Discovering unusual structures from exception using big data and machine learning techniques.

Area of Science:

• Materials Science
• Computational Materials Science
• Crystallography

Background:

• Machine learning (ML) is increasingly applied in materials science for predicting material properties and trends.
• Current ML applications often focus on general rules, potentially overlooking significant exceptions.
• Discovering unusual structures from these exceptions can yield novel scientific insights.

Purpose of the Study:

• To demonstrate the discovery of unusual crystal structures by analyzing exceptions in machine learning models.
• To investigate the relationship between atomic and electronic structures using big data from high-throughput calculations.
• To identify novel materials with potential applications, such as in energy storage.

Main Methods:

• Training a machine learning model to predict the relationship between atomic and electronic structures of crystals.
• Analyzing deviations from the ML model's predictions to identify exceptional structures.
• Utilizing a large database of high-throughput calculations for training and analysis.
• Conducting further investigation on identified unusual structures.

Main Results:

• Successfully identified an unusual crystal structure, Silver(III) difluoride (AgO2F), from ML model exceptions.
• The identified AgO2F structure exhibits unique characteristics, including the presence of Ag3+ and O2^2- ions.
• The band gap of AgO2F significantly deviated from the ML model's prediction, highlighting its anomalous nature.

Conclusions:

• Machine learning excels at identifying unusual material structures by examining deviations from predicted trends.
• The discovery of AgO2F offers a new avenue for research into anionic redox mechanisms in transition metal oxides.
• This approach holds promise for accelerating the discovery of novel materials for applications like lithium-ion batteries.