Introducing pre-trained transformers for high entropy alloy informatics

Spyros Kamnis
DOI: https://doi.org/10.1016/j.matlet.2024.135871
IF: 3
2024-01-04
Materials Letters
Abstract:High-Entropy Alloys (HEAs) offer a wealth of compositional possibilities, presenting both immense opportunities and challenges in identifying the best candidates for specific applications. This study introduces a new informatics pipeline for HEAs, utilizing Large Language Model architectures and pre-trained transformers inspired by Natural Language Processing techniques. This work demonstrates that transformers pre-trained on theoretical unlabelled thermodynamic data can be fine-tuned with labelled experimental HEA data to significantly improve predictive accuracy while negating the need for extensive experimentally driven labelled datasets.
materials science, multidisciplinary,physics, applied
What problem does this paper attempt to address?