An atomic Boltzmann machine capable of self-adaption.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Computational Neuroscience

Background:

• Current machine learning hardware implementations often combine diverse materials, leading to limitations in functionality, efficiency, and scalability.
• Developing integrated systems that directly link physical phenomena to machine learning models is crucial for advancing on-chip learning capabilities.

Purpose of the Study:

• To engineer an atomic spin system capable of directly emulating a Boltzmann machine using the orbital dynamics of a single material.
• To explore the potential of atomic-scale systems for creating efficient and scalable machine learning hardware.

Main Methods:

• Fabrication of a tunable multi-well energy landscape using individual cobalt atoms on black phosphorus, controlled via scanning tunneling microscopy.
• Utilizing the concept of orbital memory and the anisotropic properties of black phosphorus to create tuneable, multi-valued, and interlinking synapses.

Main Results:

• Demonstration of an atomic spin system emulating a Boltzmann machine directly within the orbital dynamics of cobalt atoms on black phosphorus.
• Realization of synaptic plasticity with tuneable probability distributions due to the anisotropic behavior of the material.
• Observation of autonomous reorganization of synaptic weights in response to external electrical stimuli at a distinct timescale from neural dynamics.

Conclusions:

• The developed atomic spin system offers a novel approach for implementing machine learning directly in hardware at the atomic scale.
• This self-adaptive architecture paves the way for autonomous learning capabilities in future atomic-scale machine learning devices.
• The findings highlight the potential of exploiting unique material properties for advanced computational functionalities.