Property Prediction for Complex Compounds Using Structure-Free Mendeleev Encoding and Machine Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting material properties from chemical formulas alone offers significant advantages for research and resource allocation.
• Structure-free encoding and machine learning (ML) methods can enable such predictions, but require careful consideration of processing decisions.

Purpose of the Study:

• To compare various structure-free material encodings and ML algorithms for predicting battery material properties.
• To investigate the impact of physical units of property labels on predictive model performance.

Main Methods:

• Evaluation of diverse structure-free encoding techniques for material representation.
• Application of multiple machine learning algorithms to predict material properties.
• Analysis of the influence of property label units (e.g., per weight vs. per volume) on prediction accuracy.

Main Results:

• The choice of physical units for property labels critically affects model predictive ability, irrespective of the computational approach.
• Property labels normalized by weight demonstrated excellent predictive performance.
• Property labels normalized by volume could not be reliably predicted using only chemical information, even for materials with similar physical characteristics.

Conclusions:

• The representation of material structural features and property labels is crucial for the success of machine learning models in materials science.
• While structure-free encoding is effective in some ML tasks, its performance in property prediction is highly sensitive to the chosen units of measurement.