Automatic feature engineering for catalyst design using small data without prior knowledge of target catalysis.

Area of Science:

• Catalysis
• Materials Informatics
• Machine Learning

Background:

• Descriptor design in catalyst informatics often requires prior knowledge, posing a challenge for limited datasets.
• Developing new catalysts typically involves extensive empirical testing and hypothesis generation.

Purpose of the Study:

• To introduce a novel technique for automatic feature engineering (AFE) applicable to small catalyst datasets.
• To enable machine learning model development for catalysis without relying on specific assumptions or pre-existing knowledge.

Main Methods:

• AFE generates numerous features via mathematical operations on physicochemical properties of catalytic components.
• Relevant features are extracted for specific catalytic applications, screening hypotheses computationally.
• Active learning combined with AFE and high-throughput experimentation was applied to oxidative coupling of methane (OCM).

Main Results:

• AFE demonstrated reasonable regression performance across three heterogeneous catalytic systems: OCM, ethanol to butadiene conversion, and three-way catalysis.
• The technique proved effective even when only the training dataset was interchanged.
• Active learning successfully visualized the machine's learning process in catalyst design for OCM.

Conclusions:

• AFE is a versatile and powerful technique for data-driven catalysis research.
• This method significantly advances the pursuit of fully automated catalyst discovery.