Structure-Based Synthesizability Prediction of Crystals Using Partially Supervised Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting inorganic material synthesizability is crucial for accelerated materials discovery.
• Thermodynamic decomposition stability is a common but limited predictor, often yielding too many candidates or missing metastable materials.
• Material synthesizability is a complex phenomenon influenced by factors beyond thermodynamic stability.

Purpose of the Study:

• To develop a machine learning model for quantifying the probability of inorganic material synthesis.
• To improve the accuracy of predicting synthesizability compared to traditional thermodynamic methods.
• To provide a data-driven metric for rational materials design and reduce experimental exploration space.

Main Methods:

• Implemented a partially supervised learning approach using positive and unlabeled (PU) learning.
• Utilized a graph convolutional neural network as a classifier to generate crystal-likeness scores (CLscore).
• Trained and validated the model on a large database of experimentally reported materials, including recent discoveries.

Main Results:

• The model achieved 87.4% true positive prediction accuracy on a test set of 9356 materials from the Materials Project.
• Validation on newly reported materials (2015-2019) showed an 86.2% true positive rate.
• The CLscore captures structural motifs for synthesizability beyond the capabilities of E_hull, with 71% of top-scoring virtual materials being previously synthesized.

Conclusions:

• The developed machine learning model (CLscore) effectively quantifies material synthesizability.
• This data-driven metric significantly enhances high-throughput virtual screening and generative models for materials discovery.
• The CLscore facilitates more rational materials design by reducing the experimental search space.