AVS 71 Session AC-ThP: Actinides and Rare Earths Poster Session

The f-electron systems, particularly uranium-based compounds, exhibit unconventional ground states such as coexisting ferromagnetism and superconductivity. UGe₂ was the first uranium compound where this coexistence was discovered, marking a clear departure from conventional BCS theory [1]. Studying such materials in thin-film form offers a pathway to tune quantum correlations and explore emergent behaviors in reduced dimensions. In our study, we have synthesised U-Ge films of different stoichiometries by dc sputtering from a bulk, stoichiometric single crystal in an Ar atmosphere. By varying argon pressure and dc current on the target, we prepared a series of U-Ge thin films with varied stoichiometry. Photoemission spectroscopy studies (XPS and UPS) were performed on freshly prepared surfaces of the U–Ge thin films. The experimental data were compared with available DFT results for UGe₂, which employed the relativistic FPLO method and the FP-LAPW approach (WIEN2k)[2]. The samples were further characterized using XRD, magnetisation and resistivity measurements. Preliminary magnetisation measurements revealed features not observed in bulk. Low-angle XRD data suggests an expanded unit cell volume as compared to bulk.

We thank GACR grant no. 22-19416S, Vakuum Praha for the student Vakuum Praha Grant 2024 and Grant Agency Charles University for GAUK student grant 2025.

References

[1] S. S. Saxena et al. Nature 406, 587 (2000).

[2] M. Samsel-Czekała et.al,Intermetallics 19, October (2011).

Abstract Summary:
Fission Track Analysis (FTA) is a key method in nuclear forensics for detecting fissile materials. This study proposes a novel deep learning approach to automate the segmentation and classification of star-shaped patterns in microscopic images, reducing the need for manual analysis.

Methodology:
Using a U-Net fully convolutional neural network, the research focuses on identifying star-like features in microscopy. A custom simulation tool generated artificial star shapes for training, alongside a new, diverse image database. Models were trained separately for small stars (under 60µm, fewer than 10 branches, no black center) and larger, more complex patterns. An adaptive thresholding method was introduced to improve data labeling and background noise filtering.

Key Findings:
The model reached 92.04% accuracy for small star classification and an ROC AUC of 0.84. For multi-class tasks, it achieved 86.3% accuracy in distinguishing star quality and 82.63% accuracy in recognizing stars with varying numbers of branches. Advanced classification models reached an AUC of 0.90.

Conclusion:
This study shows that deep learning can significantly enhance FTA by automating star pattern detection and classification, offering a more efficient and accurate tool for nuclear forensic analysis.