Presentation Information
[16p-PB2-10]Optimizing Machine Learning Loss Functions for SPM Nanostructure Identification
〇Soichiro Kondo1, Akihiro Tuji2, Hayato Yamashita2, Eiichi Inami1 (1.Kouchikou Univ., 2.Osaka Univ)
Keywords:
Machine Learning,Scanning Tunneling Microscopy,Atomic Force Microscope
Machine learning (ML)-based recognition of nanostructures in scanning probe microscopy (SPM) images is promising for big-data analysis, but reliable identification remains challenging due to noise and image-specific variations. In this study, we improve performance by introducing a custom loss function tailored to nanostructure analysis in SPM images.
The proposed loss function incorporates structural characteristics of target nanostructures, enabling more robust learning from noisy experimental data. Model performance was evaluated using SPM images of Pb clusters and photozipper proteins. Compared with standard implementations, the optimized model shows a clear reduction in detection errors and a significant improvement in identification accuracy. We will present detailed results for both datasets and discuss the impact of loss-function design on SPM image analysis.
The proposed loss function incorporates structural characteristics of target nanostructures, enabling more robust learning from noisy experimental data. Model performance was evaluated using SPM images of Pb clusters and photozipper proteins. Compared with standard implementations, the optimized model shows a clear reduction in detection errors and a significant improvement in identification accuracy. We will present detailed results for both datasets and discuss the impact of loss-function design on SPM image analysis.