Orbital Materials applies AI to the search for cleantech materials

Artificial intelligence consumes a huge amount of electricity, much of it generated with the fossil fuels that are warming the planet. But at a lab in New Jersey, scientists from the start-up Orbital Materials are using AI to discover materials that could slow climate change.

Founded in 2022, Orbital has created an AI model to predict the properties of new materials from public data about existing materials, computer simulations, and experiments in the company’s own laboratory.

Orbital is using its AI model to search for cleantech materials, such as catalysts for biobased chemical production and products to improve water treatment. The company’s most advanced material is a sorbent that it hopes will remove carbon dioxide from the atmosphere at a lower cost than current technologies.

Before starting Orbital with James Gin-Pollock and Daniel Miodovnik, Jonathan Godwin had been training AI to discover new materials at DeepMind, Google’s AI research subsidiary. But Godwin felt that Google wasn’t equipped to translate an AI model’s predictions into real materials, so he decided to step out on his own.

Orbital’s AI model is inspired by a machine learning framework that Godwin worked on at DeepMind. The model starts with a random scatter of atoms. It analyzes interactions between the atoms and uses the information to repeatedly rearrange them, moving closer to a useful new material with each cycle. The company then uses AI to simulate how the predicted materials will perform under certain conditions. “It first draws an outline and then adds in details over time,” Godwin says.

AI models have become good at predicting some properties, like whether a material will be able to absorb CO₂, Godwin says. But they struggle to predict other characteristics that affect a material’s utility, such as stability or manufacturability. Estimating those properties often requires the deep intuition that chemists develop over the course of a career.

“We needed a different type of company, a company that had really strong industrial chemists, people who had made materials before and brought them to market,” he says.

Orbital aims to go beyond synthesizing new materials. The company wants to put the materials into their final applications to ensure they work and then collaborate with big firms for scale-up and commercialization.

There’s a long list of engineering challenges that lie between an AI model’s prediction about a material and a functioning clean technology. To start solving those problems for the carbon capture sorbent and other materials in Orbital’s pipeline, Godwin recruited Thomas McDonald to be the firm’s chief scientific officer.

McDonald has worked on carbon capture materials since starting his PhD in 2008. He went on to cofound the carbon capture company Mosaic Materials, which was one of C&EN’s 10 Start-Ups to Watch in 2019. The oil firm Baker Hughes acquired Mosaic in 2022.

By most metrics, McDonald’s story was a tremendous success, but getting the technology from the lab to commercialization took 14 years. McDonald argues that the timeline for commercializing new materials needs to be much shorter, and AI could be a powerful tool to help. “If you get this right, you can really set an example for the way in which R&D gets done for the next 100 years,” he says.

Many companies are trying to move toward a world in which chemists can use AI to create new materials on a computer. They include big tech firms like Google and Microsoft, and other start-ups, like Materials Nexus. Big chemical firms want in too; BASF’s venture capital fund recently invested in the AI-focused materials discovery firm Solve.

But these companies face major hurdles before they reach that goal, according to Milad Abolhasani, a North Carolina State University researcher studying AI for materials discovery. He says one challenge is that without standardized data, it’s hard for AI models to make good predictions.

Robotic labs using AI to automatically carry out experiments on new materials will fill in gaps in the data, Abolhasani says, but he and other researchers have only recently begun to build these types of systems. For now, he says, AI is best used as a tool to augment a human’s chemistry expertise.

Godwin agrees that AI isn’t a replacement for chemists. He says combining AI predictions with human knowledge will help chemists move faster. “The software is a copilot for them,” he says. “It’s the hypothesis generation tool, a creativity aid.”