Practical feature filter strategy to machine learning for small datasets in chemistry.

Area of Science:

• Chemistry
• Materials Science
• Machine Learning

Background:

• Machine learning applications in chemistry and materials science often face challenges due to small dataset sizes.
• Effective model design and feature selection are crucial for reliable predictions in data-limited scenarios.

Purpose of the Study:

• To propose a practical and efficient feature filter strategy for selecting optimal input features.
• To demonstrate the strategy's effectiveness in predicting adsorption energies and sublimation enthalpies.

Main Methods:

• A feature filter strategy was developed and applied to identify relevant input features.
• The strategy was tested on predicting adsorption energies using a public dataset and sublimation enthalpies using an in-house dataset.
• Extreme gradient boosting regression models were employed and evaluated.

Main Results:

• Feature selection reduced the input dimensions for adsorption energy prediction from 12 to two, maintaining accuracy.
• Three optimal input configurations were identified for predicting sublimation enthalpies from 14 possibilities.
• The best machine learning models achieved accuracy comparable to density functional theory computations with physical interpretability.

Conclusions:

• The proposed feature filter strategy facilitates the creation of reliable, small training datasets for machine learning.
• This approach aids interdisciplinary scientists with limited AI expertise or computational resources.
• It simplifies and enhances the accuracy of machine learning model training, reducing time and improving feature selection.