Benchmark datasets incorporating diverse tasks, sample sizes, material systems, and data heterogeneity for materials...

Area of Science:

• Materials Science
• Data Science
• Computational Chemistry

Background:

• Machine learning accelerates materials discovery by predicting properties efficiently.
• A significant limitation is the absence of comprehensive benchmark datasets for evaluating machine learning models.
• This hinders the selection of optimal algorithms, architectures, and data processing methods.

Purpose of the Study:

• To address the lack of benchmark datasets in materials informatics.
• To create a diverse repository of 50 datasets for materials properties.
• To facilitate the comparison and improvement of machine learning models in materials science.

Main Methods:

• Assembled a repository of 50 diverse materials property datasets from 16 publications.
• Included both experimental and computational data for regression and classification tasks.
• Implemented standardized train-validation-test splits using cross-validation (5-fold, 10-fold, or Leave-One-Out) based on dataset size.

Main Results:

• Created a unique collection of 50 datasets covering various material systems and data modalities.
• Datasets range in size from 12 to 6354 samples, accommodating diverse machine learning applications.
• Standardized data splitting protocols were applied to ensure consistent evaluation.

Conclusions:

• The curated benchmark datasets provide a foundation for robust machine learning model evaluation.
• This resource will aid researchers in identifying optimal machine learning strategies for materials discovery.
• Further development of diverse benchmark datasets is crucial for advancing the field of materials informatics.