progress, but it's also been very much like, oh, this would be cool to do. Oh, that would be cool to do. What if I could make this graph red or orange? And how would this work? So it's a lot of exuberance in the beginning, but I think now we are settling down and being humbled a little bit and really focusing on what should we ship to the factory and to the customer.
Customer being eth, in this case, immediately, so that they can do much better.
[00:08:15] Speaker B: Yeah. Because just to break this down, the complexity of battery manufacturing in the first place is enormous.
[00:08:23] Speaker C: Oh, yeah.
[00:08:24] Speaker B: And it's still so new in terms of just this type of manufacturing and especially at the scale. So it is a tool to use something that can help out with, like you said, breaking down or even simplifying data that people can then understand. But AI, of course, and machine learning, you're able to feed something that can keep learning. Right? Yeah.
[00:08:47] Speaker C: So you have two points there. One is we don't know what we don't know. Right. And how much do we want to branch into the unknown unknowns of the problems that we think we could solve in battery manufacturing. Right. And then there are things that we know we don't know. Yeah, we've picked the second one. Things we know we don't know. Let's go actually challenge those assumptions and ask the questions of what can we learn from what we already know, how production process works and how can we inform certain decisions, certain process optimizations, predictions, anomaly detection, and so on. Then I think we are very hopeful that that will show us the path in a structured manner. Once we have the substrate and the ground built. Right. Then we can start branching into the unknown unknowns of what two intuitive areas in battery manufacturing could you connect and understand? Cause and effect analysis. Right. So I think those are the two big buckets of projects and things we think about. And the second one can pretty much span all the way from Cam to revolt. Right now, we are super focused on cell is material comes in slurry all the way to end of line and product development in labs, which is also very similar. And there are some fixed rules that.
[00:10:12] Speaker B: You can play with. Exactly.
[00:10:14] Speaker C: But the unknown unknowns is what we are very excited about. I can understand that because that opportunities are.
That's endless opportunities, right.
[00:10:23] Speaker B: Yeah. So, I mean, that kind of goes. So materials. So cathode active material is what we're talking about with CaM and revolt is the recycling material, how would you say machine learning and AI impacts or could impact material selection, cell design, and of course the manufacturing.
[00:10:43] Speaker C: So I'm not a material scientist. Right. So whatever I say may come off being insensitive to a lot of the hard work that goes into picking a material. Right.
But again, this is a lot about that domain expertise.
There is a common denominator of that domain expertise around material science. I think we'll very soon be able to have copilots that have learned that common denominator of what it means to understand certain amount of material science. Then there is the human element on top, which is the experts in Cam R D and Cam teams and experts in rebold in our own proprietary processes. I see. That's one opportunity where you can offload a lot of the heavy lifting that is time consuming and repeatable and rather not necessarily moving the needle so much, but is very essential to be known and enabling our people with those types of tools where we can really cut the latency of idea to a mind map of what do I need to do next and how could be very useful. I think the second thing is in manufacturing is both processes are extremely expensive. Right. So Cam is very expensive to produce.
And this is, again, something that we hope to start doing soon. We've not really ventured into those two areas yet as a team, except digitalization has been. That team has been supporting a lot of the manufacturing bring up for upstream. They've been having conversations there, but in the sense of us being able to think about how AI descends upon the problems and how it helps, I honestly feel, in improving our eels in material quality and the big bag production out of upstream and what goes into revolt and how to think about that. We'll find opportunities there, but I think we have to get the basics right first. Right.
[00:12:50] Speaker B: So future wise, would these sort of systems be able to identify in a much quicker manner if something is in a specific area manufacturing that's not going right or correctly?
[00:13:05] Speaker C: Yeah. I mean, that is a paradigm that already exists, I think, in different material manufacturing industries. So anomaly detection based on some things is a thing.
Yes. That's definitely more like applied machine learning. It's about putting a sensor in. And I also feel our deep understanding of our own cam can lead to even better manufacturing improvements. If we can start to think about the small scale particle level features, then.
[00:13:41] Speaker B: There'S less room for error.
[00:13:43] Speaker C: Yeah. Particle level features of what are the properties of the particles that actually lead to a problem when we are producing a ton of it? How would we change a process parameter to improve this property of a particle? Right. And there's a lot of imaging and there's a lot of offline sensing that goes into particle analysis and so on. So I think there is an opportunity there. We haven't really touched that. When I joined the company in the beginning, I was very excited about all the imaging applications, but then got humbled by our leadership and said, we got to fix this first.
[00:14:15] Speaker B: We started focusing one step at a time.
[00:14:17] Speaker C: Yeah, one step at a time. Bite as much as you can chew.
[00:14:20] Speaker B: Yeah. So we talked about before some more specific applications in the manufacturing process, and you were the one that mentioned to me that it might be interesting to talk about quality control, predictive maintenance and material selection. Yes, we talked a little about material selection, but would you like to just touch upon quality control and predictive maintenance?
[00:14:39] Speaker C: So I do want to say one thing about material selection. Go for it. One last thing is, it's about material screening. So in R and D, I'll give you an example. We have an electrolyte. The electrolyte is a very crucial component, and that's where a lot of the tuning happens across the industry. The electrolyte has a bunch of different components. It's not one molecule. It has a bunch of different molecules. Then those molecules form components, and then electrolyte is formed by mixture and proportions of those components. It's a multi tier process of molecules. We are very excited about thinking through how to screen the composing molecules that go into those components that form the electrolyte.
So you can tune the properties of what you want that molecule or that family of molecules to do. Like one example is additives. We can tune the additives by tuning the property. And that can be done using machine learning, that can be done using AI, where the properties are objectives and the molecules are the things that you tune. So there are some exciting, interesting work. We are not reinventing the wheel in any way, but we have a few secret recipes there that we are playing around with that's fun, that we think we should be able to have a significant, how do I say it? Like a jumping platform where we can jump forward whenever we want on some very specific problems. Because it's very expensive to test, to do trial and error, Monte Carlo, or like a matrix simulation of molecules and properties. And that can be done in ML. Okay, so that I'm very excited about, but coming back to your question, so.
[00:16:28] Speaker B: If you just want to break down what kind of applications you would use or how you would use AI, machine learning for quality control and maintenance as well.
[00:16:38] Speaker C: Yeah. So you start thinking about it from the perspective of, okay, what are the physical processes that are happening at the machine level or at the process level that today inform us to schedule maintenance?
And if you start thinking from that perspective, the first thing that comes to mind is we've done incredibly well. The MEs and the data team here in the digitalization team has done incredibly well, is to collect a lot of data. If you ask some people, they'll say that's a bad choice because we are collecting tons of telemetry data. That's useless. Yes, it is, but at the same time, that's a strategic choice that we made. But then when you start to look at that treasure trove of data, then you can ask questions like, oh, what are some key things that I should be studying to inform my maintenance schedules better? So that's one way to think about it, is machine telemetry data can be quite useful that you can then associate with performance of cells, which we also collect. Right?
[00:17:42] Speaker B: Yeah.
[00:17:42] Speaker C: When the cell comes out of the line, it has a few performance parameters that we measure before it goes into a wrapping section. So you have all the machine telemetry data. You exactly know which day did it come from and what process did it go through. And you can look at the maintenance schedules of that process and you can start to track how things are changing. This is no rocket science. A lot of industries do this. We are not inventing this. But that's one application we think is very interesting, is being able to control that whole data stack. This is one reason why we do several number of things in house. It's a key advantage, even though we may not be the masters at it, but the philosophy behind why we do it is in cases like this, when we start to think and try, it.
[00:18:31] Speaker B: Shows then you have all that data to play with.
[00:18:33] Speaker C: Yeah, exactly.
[00:18:35] Speaker B: Wow, that's super interesting. That's true. I mean, it becomes more predictable. If you know what is happening, you can go back.
[00:18:41] Speaker C: I'm very hopeful 2024 will be fun in some of those spaces where we start thinking and implementing something.
[00:18:49] Speaker B: Yeah, you get to play around more. So what would you say are the general challenges for the industry when it comes to implementing AI and machine learning?
[00:18:57] Speaker C: General challenges for the battery industry? I think one challenge is to choosing the correct integration strategy. And what I mean by that is all the way from designing the product, which is a messy process, to taking it to a pilot line like labs a b sample line, and then taking that to a Giga line and having traceability of all that data. Right from your coin cell pout cell experiments all the way to the giga manufacturing line? That's a huge challenge because it's sort of an implicit challenge, because it's a messy process to go through those phases. And oftentimes you want to move very fast and we drop things on the floor that don't necessarily get picked up, and the traceability of that data breaks and we lose a lot of the why we are doing something today and you have to trace it back in a manual manner. I think that's one big challenge.
[00:20:06] Speaker B: It's a lot of data and a.
[00:20:07] Speaker C: Lot of it's a lot of data which is sitting in different places and it's hard to join it and think about it in a serial fashion. I honestly feel that's a challenge.
[00:20:17] Speaker B: And do you think that that's the hopeful future, is that AI machine learning can help out with.
[00:20:21] Speaker C: Yeah, so I think with unstructured data techniques, I'm not saying, like, you could just upload data on Chat GPT and it'll solve your problems. That's not what it is, but what these new types of large language models or these transformer models that can look at very long contextual sequences of data to understand what's going on is something that is very new and that opens up a space that hasn't existed in enterprise software, which is being able to not care about how you organize your data, actually look at it in an unstructured manner. Now, that said, there are still nuances that sit in humans minds and memory on how decisions were made in coin to pouch to b, sample to giga. Right. That process is still pretty messy, and I think that's still challenging to think about. I think the other challenge is we, as a battery industry and academia, we don't necessarily understand all the physics. We understand it, we have it documented, but we don't necessarily use a ton of it in terms of accurate data driven model development. So what I mean by that is a lot of our computational understanding of battery models that we do simulations with is quite approximated. So the physics models that explain the battery physics, there is a lot of approximation that happens from taking those physics models on paper to computational simulations. I think there. There is a huge assumption bridge that becomes weaker and weaker when you start to stress it with trying to tune the battery to beat certain performance specs. So I think that's another challenge. And that field is moving very fast. I'm very excited about how some sparks of machine learning help improve that area. That we are very closely watching. And also investing in is how do we improve our physics based understanding of what's happening inside a battery at different phases. Right. From materials all the way to end of life. Yeah, I think that's super exciting. And that knowledge can be brought back to build new types of.
[00:22:40] Speaker B: So that can happen. Models of understanding simulation.
[00:22:43] Speaker C: Right?
[00:22:44] Speaker B: Wow. Okay.
[00:22:44] Speaker C: Right. We have an incredible simulations team.
[00:22:47] Speaker B: So, you know, in westeros, you know what's happening before you even.
[00:22:50] Speaker C: We can make better guesses today we make guesses, but they are sometimes constrained. We are constrained by computational complexity that is very hard to derive parameters from. I don't know if that made sense, but I think it did in my head.
So, today we have, the way I look at it is a lot of assumptions and approximations on those computational simulations so they can become realistic. Think about going from a 2d prince of Persia game in the like world of Warcraft or fortnite today.
[00:23:23] Speaker B: Okay. Yeah.
[00:23:24] Speaker C: Right.
The rendering and the accuracy of what you see on your screen has improved significantly. Right. I think that's because we have understand the physics of how we can interact with light in computers, make things look more. There are better models of how light behaves in the real world that you can implement as a program. I think that's happened because we have studied those parameters very closely. We have computational power that can do that.
[00:23:56] Speaker B: So it's the same kind of physics that you're referring to, that it's the.
[00:23:58] Speaker C: Same kind of physics, like, that's rendering and interaction of light with objects and so on. Right. And it's the same kind of physics is interaction of current and ions and materials and molecules and all the different chemistry and physics, thermodynamics, effects over time.
[00:24:16] Speaker B: That's incredible.
[00:24:17] Speaker C: Yeah. So that can be very interesting, I.
[00:24:19] Speaker B: Think that is so exciting.
[00:24:21] Speaker C: Yeah, it's like biology. I mean, there are all these different fields that you can think of, and ours is in that little living cuboid filled with electrolyte, anode and cathode.
[00:24:32] Speaker B: But is that what happens at some pharmaceutical companies? Is that what they have to do? I mean, they would have to do that if they're doing testing of certain.
[00:24:39] Speaker C: Yeah, so I think they have to study the interaction of proteins and drugs with proteins and how you do that. And that's come a long way as well. You can study it outside the body.
[00:24:51] Speaker B: First, I was going to say, instead of testing it on human.
[00:24:52] Speaker C: Right. And you can study it on the genetic code, and you can study the protein folding work that Google DeepMind has done with alpha fold, too. It's incredible.
[00:25:02] Speaker B: That's amazing.
[00:25:03] Speaker C: Is being able to unfold those proteins and understand how they fold out and how that is a property of what drugs to make is a very, very cool phenomenon. And you have similar areas in batteries that are. I'm not going to say much about it because you're not super public, but.
[00:25:18] Speaker B: No, of course.
[00:25:19] Speaker C: Yeah. There are some very interesting bets that we've made there.
[00:25:22] Speaker B: So, I mean, when I talked to Merlin about cathode act material and what the future looks like and the research and so on into, well, just making better materials and more pure materials, is there the same thing in the field of the battery manufacturing process with AI, machine learning? Are there ongoing research?
[00:25:40] Speaker C: There is some novelty that we will see. I'm not going to say like, we'll be able to make new materials.
[00:25:47] Speaker B: No, but I just meant, like, we're obviously not the only ones looking at this. I'm just curious if it's like an industry thing where there has been research done that's shared.
[00:25:57] Speaker C: I mean, we are obviously standing on the shoulders of certain giants.
I will give credit to the team. There is a fair amount of innovative thinking. I mean, it's sort of the Northwold way, right?
[00:26:10] Speaker B: Yeah.
[00:26:11] Speaker C: But it's also important to think about the way we have crafted. The talent at Northwold comes from different areas. It's not just battery manufacturing. And that brings forward a real cultural catalyst that we can bring people from semiconductors, or astrophysics, or health sensing, or biology, and have them transfer some of the knowledge that they are very familiar with to similar first principles, problems that we have in battery manufacturing. So I think there is a fair amount of chance that we see significant innovation in manufacturing. And we're already seeing that. If you go talk to Michael Regan, who leads the manufacturing engineering team, they've got some really interesting work going on around innovating manufacturing processes. And I suggest you invite him or someone from his team, Brendan or Nils or someone from his team to the podcast and you talk about manufacturing.
That would be the right set of people to think about manufacturing innovation.
I think we can add a lot in terms of optimizing certain ways of handling information and data. I think that's exciting. I think sensing is a very underappreciated area. We have run of the mill sensors that we deploy. The vendors sell them at 5610 times the marking the cost at which they build. And we know that that's a little bit of a monopoly because it's a challenging thing to integrate sensors in an industrial scale at that speed.
I think there we are really thinking hard on what bets to make in innovating certain sensing technologies. We are public with our investment in a company like liminal, which is using ultrasound, 2d ultrasound sensing, to sense battery defects.
So that's one example of how we are thinking about.
[00:28:21] Speaker B: Yeah, I saw. I actually saw some of those machines.
[00:28:25] Speaker C: Did you see the big arm Sean Stevenson has? It's his baby. He's brought it from a little desktop lab set up to that big setup under his flesh. And Sean Stevenson, that's crazy.
[00:28:41] Speaker B: It's in its own little sort of space now. But that's the thing is exciting when you have people coming from different parts of different industries, but they might have been exposed to or worked with something similar, but just not with the battery manufacturing process. And like you said, they could come in with very unique perspectives and innovation. I think that's super exciting. So this is going to be. I don't know, maybe you have a clear answer to this, but this is my final big picture question for you. What do you think the future looks like?
What does Norfolt et look like in 15 years?
[00:29:15] Speaker C: 15 years? Wow. With AI machine learning, I haven't thought about that. Peter keeps saying, focus close.
[00:29:24] Speaker B: But if you just think big picture for yourself, how you've seen things grow.
[00:29:28] Speaker C: And what you've seen it at 60 gigawatt hours of battery lines humming upstream.
[00:29:36] Speaker B: With data just seamlessly going through everywhere.
[00:29:39] Speaker C: Producing our active material. The robots in Reworth going twenty four seven. The full on humming factory. Yeah, that's what. And I hope it doesn't take us 15 years, but sooner than that, do.
[00:29:53] Speaker B: You see everything going seamlessly, like you said, from cell design to r and d up to manufacturing?
[00:29:59] Speaker C: I don't see another choice. Just got to be there.
We have to be there.
That is the commitment and that's the challenge that we have to take.
And we have to accept and be optimistic that we will get there together. That's where I see us going in terms of AI and data and how we use data. I think we will keep up with the rest of the industry. Maybe we'll be ahead in some areas, because I do see the culture is very different from a traditional commodity manufacturing company. Culture here is very different. So I think that will be our key enabler to really think out of the box. And my previous employer, there's a lot said about culture on why that company really excels at producing a new phone every year or a new MacBook every year, so on. So I think culture is a huge driver. Yeah, it is very unique here. That is definitely going to be a big enabler for the company and people embracing and our customers embracing AI, machine learning, modern technologies into our process and design.
[00:31:22] Speaker B: Awesome. Thank you so much.
[00:31:24] Speaker C: It is awesome.
[00:31:25] Speaker B: Is there anything else you want to add?
[00:31:27] Speaker C: No. I hope this was fun and you got a few names to bring on next.
[00:31:32] Speaker B: I do, yeah.
[00:31:33] Speaker C: You should bring every single person I mentioned. Marcus, you should bring Mike O'Regan.
[00:31:41] Speaker B: You're name dropping everyone.
[00:31:42] Speaker C: Yeah, I'm name dropping everyone. You should bring Sean Stevenson and Andreas from ETH, a few people who are really driving hard on vision.
It's Julian Eiler and my team and Eda. And there are a bunch of folks in ETh.
[00:31:59] Speaker B: That's what I love about group podcast.
[00:32:02] Speaker C: Exactly.
[00:32:02] Speaker B: We just talk about the future and what everything will look like.
[00:32:05] Speaker C: I've been watching some of those YouTube podcast things where they do a group podcast. It's actually quite fun.
[00:32:10] Speaker B: I know.
[00:32:11] Speaker C: Yeah.
[00:32:11] Speaker B: That's the goal. The goal is to get there.
[00:32:14] Speaker C: Cool.
[00:32:14] Speaker B: Okay. Thank you so much. Thanks for joining me.
[00:32:17] Speaker C: Thank you.
