Bonn 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

A: Fachverband Atomphysik

A 21: Poster – Atomic Systems in External Fields

A 21.5: Poster

Mittwoch, 12. März 2025, 17:00–19:00, Tent

Leveraging of self-supervised machine learning over supervised machine learning for crystalline materials properties prediction. — •Moses Adasariya — Kwame Nkrumah University of Science and Technology (KNUST), Kumasi, Ghana

The accurate prediction of material properties is essential for the progress of materials science. However, the limited availability of labeled datasets presents a considerable obstacle. This research investigates the capabilities of self-supervised learning (SSL) models to overcome this challenge by utilizing the bulk unlabeled data available for predicting the properties of crystalline materials. Three SSL models were assessed alongside four different supervised learning (SL) model their ability to predict bandgap, formation energy, bulk modulus, and shear modulus. The findings revealed that SSL models consistently outperformed or equaled the performance of SL models across all evaluated tasks. CrysAtom was identified as the most effective model, achieving improvement percentage of 15.1% over orbital graph convolutional neural network (OGCNN) for bandgap, and 9.7% for formation energy over OGCNN. The other SSL models, CT-Barlow and CT-SimSiam, also demonstrated competitive results, particularly in the predictions of bandgap and formation energy. These results underscore the potential of SSL models to diminish dependence on labeled datasets while preserving high levels of prediction accuracy

Keywords: Self-supervised learning (SSL); Supervised learning (SL); Crystalline materials; Bandgap; Formation energy