To narrow down the huge range of possibilities, the researchers made a preselection based on experience and economic requirements. A catalyst that can be used on a large scale needs to be not only effective but also inexpensive. For that reason, the main active ingredients for the catalyst were limited to three comparatively cheap metals: iron, copper and cobalt.
In addition to these main metals, the researchers considered three elements that are traditionally added to catalysts in small quantities for the purposes of doping, as well as potassium, which is also contained in many catalysts. As to carrier materials, the researchers limited themselves to four typical metal oxides. Multiplied by the different mixing ratios, this still resulted in 20 million possible combinations.
Taking iterative steps with AI-supported statistics
At this point, the researchers brought an AI algorithm into play that uses what is known as Bayesian optimisation to find the best possible solutions. This special form of statistics is particularly suitable when only a small amount of data is available. Unlike in classical statistics, the probability doesn’t derive from the relative frequency as calculated from numerous experiments. Instead, the calculation takes into account the probability that can be expected based on the current state of knowledge.
In the initial round, the algorithm randomly selected 24 catalyst compositions that met the specifications drawn up for the purposes of limiting the complexity. These catalysts were produced directly using the Swiss Cat+ automated laboratory infrastructure and then tested.
Delivering lots of highly reliable results quickly
The results of this initial selection served the researchers as the starting point for an AI prediction; the catalyst compositions thus predicted were in turn automatically synthesised and tested. For this first demonstration test, the scientists had their integrated system complete a total of six such rounds.
The fact that the results improved between rounds not in a linear fashion, but rather by leaps and bounds, was entirely intentional: not only does the algorithm optimise the results of earlier rounds, it also includes an exploratory component that feeds completely new compositions into each round and learns about the chemical space. This is how the researchers prevented the calculations from getting stuck in an optimisation dead end amongst all the possibilities.
Generating data beyond petrochemicals
In this first project, though, the researchers’ primary concern wasn’t to come up with the best possible catalyst for methanol synthesis. “At present, knowledge about catalysts for fuel production is based predominantly on expertise from the oil industry,” Copéret says. “When it comes to reactions for use in the sustainable energy industry, reliable data is still largely lacking.” However, AI algorithms and human research intelligence need that data before they can search in a more targeted way in the vast space of chemical possibilities. “And that’s precisely the kind of high-quality, reproducible data our AI-assisted robot laboratory now delivers. It’s certain to take catalyst research a long way forward,” Laveille adds.
