Graph-Based Deep Learning Models for Thermodynamic Property Prediction: The Interplay between Target Definition, Data...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate prediction of thermodynamic properties is crucial for materials discovery and design.
• Graph-based deep learning models offer a promising approach for predicting these properties.

Purpose of the Study:

• To investigate the impact of target definition, data distribution, featurization, and model architectures on graph-based deep learning for thermodynamic property prediction.
• To identify key factors influencing model accuracy and robustness.

Main Methods:

• Evaluation of five diverse datasets with varying elemental composition, multiplicity, charge state, and size.
• Analysis of different target definitions (formation vs. atomization energy/enthalpy).
• Comparison of various featurization approaches and model architectures.

Main Results:

• Target definition (formation energy) and featurization approach are critical for model accuracy.
• Modest accuracy gains were observed through direct modification of model architectures.
• Molecule-level predictions outperformed atom-level increment predictions, contrary to prior findings.

Conclusions:

• Development of robust graph-based thermodynamic models requires careful consideration of target definition and featurization.
• The findings suggest a path towards more universal graph-based models with enhanced accuracy across diverse datasets and compound domains.