Thumbnail of this research highlight
Project Summary: Transition metal dichalcogenides (TMDs) samples are often characterized using atomic force microscopy (AFM), which generates large amounts of historical data. To effectively use this data, models for automated detection and classification of TMD samples are needed. These trained models can then help create simplified, low-dimensional representations of the data that are more accessible. In this work, transfer learning with convolutional neural networks (CNNs) is used to classify five TMDs: MoS2, WS2, WSe2, MoSe2, and Mo-WSe2. The model achieves high classification accuracy and enables us representation of AFM data in just two dimensions using principal component analysis. When visualized, these two-dimensional representations show clear boundaries between different TMD classes. Two transfer learning approaches are evaluated: sequential learning, where to more TMD classes are gradually added by fine-tuning models trained on fewer classes; and ImageNet, where the original ImageNet weights are fine-tuned each time new TMD class are added. It is shown that transfer learning can classify TMDs more accurately than human experts and provide simplified representations of AFM images that are useful for interpretation and further analysis.
Publication: Moses, I. A., & Reinhart, W. F. (2025). Transfer learning for multi-material classification of transition metal dichalcogenides with atomic force microscopy.Machine Learning: Science and Technology,5(4), 045081.
Dataset; Instrument
2DCC Role:The 2DCC MOCVD thin films facility was used to grow TMDs that were characterized and data science personnel contributed transfer learning accessing the growth and characterization data through the Lifetime Sample Tracking (LiST) tool.