Highly efficient computer algorithm for identifying layer thickness of atomically thin 2D materials

다국어 초록 (Multilingual Abstract)

<P>The fields of layered material research, such as transition-metal dichalcogenides (TMDs), have demonstrated that the optical, electrical and mechanical properties strongly depend on the layer number <I>N</I>. Thus, efficient and accurate determination of <I>N</I> is the most crucial step before the associated device fabrication. An existing experimental technique using an optical microscope is the most widely used one to identify <I>N</I>. However, a critical drawback of this approach is that it relies on extensive laboratory experiences to estimate <I>N</I>; it requires a very time-consuming image-searching task assisted by human eyes and secondary measurements such as atomic force microscopy and Raman spectroscopy, which are necessary to ensure <I>N</I>. In this work, we introduce a computer algorithm based on the image analysis of a quantized optical contrast. We show that our algorithm can apply to a wide variety of layered materials, including graphene, MoS<SUB>2</SUB>, and WS<SUB>2</SUB> regardless of substrates. The algorithm largely consists of two parts. First, it sets up an appropriate boundary between target flakes and substrate. Second, to compute <I>N</I>, it automatically calculates the optical contrast using an adaptive RGB estimation process between each target, which results in a matrix with different integer <I>N</I>s and returns a matrix map of <I>N</I>s onto the target flake position. Using a conventional desktop computational power, the time taken to display the final <I>N</I> matrix was 1.8 s on average for the image size of 1280 pixels by 960 pixels and obtained a high accuracy of 90% (six estimation errors among 62 samples) when compared to the other methods. To show the effectiveness of our algorithm, we also apply it to TMD flakes transferred on optically transparent <I>c</I>-axis sapphire substrates and obtain a similar result of the accuracy of 94% (two estimation errors among 34 samples).</P>