High entropy alloy property predictions using a transformer-based language model.

Area of Science:

• Materials Science
• Computational Materials Science
• Materials Informatics

Background:

• High-entropy alloys (HEAs) present unique challenges for property prediction due to their complex compositions and limited experimental data.
• Traditional machine learning models often struggle with the intricate elemental interactions inherent in HEAs.
• Developing accurate predictive models is crucial for accelerating the discovery and optimization of novel HEA materials.

Purpose of the Study:

• To introduce a language transformer-based machine learning model for predicting key mechanical properties of high-entropy alloys.
• To address the challenges of data scarcity and complex elemental interactions in HEAs.
• To enhance the accuracy and interpretability of property predictions for materials informatics.

Main Methods:

• A language transformer model was pre-trained on synthetic materials data and fine-tuned on specific HEA datasets.
• Self-attention mechanisms within the transformer were utilized to capture intricate elemental interactions.
• Transfer learning was employed to mitigate issues arising from limited experimental data.
• Attention weights were visualized to enhance model interpretability.

Main Results:

• The transformer model demonstrated enhanced predictive accuracy for properties such as elongation and ultimate tensile strength compared to random forests and Gaussian processes.
• The model effectively captured complex elemental interactions, aligning with established metallurgical principles.
• Visualization of attention weights provided insights into significant elemental relationships.

Conclusions:

• Transformer models show significant potential for accelerating materials discovery and optimization in the field of materials informatics.
• The developed model enables accurate prediction of mechanical properties, overcoming limitations of traditional methods.
• Future work should focus on advanced preprocessing, realistic synthetic data generation, and refined tokenization for practical applications.