How big is big data?

Area of Science:

• Materials Science
• Computer Science
• Data Science

Background:

• Machine learning models are increasingly used for predictive tasks in materials science.
• The definition and implications of 'big data' in this domain require careful examination.

Purpose of the Study:

• To define 'big data' in the context of materials science machine learning.
• To investigate challenges related to data volume, quality, veracity, and infrastructure.
• To explore model generalization, data aggregation, feature engineering, and computational requirements.

Main Methods:

• Analysis of typical materials science machine learning problems.
• Evaluation of model generalization across datasets.
• Case studies on gathering high-quality data from heterogeneous sources.
• Assessment of feature set and model complexity impact on expressivity.
• Examination of infrastructure needs for large-scale data and model training.

Main Results:

• 'Big data' in materials science involves complex interplay of data volume, quality, and veracity.
• Model generalization varies significantly with dataset characteristics.
• Effective aggregation of heterogeneous data sources is challenging but feasible.
• Feature engineering and model complexity are critical for predictive accuracy.
• Substantial infrastructure is required for managing and training on large materials datasets.

Conclusions:

• Big data in materials science machine learning poses multifaceted challenges.
• Further research is needed to address data quality, infrastructure, and model development.
• Optimizing data handling and model training is essential for unlocking predictive potential.