Predicting Synthesizability using Machine Learning on Databases of Existing Inorganic Materials.

Area of Science:

• Materials Science
• Computational Materials Science
• Data-Driven Discovery

Background:

• Predicting the synthesizability of new compounds is a critical challenge in materials science.
• Advancements in machine learning (ML) and computational power offer new approaches to address synthesizability.
• Data-driven research requires efficient methods for identifying experimentally viable materials.

Purpose of the Study:

• To develop a deep learning model for predicting material synthesizability.
• To create a synthesizability score (SC) model for filtering and identifying promising new material candidates.
• To evaluate the model's performance on existing and newly discovered materials.

Main Methods:

• Utilized the Inorganic Crystal Structure Database (ICSD) and Materials Project (MP) databases.
• Employed Fourier-transformed crystal properties (FTCP) representation for crystal structures.
• Developed and trained a deep learning model to predict a synthesizability score (SC).

Main Results:

• The SC prediction model achieved 82.6% precision and 80.6% recall for ternary crystal materials.
• Testing on materials added after 2015 showed an 88.60% true positive rate, indicating high synthesis potential in unexplored materials.
• A list of 100 high-SC materials from the post-2019 dataset was generated for future research.

Conclusions:

• The developed SC model serves as an effective filter for material screening and discovery.
• The model demonstrates high accuracy in predicting synthesizability, particularly for novel materials.
• This approach facilitates efficient identification of new, experimentally realizable compounds.