Computational Material Science Has a Data Problem.

Area of Science:

• Computational Materials Science
• Data Science
• Machine Learning

Background:

• Machine learning models are increasingly used in materials science.
• The reliability of training data is crucial for accurate predictions.
• Existing materials databases may contain inherent limitations.

Purpose of the Study:

• To critically evaluate the suitability of computational materials science data for machine learning.
• To identify potential issues within commonly used materials datasets.

Main Methods:

• Analysis of energy above the convex hull (Eh).
• Examination of electronic bandgap data.
• Assessment of formation energy values.
• Investigation of data consistency and representation within the Materials Project database.

Main Results:

• Energy above convex hull (Eh) and DFT-computed voltages are unsteady due to insufficient chemical space representation for crystal decomposition.
• Discrepancies exist in reported electronic bandgap values.
• Formation energy data can be unstable based on optimization parameter variations.

Conclusions:

• Current computational materials science datasets may not be fully suitable for training robust machine learning models.
• Further data curation and validation are necessary for reliable materials discovery.
• The identified inconsistencies highlight the need for improved data generation and quality control.