Classification-Based Detection and Quantification of Cross-Domain Data Bias in Materials Discovery

Area of Science:

• Materials Science
• Data Science
• Artificial Intelligence

Background:

• Accurate artificial intelligence (AI)-driven models rely heavily on data quantity and quality.
• Bias in sample selection during database generation can lead to unreliable predictions when AI models are applied to data with different origins.
• This bias can hinder AI-based materials discovery, even with large, high-quality datasets.

Purpose of the Study:

• To develop a method for detecting and quantifying data bias in AI models.
• To reduce the impact of data bias on materials discovery.
• To enhance the reliability of AI models for diverse and unseen materials.

Main Methods:

• A classification strategy is employed to identify and quantify data bias.
• The method identifies 'out-of-the-box' materials where pretrained model predictions are likely unreliable.
• The approach is validated using superconductor and thermoelectric materials.

Main Results:

• The proposed method effectively detects and quantifies data bias.
• It successfully identifies materials outside the training data distribution for which predictions are unreliable.
• The methodology enhances the reliability of AI models for materials discovery.

Conclusions:

• The developed method offers a simple, flexible, and adaptable solution for bias detection in AI.
• It improves the trustworthiness of AI models applied to new materials, advancing AI-driven materials discovery.
• This approach is compatible with various AI architectures, including graph equivariant neural networks.