MD
M.P. Dubbelman
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Deep learning is a powerful tool that can be used to automate various tasks. Here, deep learning is used in the field of quantum nanoscience to predict how thick 2D material flakes are, to locate them and to create high resolution masks. Mask R-CNN, a deep learning computer vision model, trained on Graphene, MoS2, WTe2 and hBN, is fine-tuned on NbSe2 using transfer learning. To accomplish this, a dataset of optical images and AFM measurements is created. All data is gathered in a glove box at room temperature. This dataset is expanded by an image splitting algorithm. All four starting models produce an approximately equal performing NbSe2 model. With the hBN model being the best starting model. Which was not expected based on the optical properties of the materials. This might be caused by some unknown similarities between the datasets. For large flakes (> 96px2) the model performs sufficiently well for implementation in automated flake searching (AP@IoU50%:95% of 0.54). The model is strongly biased towards one class due to the unbalanced dataset. With a balanced dataset, this project enables researchers to build an automated flake searching setup.
...
Deep learning is a powerful tool that can be used to automate various tasks. Here, deep learning is used in the field of quantum nanoscience to predict how thick 2D material flakes are, to locate them and to create high resolution masks. Mask R-CNN, a deep learning computer vision model, trained on Graphene, MoS2, WTe2 and hBN, is fine-tuned on NbSe2 using transfer learning. To accomplish this, a dataset of optical images and AFM measurements is created. All data is gathered in a glove box at room temperature. This dataset is expanded by an image splitting algorithm. All four starting models produce an approximately equal performing NbSe2 model. With the hBN model being the best starting model. Which was not expected based on the optical properties of the materials. This might be caused by some unknown similarities between the datasets. For large flakes (> 96px2) the model performs sufficiently well for implementation in automated flake searching (AP@IoU50%:95% of 0.54). The model is strongly biased towards one class due to the unbalanced dataset. With a balanced dataset, this project enables researchers to build an automated flake searching setup.