AI-Powered Intelligence Could Transform Materials Design

I have always been fascinated by very fundamental science – the equations that describe reality and matter, and so on – but I’ve also been drawn to the applied side of it as well, where we can look at how to use this knowledge to solve big, important problems.

I was doing research in semiconductor physics, and it was around this time that I started getting interested in the increasingly data-driven methods we use for science today. This interest led me to become a machine learning engineer. I started to work on some different problems that were outside my core experience – things like self-driving autonomous vehicles or industrial Internet of Things solutions. I was familiar with software engineering, machine learning and applied AI principles, and when I took a step back, I really wanted to bring this together with the applications that I really care about – solving challenges around climate change, creating novel materials for the future of electrification, decarbonizing industry and this emerging world of AI in science.

I did my PhD with Dr. Sam Cooper, a professor at Imperial College London, and that is where I met my co-founder, Dr. Steve Kench, as well. In our research group, we were working at the forefront of these areas, looking at how we can develop and use generative AI models to solve some of the biggest challenges in materials science, with a focus on battery materials and electrochemistry.

Over the past six or seven years, our group’s research has had some great traction in the academic community, including a cover article in Nature Machine Intelligence with some of our early models. What gave birth to Polaron was the interest we started getting from big industrial manufacturers and engineering teams who had seen our research and wanted to talk to us. Once we’d had a glimpse at what was happening in industry circles and where the biggest challenges were, we realised that we could do so much more than just publishing cool papers, but to get our powerful AI models into the hands of engineers, we would need to build them into industry-grade tools.

Simply put, it’s about length scale.

Most companies in the AI for science space today are focusing on atoms and molecules - discovering new arrangements of these tiny building blocks that lead to desirable properties. At the other extreme, you’ve got companies focused on the design of components and products, like the optimal shape of an airplane wing.

Then you have the in-between length scale, which is what we call the “microstructural” length scale. Here you have billions and billions of atoms (too many to simulate directly!), but it’s not yet millimeters or meters long. At this length scale, you have many important materials science phenomena occurring that fundamentally change the way that materials behave.

For example, in a battery material, its microstructure – the shapes and arrangements of the particles  – will come together and fundamentally change how fast you can charge and how long it will last.

The way that we manipulate this length scale is by controlling industrial processes used to make the materials      (think mixing, heating, cooling, crushing, rolling, etc.). This is super messy, but this is where some serious science happens. What we are doing is understanding the relationships between processing, structure and performance at this length scale using image data.

Crucially, we’re using image data to train our models. There is so much underutilized image data of materials out there in the world – billions of images of materials are being collected every year to make decisions. There is a lot of hidden value in this microstructural image data, and what we are doing is turning that data into something that is much more searchable and learnable, so then we can understand more quickly how these factors link together and then optimize the way these materials behave at that length scale.

The real “North Star” question here is working out how we can connect processing, structure and performance together. Polaron’s approach is to split it into two steps.

Firstly, we have characterization. Characterization is all about trying to understand what you have already made. This means taking raw microscopy and image data and turning that into quantifiable, reliable insights about the material.

In industry today, this is still nearly all of this is done by eye, so it takes a very long time and it can also be quite subjective. With AI, we can rapidly accelerate this; we have saved one of our customers around 1,000 engineer hours so far this year with our automated image analysis workflow. To achieve this, we have vision models that understand materials science and put this into a workflow where we have human expertise in the loop where necessary, so we can get quality insights more quickly.

Crucially, Polaron can generate 3D microstructural data from 2D microscopy images. This means you don’t need 3D imaging to get those insights, which saves a lot of time and money. You can start learning things from this data that you couldn’t know before.

The next step is design. While characterization asks the question “what have we made?”, design asks “how do we make this better?”. And specifically, how can we do this within the constraints of what is realistically manufacturable today?

We use this image data to learn the relationships between the way that these materials are made, the structure seen in the microscopy data and their performance from either real-world performance data or physics simulations based on the material’s structure. Once we have that model, we can start to explore the space of possible designs. We aren’t just identifying whether a structure performs better, but also the “recipe” in terms of how to actually make it.  

Materials design problems are actually very “local” – the question is really “how can I get the performance I want with the materials and machine that I have?” The exact interactions between material components, the manufacturing processes, the equipment, the geography, the operator – all of these things influence the design and mean that it cannot be trivially mapped to a new context. There simply is no “best” electrode recipe. That makes it hard to make a generalized model for manufacturing.

Customer data is never used to train any base models or foundation models. Customer data is entirely siloed, and the models that are trained on that data are only ever used to deliver that customer their insights. Then, separately, we are also collecting some proprietary data of different types of material systems in order to make our base models better.

At the moment, our core focus as a company is the battery space. So, looking at the design and manufacturing of battery electrodes, regardless of battery chemistry or manufacturing technique. We also have existing collaborations and partnerships with customers working in construction materials and in alloys.

This technology is super generalizable into other kinds of sectors; it is entirely materials agnostic. Any area where you can image the structure of a material and that structure can be changed by how it is made – so, essentially all solid material systems – that is where we can have an impact. Think alloys, composites, catalysts and concrete.

We are even looking at industries such as pharmaceuticals. If you think about it, those oral slow-release pills, it is the microstructure of the pill coating that determines how that drug is released. It is similar for the food industry as well, both with packaging and food itself.

As a scientist and engineer, I find it very difficult to commit to something without any kind of uncertainty attached to it! But I will give you one hypothesis.

As I’ve said, materials science and more generally, with engineering and the chemical industry too, is not a simple, well-constrained, well-scoped problem. You can find simple problems within these spaces, but that isn’t the whole space. I think we need to solve the whole thing if we are going to really change the way that industry works.

To do that, we need to start thinking about this as a multi-scale, interconnected design problem. Right now, we have tens of different professions for all of these different parts [of the industry]. This is where the power of these intelligent systems that we are developing can come in. They can look at these huge amounts of information and synthesize it and hold all of that in memory. As humans, we struggle to do that.

Where AI can be a sort of connective tissue between discovery, processing and manufacturing, that is where we are going to see the long-term benefits of AI. We will have fewer experiments, smarter design loops and better connections between all of these different stages. That will really enable this continuous data-driven optimization. That is how I see the future of the industry.