Developing Machine Learning in Atomic Ensembles Using Orbital Memory
Brian Kiraly
When adsorbed onto the surface of black phosphorus (BP), a single cobalt atom can take two distinct orbital configurations or valencies [1]. These bistable valencies, arising due to differences in the screening of the cobalt’s 3d orbitals, are separated by a large energy barrier and electrically addressable using the tip of a scanning tunneling microscope (STM). This makes the cobalt on BP system the first to exhibit so-called orbital memory, where information can be stored within the atomic valency. This new degree of freedom can also be coupled through the underlying BP substrate to realize an analog of the Ising spin system. Furthermore, because of the intrinsic electronic anisotropy of the substrate [2], the interactions of such a model system can be non-locally tuned to realize a reconfigurable stochastic system equivalent to the machine learning construct of a Boltzmann machine [3]. Finally, I will illustrate how this Boltzmann machine, an assembly of just seven cobalt atoms, autonomously adapts to environmental stimuli, mimicking the learning processes of the human brain.
[1] Kiraly, Rudenko, Weerdenburg, Wegner, Katsnelson, Khajetoorians, Nature Commun. 9, 3904, (2018).
[2] Kiraly, Knol, Volckaert, Biswas, Rudenko, et. al., Phys. Rev. Lett. 123, 216403 (2019).
[2] Kiraly, Knol, Weerdenburg, Kappen, Khajetoorians, Nature Nanotech. 16, 414, (2021).