Chemists have discovered four new materials based on ideas generated from a neural network, according to research published in Nature.
Uncovering new materials is challenging. Scientists have to search for combinations of molecules that lead to useful compounds that can be manufactured.
Traditional methods rely on fiddling around with known materials, and although these techniques narrow down the search for materials that work well, they don’t always produce something useful, according to Matt Rosseinsky, a chemistry professor at England’s University of Liverpool who co-wrote the research paper.
“To date, a common and powerful approach has been to design new materials by close analogy with existing ones, but this often leads to materials that are similar to ones we already have,” he explained
“We therefore need new tools that reduce the time and effort required to discover truly new materials, such as the one developed here that combines artificial intelligence and human intelligence to get the best of both.”
Rosseinsky and his colleagues turned to a neural network made up of nine layers and over 50,000 parameters. The team fed the software examples of known solid state materials from the Inorganic Crystal Structure Database, a dataset containing at least 200,000 inorganic compounds.
The neural network shuffles the combinations of chemicals to generate new ones for scientists to study. These outputs are ranked by the software on how likely they are to produce materials that are novel and possible to produce in a lab.
“A material is only discovered once it has been made and measurements have confirmed what it is,” Rosseinsky told The Register. “Four materials with elements highlighted by this model have already been synthesised in the laboratory. Material discovery cannot be a purely computational exercise, and experimental realization is the only true validation of the model.”
The four materials crafted from hundreds of possible outputs by the AI model are a family of crystalline solids that conduct lithium atoms, according to the academics; the team believes they could be useful in batteries for electric cars one day.
Coming up with new materials using AI is a collaborative effort that requires human expertise. Scientists still need to rely on their expert knowledge to select the best training inputs to give to the neural network and scrutinize its outputs to decide which potential materials are worth synthesizing in real life, Rosseinsky explained.
“Human expertise evaluates the ranking produced by the model, and selects the element combinations of interest for evaluation by the model. The model offers complementary numerical guidance that is distinct from the understanding and expertise of the human researchers. The human researchers make the decisions,” the professor added.
The team hopes to explore how humans and machines can work together on science. “The motivation of the study is to support and guide expert human researchers in realizing materials in the laboratory. We will apply the model to a broader range of chemistries, aiming to discover new materials in other application areas. We are also interested in extending the model to rank materials based on their likely performance as well as synthetic accessibility.” he concluded. ®