AI models are dreaming up the materials of the future

Made of metal ions held together by compounds containing carbon, MOFs come in a dizzying array of structures, each with its own distinct properties. A MOF capable of absorbing CO2 at a humid sea-level location, for example, will have a different structure from one that can operate in a dry, high-altitude climate. Sorting through the billions of possibilities to find the right MOF for the job is an almost impossible task for a human chemist. It is, however, a perfect task for an artificial-intelligence (AI) model.

One startup that is building such a system is CuspAI. It uses a multitude of AI models in concert: some are trained to generate candidate molecules with prescribed properties, which get passed to a specially trained foundation model to assess their properties. CuspAI’s goal isn’t simply to find a good MOF, but to build a system that can spit out the right one for any environmental conditions—and, from there, to demonstrate that AI can be used to tackle any problem in materials science. Better batteries, cleaner bioplastics, more powerful semiconductors and, potentially, even room-temperature superconductors might soon be up for grabs.

This is no pipe-dream. In a conference paper in November 2024, Aidan Toner-Rodgers, a doctoral student in economics at the Massachusetts Institute of Technology (MIT), analysed the effects of a new AI tool on the productivity of materials researchers at an unnamed American company. Thanks to a staggered launch, with the thousand-odd scientists at the firm getting access to the tool in three distinct groups, Mr Toner-Rodgers was able to treat its introduction like a randomised experiment, and estimate its impact. The results were impressive: a 44% increase in the number of materials discovered, a 17% rise in product prototypes that used those new materials and a 39% increase in the number of patents filed.

Insofar as it can be measured, Mr Toner-Rodgers says, the resulting innovations also seemed to be more genuinely novel. AI-assisted patents were more likely to mention new technical terms, and the materials themselves boasted more unfamiliar physical structures.

But whether you use AI models or not, notes Aaike van Vugt, a Dutch chemical engineer, materials design remains “a pain in the ass”. Some challenges are technical, with the production of new materials often requiring bespoke manufacturing facilities capable of pumping out vast quantities at speed. Others are financial, with companies focusing on early research and development struggling to find a way to turn discoveries into profit.

Other industries have already squared this circle. Stef van Grieken, the co-founder of Cradle, an AI protein lab with offices in Amsterdam and Zurich, describes the pharmaceutical industry as “private equity with laboratories attached”. Clinical trials for new medications may be unforgiving, but they encourage investments that distribute risk and reward throughout the industry, funnelling resources back to the researchers in the trenches. There is no such luxury for materials scientists: those in the business of designing a material must inevitably work out how to test, manufacture and sell it too.

That has not deterred CuspAI. It hopes to build a platform that can design materials to order, leaving it to larger companies with labs and fabrication facilities to do the testing and manufacturing.

MOF to the races

London-based Orbital Materials is also using AI to try to build a MOF. The company has trained its own model from scratch, using supercomputer simulations to generate training data, says Jonathan Godwin, a former researcher at Google DeepMind who co-founded the business. The end result is hundreds of millions of simulated chemical interactions, each made up of just a couple of hundred “tokens”: advanced versions of the terse chemical reactions that fill a high-school textbook. That is orders of magnitude less training data than is required to train a large language model but, Mr Godwin hopes, more than enough to build a small and efficient model that can accurately predict chemical interactions.

But rather than operate as a purely virtual lab like CuspAI, building an AI and selling the discoveries it makes, Orbital is prepared to get its hands dirty. Its foundation model has already spat out a number of candidate MOFs, and Orbital has invested the time and money in in-house labs and chemical engineers to verify that they work and can be manufactured at scale. In December it announced a deal with Amazon Web Services, a hyperscaler, to integrate one discovery into one of the company’s vast data centres, where the waste heat of the air-cooling system will power the chemical reaction that scrubs CO2 from the air. The goal is to turn the data centre carbon negative, for a cost of 20 cents per hour per chip. If it works, Orbital will have turned an AI-generated invention into a functional product faster than anyone in the pharmaceutical industry.

Other companies are trying to automate away the need for laboratories entirely. Mr Van Vugt, the chemical engineer, is one. His startup, VSParticle, offers what is, in effect, a nanoscale 3D printer: using a technique called spark ablation, it builds up a thin film of novel materials one nanoparticle at a time, following a recipe unique to each material. Such films can be used in batteries or as catalysts. If widely adopted, Mr Van Vugt argues, it could save materials scientists the hard work of figuring out how to physically produce a desired candidate. Instead of worrying about synthesis, they could simply email the recipe to VSParticle’s lab and wait for the end product to be printed in one of the company’s automated fabricators.

Automation has gone further still. In 2023 scientists from MIT showed that an AI-enabled robot could predict, make and analyse almost 300 new chemical dyes, leading to nine engineered to have properties highly desirable in biomedical imaging. In 2024 a group led by researchers at the University of Toronto presented an AI agent that managed (albeit with some help from humans) to create a world-beating gain material—the light-amplifying substance—for a laser.

Using a combination of AI and robotics as a shortcut to synthesising new materials would be huge, says Max Welling, a co-founder of CuspAI. But, he warns, “Recipes are very finicky.” Even minor differences in humidity or air quality can scupper a lab’s chances of making the desired product. That is even truer for labs run by robots, which has led some to question their results. In 2023 researchers at A-Lab, an automated lab at the Lawrence Berkeley National Laboratory, claimed to have made 41 new materials predicted using data from Google DeepMind and the Materials Project, an initiative looking to simulate the properties of all inorganic materials. The announcement was impressive, but questions regarding the model’s analysis have led some chemists to question whether any new materials were actually produced. The A-Lab team stands by their approach.

For now, there is reason for cautious optimism. In November 2024 Meta, a technology giant, announced a partnership with VSParticle and the University of Toronto that has funded the creation, analysis and digitisation of more than 500 experimental electrocatalysts—a category of materials that could be crucial to powering next-generation batteries. The company’s big data centres aren’t always running at maximum capacity, said Larry Zitnick, research director at Meta’s AI division. That left spare computing power which Meta was able to donate to the project to provide the initial simulations for those electrocatalysts.

For Chad Edwards, CuspAI’s other founder, more is at stake than just a new carbon-capture material. If his company’s bet pays off, it would be a chance to show that AI can actually make meaningful contribution to science.