Machine learning (ML) transforms the design of heterogeneous catalysts, traditionally driven by trial and error due to the complex interplay of components. BIFOLD researcher Parastoo Semnani from the ML group of BIFOLD Co-Director Klaus-Robert Müller (TU Berlin) and additional researchers from BASLEARN, BASF SE, and others have introduced a new ML framework in the Journal of Physical Chemistry C.

Machine learning (ML) models have recently become popular in the field of heterogeneous catalyst design. The inherent complexity of the interactions between catalyst components is very high, leading to both synergistic and antagonistic effects on catalyst yield that are difficult to disentangle. Therefore, the discovery of well-performing catalysts has long relied on serendipitous trial and error.

In the paper, the researchers introduce a machine learning framework that deals with the challenges of experimental data and provides robust predictions of catalyst performance. Additionally, they incorporate explainable AI methods in the framework that help determine which catalysis components contribute more strongly towards high-performance catalysts.

The high costs associated with generating experimental catalyst data often result in small datasets biased towards low-performance catalysts.

"We believe that our framework combines best practices in the field and provides a conceptual blueprint on how to work with and analyze experimental catalyst data, which should prove useful to future machine learning research efforts in this field, and help push AI-assisted Catalyst design forward," concludes Semnani.

This framework tackles small, unbalanced datasets and predicts catalyst performance robustly. By integrating explainable AI, it identifies key catalyst components driving efficiency. This innovative approach offers a blueprint for future AI-driven breakthroughs in catalyst discovery.