Establishing Trust in Machine Learning for Battery Testing and Development

Monolith is a company whose mission is to empower engineers to use AI and ML – always with the goal of bringing products to market faster. That last part is important because we operate in the commercial space, essentially taking the latest and greatest developments from academia, validating them, seeing if we can reproduce them and make that applicable in an industrial setting, and then make them available to the greatest R&D teams across the globe. That’s the value we add.

When it comes to working with AI and ML, the majority of the work we do is still grounded in traditional machine learning and first-generation AI. We look at things such as material selection, data analysis, finding new insights, automating data review and running simulations and predictions of intractable physics based on data-driven models within a confined design space. This is what we do, and we have specialized across multiple verticals now, one of those being the battery world.

I think the main issue we're seeing right now – and you do have to look at this from a political perspective – is that in Europe and the US, there’s a scramble to ramp up new variants and forms of lithium iron phosphate (LFP) batteries. If LFP is the dominant chemistry, we have more-or-less zero production of it in the West.

That means we not only need to develop new LFP batteries but also ramp up manufacturing. That requires incorporating the criteria for manufacturability into the R&D process. In Europe and the US, we don’t have experience with that – we’re way behind, and there’s more than 20 years of work to catch up on.

At the opposite end of the spectrum, you also have these very early-stage, completely new types of batteries. There, the challenge is more around improving how we select materials, screen them and formulate them – how to do that faster and better.

I think we can divide this into two: the very beginning and the very end of the process.

At the very start, the challenge when you're experimenting with novel materials or new combinations of materials is that you often don’t really know what to test. You have foundational knowledge in chemistry, physics and electrochemistry, of course, and you do have people with experience in this, but when you’re working with something completely new, you cannot apply your experience to make the R&D process faster.

This is a perfect example of a problem where ML and AI can be applied – to help narrow down what you’re going to do and to learn from the data you get back much more quickly. Crucially, you can do this iteratively. Rather than following the typical human approach, which typically involves running longer, experience-based tests before adjusting course, AI and ML let you iterate much, much faster, and that can help you tremendously.

Then, towards the end of the process, you can use AI/ML models based on the battery cells – essentially building digital twins of the battery cells – that are data-driven models, or sometimes hybrid physics- and data-driven models. You can combine this with the test data from the ramp-up of your manufacturing lab.

The battery in the lab is very different from the battery in manufacturing, but you need to simulate manufacturing conditions in the lab to make sense of what's going to happen at scale. So essentially, you’re running two types of experiments at the same time, and you can use AI and ML to combine those two in ways that human beings typically can't. We don’t have the expertise to go across these two domains and spot correlations or causal relationships across these stages so easily. That’s another space where AI and ML can add tremendous value.

When people talk about trust in AI, they typically focus on the performance of the model – how well does it work? What are the statistics? And then from that, they say, “Here are the numbers, this is the performance, this is why you should trust it or not.”

But in my opinion, trust in AI isn’t just about how well the model performs. It is just as much about how you interact with your model and your AI.

For example, at Monolith, we initially started as a pure AI and ML company – we weren’t fundamentally battery experts, but we could see that batteries were a great use case for our technology. So, we started applying our AI models to batteries, but we had to work very closely with customers to validate every step of the process.

Naturally, our customers didn’t always enjoy that because it takes time, and it’s cognitively taxing and “AI should just fix everything,” right? But what happened was, by interacting with the model over time, the customers began to build a relationship with the model and started to trust it. They knew where it was good, where it was bad and they could also see that performance was improving over time – and so, they trusted it.

It’s more about how you interact with AI and how the model changes its behavior over time when you are giving it this additional input, as opposed to how good AI or ML is.

To me, I think the question for our industry and vendors like ourselves is, how do we innovate and enable this interaction with the models and the outcomes without this becoming a very time-consuming and cognitively taxing task? If we can get that down and really innovate on user interfaces, innovate on interactions and make sure that we are showing you only relevant information for review, I think that will be the key determining factor for AI adoption in R&D.

Data quality is a huge problem. The good thing is that a lot of the technologies and toolkits used to improve data quality have become significantly cheaper, easier to use and more accessible. There are more people now who know how to use them, so the cost of producing good data has dramatically decreased over the years. This is something that benefits our side of the industry a lot, but nevertheless, it is still a big problem.

Another development we’re seeing now is that a lot of the very manual, tedious data-cleanup processes can now be addressed with LLMs and AI agents. This is still very nascent, and some people don’t fully trust it yet. But I do. I believe it’s going to be a huge game-changer in terms of data quality. There are certain things that you currently need to do manually, and that increases the cost of producing high-quality data to the level where some people can’t, or simply won’t, do it. AI agents based on LLMs will change that – it will reduce the cost of producing high-quality data.

My opinion is that with the usage of AI, engineering – and especially in certain areas such as batteries – will be completely unrecognizable in five years.

It’ll take a bit of time until things like Sam Altman’s ChatGPT fully find their way into labs – it will start slowly, but once it gains traction, it will most likely change everything fundamentally. The reason I say that is because machine learning will still be the main way we can improve the R&D process, but the ease of use and the relevance of these tools will increase by an order of magnitude.

For example, let’s say you’ve spent 25 years developing a new cathode. You’ve done the research, and now you have a material with certain performance parameters that are better than anything else we have seen. Maybe you’ve got a team of five researchers working in this lab, so it’s very small-scale, but you can prove that it works. But now you need to bring it closer to production and commercialization. You need to combine that cathode with an anode, electrolyte, separator, all these other pieces and you need to optimize them to make the best use of your cathode. But you’re a cathode researcher; you’ve spent 20 years looking at cathodes, not these other problems.

Now, what you can do by incorporating LLMs and working with PhD candidates who are in all these various fields, is you can narrow down all these other parameters that you need to know about much faster, to produce a great product.

It could flip the competitive landscape: where once conglomerates had a monopoly on all these adjacent fields, now startups can do this too.

I believe that in five years, we will have the software to support all of this. And in 10 years, you will have the results – there will be a flood of new products in the market because of this.