Automation of Atom-Scale Device Patterning using Machine Learning

In recent years, research involving dangling bonds (DBs) patterned on hydrogen-terminated silicon (H-Si) has reached several significant benchmarks, including their applications as logic gates, binary wires¹ and rewritable memory². These newly developed devices show promise for the implementation of atom scale DB circuitry. As device applications become realizable, automation of device fabrication is necessary to facilitate the transition to commercial applications. We show that by incorporating a deep neural network in our patterning process, we can greatly reduce the amount of active user time needed for device fabrication. Semantic segmentation is used with a convolution-deconvolution network to properly map and label surface defects. By combining this neural network with libraries from a commercial scanning probe microscope controller and a previously implemented probe tip conditioning suite³, complete automation of the patterning process is realized.

1. Huff, T. et al. arXiv:1706.07427 (2017).
2. Achal, R. et al. Nat. Commun. 9, 2778 (2018).
3. Rashidi, M. et al. ACS Nano 12, 56 (2018).