CuspAI: A Search Engine for Materials That Don’t Exist Yet

[00:00:00]At Cusp, we've built what can be described or thought of as a search engine for materials. And so, for many years, as a human race and species, we've relied on serendipity and trial and error to discover new materials.

[00:00:13]And the space in which we have to explore for new materials, which we call chemical space, is extremely large. In fact, it's larger than all possible grains of sand on the Earth. Now, with the advent of AI, and in particular generative AI, we're able to move to a new era where you're able to generate the exact materials that you're looking for with the properties that you need for your application. And so, the search engine that we built at Cusp enables you to do exactly that. And this, of course, has very large potential implications across the semiconductor industry, the energy industry, and also a very positive impact on the climate side of things, too.

[00:00:48]>> So, materials discovery this typically takes 10 to 20 years to actually do it correctly and and find kind of a novel material that's useful. Cusp is doing this in months.

[00:00:57]>> Well, ultimately, it's the advent of AI and the progress in AI. So, the way it happens is the material that you're interested in, including like all the properties that you want optimized, you would type it in the search engine. And then, an agent receives that information and it builds a plan, a workflow.

[00:01:15]The first step typically is to search the literature. We have a large proprietary database of all sorts of experimental data as well as the scientific literature data. It will search that for good candidates, but often it's not good enough. And then, it will start to generate entirely new materials, materials that have never seen the light of day in the universe.

[00:01:35]And we can generate, you know, hundreds of thousands to millions of those. Those get then subsequently evaluated by many of our property predictors. And then, after that, we'll do a molecular dynamics simulation to further explore the properties of this particular material. And then, finally, that gets sent to an experimental facility uh, where it's actually made and characterized by, uh, sort of experimental setup and measuring machines.

[00:02:00]>> Past AI attempts often generated materials that couldn't actually be made. So, it's one thing to say, "Hey, we're going to go explore and find some novel material." but it's another one to actually solve synthesizability.

[00:02:13]Like, what allows Cusp platform to help solve synthesizability in addition to whatever this novel material that will solve the problem is?

[00:02:20]>> There's a number of things we can compute to make sure that the material is what we call in a low energy state, which means basically it's stable. It will actually form and stay in that particular state. But, we also have dedicated predictors. So, we typically look up the material in the database and we find similar materials. And then we look up what the synthesis process has been for that particular material. And then our agents will reason about a new synthesis process for the new material that we generated and then we recommend that to the experimental setup.

[00:02:52]>> So, Chad, you grew up in a Welsh farming village and were the first in your family to attend university. Max, you did your PhD under a Nobel laureate and co-invented VAEs. How did you two end up building this together?

[00:03:04]>> I was working in the quantum computing space and then I saw a paper that came out of Microsoft Research called Matter Gen, which was the first generative model or one of the first generative models for material science. And of course, the whole world knows who Max Welling is and his reputation precedes him in AI and machine learning. And so, I reached out to Max to learn more about the Matter Gen project and in the process learned that Max had actually just resigned from Microsoft and had ambitions to also think about building a startup in this space. And so, when I reached out to Max, he'd also had a very similar philosophy. He wanted to build a startup and he wanted to focus on material science because of the work that had started to come out of Microsoft Research. After several conversations and iterations, um, um, Max and I decided to make the leap.

[00:03:49]>> Cusp today works with enterprises like Meta, Hyundai, Kemira.

[00:03:54]Can you give a concrete example of a material Cust help discover and its real-world impact?

[00:03:59]>> We're the company where a little over 2 years old and so in that time we've built our team and our technology.

[00:04:05]And now I have it deployed across many different deployments and customers. But one of the most forthcoming use cases I think that's worth talking about here and in the public domain is a project that we have with a Finnish water chemical company. They're a multi-billion euro company called Kemira. And Kemira recognized this was led by the CEO of Kemira. They recognized that global regulation around these things called PFAS, so forever chemicals, was coming into force.

[00:04:29]Now these are things that have for the last 50 years or so of lined things like our frying pans. They're very heavily used in the semiconductor industry. And by design, they're designed to never degrade. And so they're really, really difficult. And now they're aggregating and accumulating in our waterways and even in our bloodstream. I think around 99% of us now have PFAS molecules building up inside our bloodstreams.

[00:04:51]And we've started to realize the toxic effects of these.

[00:04:54]And so they're also notoriously difficult to remove from wastewater or other streams. And so Kemira didn't have existing materials that were capable of removing these from the water in such an efficient way. And so the task set to us, which is a very difficult task, was can you come up with an entirely new material set that allows us to efficiently remove these PFAS molecules from water and that will have of course a very positive environmental impact, but also a very large commercial opportunity.

[00:05:23]>> So Cust is in building, I guess what you're calling materials foundry partnerships. What exactly are those and what do they signify about where the industry is heading?

[00:05:30]>> As we think about this new paradigm or this new era of materials discovery, we call it the materials on demand era, there are five core components that I think you need in order to make this happen or make it feasible. But the philosophy at Cust has been how do we bring together these five components globally to enable not just individual companies like ourselves to discover new materials, but an entire new ecosystem across many different sectors. And so, what we're doing is we're launching our AI materials foundry in three regions, European hub, US hub, and also in APAC.

[00:06:04]And the foundries bring together the five components that I talked about. So, one would be of course the first component you need for any AI system is high-quality data sets. And at Cast we now believe we've aggregate the world's largest curated data set for material science, and we're making this available via the foundry network, and making this available to our AI models to do the discovery.

[00:06:24]The second thing of course you need is a platform, and that's the search engine that we've described, that we've built here at Cast. This is an agentic platform that is able to do generation, simulation, and evaluation of candidates using that data, and also experimental data that we generate.

[00:06:40]And then the third thing that we need of course is actual synthesis capability, so lab capabilities where you ground your simulations now in reality by running physical experiments.

[00:06:50]And then the fourth thing you need of course is compute, and so we're using local computing infrastructure and these AI factories that Nvidia have been busy building.

[00:06:58]And then also, we also need customers with the right domain knowledge to say, "Hey, we really need this material to solve our problem."

[00:07:06]And of course the companies that we work with, the likes of Kemira's and the Meta's of this world, they know their domains far better than we ever will.

[00:07:13]And so, if they bring their domain experts we combine that with our platform, our team, and our data, and the compute, we're able to develop new materials in a much faster way. And these foundries are underpinned both by Meta and Nvidia who are kind of two founding members, and we now have of the order of about 40 industrial partners across the semiconductor industry and other spaces as well. So, it's a really large global initiative. It's an ecosystem. And one of the key things we're also finding with the foundry is that very often even in individual spaces like the semiconductor industry, the communication channels between different parts of the value chains are sometimes and more often than not broken. And materials is kind of the currency of the semiconductor industry all the way from the likes of ASML through to the ALD providers, the foundries, materials really underpin the whole space. And so we believe at Cusp we have a really integral role to play in shaping materials of the future but also making sure that each part of the value chain understands which materials are required for the next node and linking all of that together. So it's really a an aggregation but also more of an ecosystem that brings together all of these different players.