1. How does AFE technique work on small data for complex materials?
The AFE technique, developed in this study, works on small data for complex materials such as solid catalysts without requiring any prior knowledge of the target system. It consists of a pipeline that includes assigning a series of features to materials of arbitrary compositions, synthesizing a large number of higher-order features considering nonlinear and combinatorial effects, and selecting a feature subset in the context of supervised ML. This approach allows for automatic feature engineering, overcoming challenges in descriptor design and enabling the study of various heterogeneous catalysis with different catalyst designs. Additionally, the AFE technique can be extended to active learning in combination with high-throughput experimentation (HTE) to refine a feature set and obtain a globally fit model.
read more
2. How does AFE contribute to ML for small data of solid catalysts?
AFE, or Active Feature Engineering, is a versatile technique that enables effective machine learning (ML) for small data of solid catalysts with diverse compositions. It designs highly expressive features specific to a given catalyst system without requiring prior knowledge of the system. The availability of process-consistent datasets obtained through High-Throughput Experimentation (HTE) was crucial in the development of AFE. Active learning that integrated AFE, Feature Selection Process (FPS), and HTE in a loop helped eliminate alternative hypotheses and identified a true hypothesis set that applies to diverse catalysts. This is attributed to the ability of the machine to develop a feature or knowledge space for recognizing composition-performance relationships of catalysts. Incorporating AFE into automated experiments can enable highly efficient autonomous catalyst design. Moreover, the knowledge acquired for a specific system will not only help predict the performance of unknown compositions in the same system but also assist in acquiring knowledge for different systems through transfer learning. As the machine accumulates knowledge across many catalytic systems, it will ultimately develop comprehensive catalytic knowledge and achieve catalyst development freed from researchers' experiences and knowledge.
read more