Predicting synthesis recipes of inorganic crystal materials using elementwise template formulation.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Virtual materials design is advancing rapidly, but experimental synthesis remains a bottleneck.
• Existing methods for predicting inorganic synthesis routes lack prioritization and can suggest unrealistic precursors.

Purpose of the Study:

• To develop a novel computational model for predicting inorganic synthesis recipes.
• To improve the accuracy and reliability of predicted synthesis pathways.
• To provide a confidence measure for prioritizing synthesis predictions.

Main Methods:

• An element-wise graph neural network (GNN) was developed to predict inorganic synthesis recipes.
• The GNN model was trained on a comprehensive dataset of materials synthesis data.
• Model performance was evaluated using top-k exact match accuracy and publication-year-split tests.

Main Results:

• The GNN model significantly outperformed a baseline statistical model in predicting synthesis recipes.
• The model demonstrated predictive capability for materials synthesized after its training data cutoff.
• A high correlation was observed between the model's probability score and prediction accuracy.

Conclusions:

• The proposed element-wise GNN is a valid and effective approach for predicting inorganic solid-state synthesis.
• The probability score from the model can serve as a reliable indicator of prediction confidence.
• This work accelerates the discovery and synthesis of novel inorganic materials.