Benchmarking the acceleration of materials discovery by sequential learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Sequential learning (SL) iteratively updates models to guide experiments, promising accelerated materials discovery.
• Previous applications on computational data and limited experiments show SL's potential.

Purpose of the Study:

• Quantitatively evaluate SL's ability to accelerate physical materials discovery.
• Benchmark SL algorithms against diverse research goals: finding any "good" material, all "good" materials, or accurate predictive models.

Main Methods:

• Tested diverse SL schemes on four pseudo-quaternary metal oxide chemical spaces (2121 catalysts each).
• Used high-throughput synthesis and electrochemistry data for oxygen evolution reaction (OER) overpotential.
• Compared SL performance against random acquisition strategies.

Main Results:

• SL can accelerate research by up to 20x in specific scenarios.
• Certain SL models significantly underperform random acquisition for particular research goals.
• Performance varies greatly depending on the SL strategy and research objective.

Conclusions:

• Provides quantitative guidance for tuning SL strategies based on research goals.
• Highlights the need for materials-aware SL algorithms to optimize discovery acceleration.
• Demonstrates the critical importance of selecting appropriate SL models for efficient materials research.