Synthetic fuels and more thanks to machine learning

In addition to laboratory experiments, a considerable amount of work is carried out on the computer. Image: Empa

The most common electrodes used in CO₂ electrolysis are gas diffusion electrodes. These are porous structures that are coated with the active catalyst material – usually copper or silver. The shape, arrangement and density of the pores can influence the results of the catalysis. However, it is not yet clear exactly how they do this.

This is what Bozal-Ginesta has set out to change. To do so, she is combining two things: machine learning and the high-throughput system for research into CO₂ electrolysis, which was established in the Materials for Energy Conversion laboratory (to the article). “To use ML, you need large data sets,” explains the researcher. The system, which can characterize eight to ten catalyst candidates simultaneously, is the key to this. Bozal-Ginesta plans to combine the resulting data with various microscopy and spectroscopy methods.

Machine learning will be used in three different ways. “First, I want to understand the correlation between the structure of the electrode and its electrochemical performance. Secondly, I want to develop ML-based tools that can identify useful structural features in microscopy images of the electrodes. And finally, I want to use this data to derive realistic new electrode structures that could have promising properties,” explains the researcher.

Bozal-Ginesta is particularly mindful to use the ML technology in a responsible fashion. “Machine learning and artificial intelligence are often used in a 'brute force approach' – with limited data curation and little consideration of existing knowledge,” she explains. If the data set is incomplete or the models are inaccurate, the results will be misleading. “It is still up to us, the scientists, to formulate hypotheses and set priorities, i.e. propose and validate which information matters most in performance modelling and prediction. ML is a powerful tool for testing our hypotheses. But only we can decide what's really important,” Bozal-Ginesta is convinced.