Science-Driven Atomistic Machine Learning.

Area of Science:

• Computational chemistry and materials science
• Application of artificial intelligence in scientific research

Background:

• Machine learning (ML) is increasingly vital across scientific disciplines.
• Traditional ML relies heavily on extensive datasets, which are often scarce in chemistry.
• This limitation hinders the application of ML in atomistic modeling of materials and molecules.

Purpose of the Study:

• To review science-driven ML approaches suitable for chemistry, particularly when "big data" is unavailable.
• To highlight methods that integrate scientific inquiry with ML model development.
• To emphasize efficient data utilization and robust evaluation in ML for chemical applications.

Main Methods:

• Focus on science-driven ML, where scientific questions guide data collection and model design.
• Discussion of automated and purpose-driven data acquisition strategies.
• Integration of chemical and physical "priors" (existing knowledge) to enhance data efficiency.

Main Results:

• Science-driven ML enables effective atomistic modeling without large databases.
• Key features include tailored data collection and the incorporation of scientific knowledge.
• Emphasis on rigorous model evaluation and error estimation is crucial for reliable results.

Conclusions:

• Science-driven ML provides a viable and efficient alternative to data-intensive ML in chemistry.
• This paradigm shift allows for powerful applications of ML even with limited data.
• Future research should focus on developing and validating these science-informed ML techniques.