Composition Based Oxidation State Prediction of Materials Using Deep Learning Language Models.

Area of Science:

• Computational materials science
• Artificial intelligence in chemistry
• Predictive modeling for inorganic compounds

Background:

• Oxidation states (OS) are crucial for understanding chemical properties and material behavior.
• Current methods for predicting OS rely on heuristic rules with many exceptions, limiting accuracy.
• Existing machine learning models primarily focus on metal ions and require structural information.

Purpose of the Study:

• To develop a novel deep learning model for predicting oxidation states of all elements in inorganic compounds.
• To enable composition-based oxidation state prediction, eliminating the need for determined crystal structures.
• To facilitate large-scale screening of hypothetical materials for accelerated materials discovery.

Main Methods:

• Implementation of a novel deep learning-based BERT transformer language model, named BERTOS.
• Training and benchmarking the BERTOS model on the cleaned Inorganic Crystal Structure Database (ICSD).
• Evaluating the model's accuracy for all-element oxidation state prediction and specifically for oxide materials.

Main Results:

• BERTOS achieved 96.82% accuracy for all-element oxidation states prediction on the ICSD dataset.
• The model demonstrated high accuracy (97.61%) for predicting oxidation states in oxide materials.
• The study showcased BERTOS's utility in screening hypothetical material compositions.

Conclusions:

• The BERTOS model offers a highly accurate, composition-based approach for predicting oxidation states in inorganic compounds.
• This method overcomes limitations of heuristic rules and structure-dependent models.
• BERTOS significantly advances the potential for discovering novel materials by enabling rapid computational screening.