Regensburg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 23: Poster Monday: Nanostructures 1
O 23.1: Poster
Montag, 5. September 2022, 18:00–20:00, P4
An Atomic Boltzmann Machine capable of Self-Adaption — Brian Kiraly1, Elze J. Knol1, •Werner M.J. van Weerdenburg1, Hilbert J. Kappen2, and Alexander A. Khajetoorians1 — 1Institute for Molecules and Materials, Radboud University Nijmegen, the Netherlands — 2Donders Institute for Neuroscience, Radboud University Nijmegen, the Netherlands
A grand challenge in creating materials with brain-like functionality is understanding multi-well systems. Such multi-well landscapes are linked to energy-based machine learning models, often based on Ising spins. However, the typical short-ranged exchange coupling of Ising spins in real materials prohibit the connectivity required for multi-well landscapes. Therefore, understanding how to create multi-well systems and link these to attractor network models is vital [1].
We present an atomic Boltzmann machine capable of self-adaption using single Co atoms on Black Phosphorus (BP). Using the concept of orbital memory in Co atoms [2], we design a tunable multi-well energy landscape by patterning atoms with atomic manipulation in a scanning tunneling microscope (STM). By electrically gating the structure with the STM tip, we allow the dynamical exploration of its configurations. Due to the anisotropic BP, we find two different timescales that emulate a fast "neural" and a slow "synaptic" timescale. We demonstrate the self-adaption of the synaptic weights to electrical stimuli and explore frequency-based input signals in new types of orbital memory.
[1] Kolmus et al., New J. Phys. 22 (2020);
[2] Kiraly et al, Nat. Comm. 9 (2018)