Scientists at an unnamed corporate lab were randomly assigned a machine learning tool.Credit: Eugenio Marongiu/Getty
Artificial intelligence (AI) is becoming ubiquitous in applied research, but can it really invent useful materials faster than humans? It’s still too early to tell, but a massive study suggests this could be the case.
How close is AI to human intelligence?
Aidan Toner-Rodgers, an economist at the Massachusetts Institute of Technology (MIT) in Cambridge, tracked the deployment of a machine learning tool at an unnamed corporate lab employing more than 1,000 researchers. Teams randomly assigned to use the tool discovered 44% more new materials and filed 39% more patent applications than those who stuck to their standard workflow, it found. Toner-Rodgers put the results online last month, and submitted them to a peer-reviewed journal.
“It’s a very interesting paper,” says Robert Palgrave, a solid-state chemist at University College London, adding that limited disclosure of trial details makes the results of AI deployment difficult to assess. “It perhaps doesn’t surprise me that AI can come up with lots of suggestions,” Palgrave says. “What we’re kind of missing is whether those suggestions were good suggestions or not.”
Material manufacturer
Toner-Rodgers was given access to the lab’s internal data and interviewed the researchers on the condition that it not disclose the name of the company or the specific products it designed. He writes that it is an American company that develops new inorganic materials – including molecular compounds, crystal structures, glasses and metal alloys – for use in “health care, optics and industrial manufacturing.
Researchers have built an “AI scientist”: what can it do?
From 2022, the company has systematically adopted an AI tool that it has personalized to meet its needs. According to Toner-Rodgers, the tool combines graph neural networks – a popular approach in materials discovery that has been used, among others, by DeepMind, Google’s London-based AI company – with reinforcement learning. . The neural network was pre-trained using data from large existing databases, including crystal structures and their properties from the Materials Project and molecular structures from the Alexandria Materials Database.
Researchers input requirements for a material’s desired properties into the neural network, and the system suggests structures for new materials that might have those properties. Teams then eliminate potential hiccups – such as formulas that would not lead to a stable compound – using their own specialist knowledge and computer simulations. They then attempt to synthesize candidate structures and, if successful, test them experimentally and even on prototypes of finished products. The results are fed back into the neural network – the “strengthening” stage that helps it improve its predictive capabilities.