Melting temperature prediction using a graph neural network model: From ancient minerals to new materials

Area of Science:

• Materials Science
• Computational Chemistry
• Planetary Science

Background:

• Melting point determination is crucial for materials science and geology but is often time-consuming.
• High-throughput analysis of melting relations and phase diagrams is hindered by slow measurement and computation.
• Predicting melting temperatures is essential for discovering new materials and understanding planetary composition.

Purpose of the Study:

• To develop a rapid and accurate machine learning model for predicting material melting temperatures.
• To enable high-throughput screening of candidate compounds for materials design.
• To analyze large datasets of minerals for applications in planetary science and geology.

Main Methods:

• A machine learning model was trained on a database of approximately 10,000 compounds.
• The model utilizes graph neural network and residual neural network architectures.
• The predictive model was made publicly available online for broad accessibility.

Main Results:

• The machine learning model predicts melting temperatures in a fraction of a second.
• Novel multicomponent materials with high melting points were discovered using the model.
• Predictions were validated using density functional theory calculations and experimental methods.
• Analysis of ~4,800 minerals revealed correlations relevant to mineral evolution.

Conclusions:

• The developed machine learning model significantly accelerates the prediction of melting temperatures.
• The model is a valuable tool for materials design, discovery, and planetary science research.
• The public availability of the model promotes wider scientific application and collaboration.