Determining usefulness of machine learning in materials discovery using simulated research landscapes

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Data acquisition bias significantly impacts machine learning (ML) model performance in materials science.
• High costs often limit the creation of new, unbiased datasets, while material properties depend on a few unknown parameters.

Purpose of the Study:

• To introduce simulated research landscapes for evaluating ML usefulness under common data acquisition constraints.
• To quantitatively compare standard materials exploration with ML-guided strategies.

Main Methods:

• Development and application of simulated research landscapes to model dataset evolution.
• Quantitative comparison of ML-guided versus standard materials discovery strategies.
• Analysis of factors influencing ML prediction capabilities, including data dimensionality and collection strategy.

Main Results:

• A specific window of opportunity exists for significant ML benefit; adopting ML too early or too late can diminish its advantages or even hinder discovery.
• ML-guided strategies can outperform standard exploration, but the benefit is contingent on specific conditions.
• Data dimensionality, landscape corrugation, and data collection strategy critically affect ML prediction accuracy and discovery acceleration.

Conclusions:

• Simulated research landscapes provide a framework for understanding and optimizing ML adoption in materials discovery.
• A qualitative guide is offered for identifying conditions under which ML can effectively accelerate the discovery of new materials.
• Strategic timing and careful consideration of data characteristics are crucial for maximizing the benefits of ML in scientific research.