When ChatGPT launched three years ago, it became abundantly clear that the power of generative artificial intelligence had the capacity to extend far beyond clever chatbots. Companies raised huge amounts of funding based on the idea that this new, more powerful AI could solve fundamental problems in science and medicine — design new proteins, discover breakthrough drugs, or invent new battery chemistries.
Citrine Informatics, however, has largely kept its head down. The startup was founded long before the AI boom, back in 2013, with the intention of using simple old machine learning to speed up the development of more advanced, sustainable materials. These days Citrine is doing the same thing, but with neural networks and transformers, the architecture that undergirds the generative AI revolution.
“The technology transition we’re going through right now is pretty massive,” Greg Mulholland, Citrine’s founder and CEO, told me. “But the core underlying goal of the company is still the same: help scientists identify the experiments that will get them to their material outcome as fast as possible.”
Rather than developing its own novel materials, Citrine operates on a software-as-a-service model, selling its platform to companies including Rolls-Royce, EMD Electronics, and chemicals giant LyondellBassell. While a SaaS product may be less glamorous than independently discovering a breakthrough compound that enables something like a room-temperature superconductor or an ultra-high-density battery, Citrine’s approach has already surfaced commercially relevant materials across a variety of sectors, while the boldest promises of generative AI for science remain distant dreams.
“You can think of it as science versus engineering,” Mulholland told me. “A lot of science is being done. Citrine is definitely the best in kind of taking it to the engineering level and coming to a product outcome rather than a scientific discovery.” Citrine has helped to develop everything from bio-based lotion ingredients to replace petrochemical-derived ones, to plastic-free detergents, to more sustainable fire-resistant home insulation, to PFAS-free food packaging, to UV-resistant paints.
On Wednesday, the company unveiled two new platform capabilities that it says will take its approach to the next level. The first is essentially an advanced LLM-powered filing system that organizes and structures unwieldy materials and chemicals datasets from across a company. The second is an AI framework informed by an extensive repository of chemistry, physics, and materials knowledge. It can ingest a company’s existing data, and, even if the overall volume is small, use it to create a list of hundreds of potential new materials optimized for factors such as sustainability, durability, weight, manufacturability, or whatever other outcomes the company is targeting.
The platform is neither purely generative nor purely predictive. Instead, Mulholland explained, companies can choose to use Citrine’s tools “in a more generative mode” if they want to explore broadly and open up the field of possible materials discoveries, or in a more “optimized” mode that stays narrowly focused on the parameters they set. “What we find is you need a healthy blend of the two,” he told me.
The novel compounds the model spits out still need to be synthesized and tested by humans. “What I tell people is, any plane made of materials designed exclusively by Citrine and never tested is not a plane I’m getting on,” Mulholland told me. The goal isn’t to achieve perfection right out of the lab, but rather to optimize the experiments companies end up having to do. “We still need to prove materials in the real world, because the real world will complicate it.”
Indeed it will. For one thing, while AI is capable of churning out millions of hypothetical materials — as a tool developed by Google DeepMind did in 2023 — materials scientists have since shown that many are just variants of known compounds, while others are unstable, unable to be synthesized, or otherwise irrelevant under real world conditions.
Such failures likely stem, in part, from another common limitation of AI models trained solely on publicly available materials and chemicals data: Academic research tends to report only successful outcomes, omitting data on what didn’t work and which compounds weren’t viable. That can lead models to be overly optimistic about the magnitude and potential of possible materials solutions and generate unrealistic “discoveries” that may have already been tested and rejected.
Because Citrine’s platform is deployed within customer organizations, it can largely sidestep this problem by tuning its model on niche, proprietary datasets. These datasets are small when compared with the vast public repositories used to train Citrine’s base model, but the granular information they contain about prior experiments — both successes and failures — has proven critical to bringing new discoveries to market.
While the holy grail for materials science may be a model trained on all the world’s relevant data — public and private, positive and negative — at this point that’s just a fantasy, one of Citrine’s investors, Mark Cupta of Prelude Ventures, told me over email. “It’s hard to get buy-in from the entire material development world to make an open-source model that pulls in data from across the field.”
Citrine’s last raise, which Prelude co-led, came at the very beginning of 2023, as the AI wave was still gathering momentum. But Mulholland said there’s no rush to raise additional capital — in fact, he expects Citrine to turn a profit in the next year or so.
That milestone would strongly validate the company’s strategy, which banks on steady revenue from its subscription-based model to compensate for the fact that it doesn’t own the intellectual property for the materials it helps develop. While Mulholland told me that many players in this space are trying to “invent new materials and patent them and try to sell them like drugs,” Citrine is able to “invent things much more quickly, in a more realistic way than the pie in the sky, hoping for a Nobel Prize [approach].”
Citrine is also careful to assure that its model accounts for real world constraints such as regulations and production bottlenecks. Say a materials company is creating an aluminum alloy for an automaker, Mulholland explained — it might be critical to stay within certain elemental bounds. If the company were to add in novel elements, the automaker would likely want to put its new compound through a rigorous testing process, which would be annoying if it’s looking to get to market as quickly as possible. Better, perhaps, to tinker around the edges of what’s well understood.
In fact, Mulholland told me it’s often these marginal improvements that initially bring customers into the fold, convincing them that this whole AI-for-materials thing is more than just hype. “The first project is almost always like, make the adhesive a little bit stickier — because that’s a good way to prove to these skeptical scientists that AI is real and here to stay,” he said. “And then they use that as justification to invest further and further back in their product development pipeline, such that their whole product portfolio can be optimized by AI.”
Overall, the company says that its new framework can speed up materials development by 80%. So while Mulholland and Citrine overall may not be going for the Nobel in Chemistry, don’t doubt for a second that they’re trying to lead a fundamental shift in the way consumer products are designed.
“I’m as bullish as I can possibly be on AI in science,” Mulholland told me. “It is the most exciting time to be a scientist since Newton. But I think that the gap between scientific discovery and realized business is much larger than a lot of AI folks think.”
Zanskar's drilling site in northern Nevada. Zanskar
