Introducing Materials Fingerprint (MatPrint): A novel method in graphical material representation and features compression

This research encompasses a comprehensive exploration of feature compression and graphical representation in the domain of single crystal materials. The study introduces a novel framework known as Material Fingerprint (MatPrint), leveraging crystal structure and composition features generated via the Magpie platform. MatPrint incorporates 576 crystal and composition features, transformed into 64-bit binary values through the IEEE-754 standard. These features contribute to a nuanced binary graphical representation of materials, emphasizing sensitivity to both composition and crystal structure, particularly beneficial in distinguishing unique graphical profiles for each material, including polymorphs. Additionally, the current MatPrint representations of 2021 compounds and their formation energy were used in a learning process using a pretrained ResNet-18 model to establish a baseline for the efficiency of the representation in data-driven tasks regarding material property prediction, the employed model exhibited a validation loss of 0.18 eV/atom which proposes that the current model can be used extensively with a larger dataset that can be used in different areas of material informatics. Finally, the proposed methodology plays a crucial role in the reversible compression of tabular data derived from the feature generation process, facilitating its use in diverse machine and deep learning models.