Artificial intelligence and machine learning algorithms provide a path to train models that can guide experiments in scanning probe microscopy to the performance level of an experienced scientist or engineer, for certain tasks. We collect hyperspectral data autonomously via Gaussian process regression, where one-dimensional convolutional neural networks (CNN) are used in tandem for defect identification and subsequent feedback.[1] We explore monolayer transition metal dichalcogenide (TMD) sulfur vacancies within tungsten disulfide (WS₂), to provide hyperspectral insight into available sulfur-substitution sites within a TMD that is combined with spectral verification on the Au{111} herringbone reconstruction. Here, scanning probe microscopy methods that take advantage of dense tunneling bias spectroscopy acquisition can both image and identify the atomic and electronic landscape without any user-driven bias. Additionally, we delve into building a two-dimensional CNN for image recognition over Au{111} substrates.

[1] Thomas, J.C., Rossi, A., Smalley, D. et al. Autonomous scanning probe microscopy investigations over WS₂ and Au{111}. npj Comput Mater 8, 99 (2022).

Schematic depicting the control of a scanning tunneling microscope with artificial intelligence and machine learning algorithms. The open-access library, gpSTS, enables optimized hyperspectral data collection and a path for autonomous discovery that can be implemented over the large library of available quantum materials.

Eli Rotenberg

Appearances