for Neuromorphic Computing*

The brain-inspired codes of the AI revolution primarily run on conventional silicon computer architectures that were not designed for it. This will lead to unrealistic energy consumption as AI continues to grow. Neuromorphic architectures offer the promise of lower energy consumption by mimicking the basic components of the brain: neurons and synapses. While numerous materials are being considered for mimicking synapses, only a few quantum materials are suitable for replicating neurons. Surprisingly, these "neuristors" often present multiscale fractal electronic patterns, which must be understood to fully control them. To address this, we will explore three key questions related to neuristors:

(i) How inhomogeneous are these materials? We have developed a new optical microscopy method that allows for the precise sub-micron recording of the insulator-to-metal transition in VO₂ neuristors [1]. Using thousands of fully image-stabilized fractal patterns, we have reconstructed and analyzed critical temperature Tc maps and hysteresis width ∆Tc maps.

(ii) What underlying interactions are present? Using image recognition techniques based on a deep learning framework, we have analyzed the entire recorded series of patterns [2]. We find that a two-dimensional Hamiltonian with both nearest neighbor interactions and random field disorder explains the rich spatial structure of the fractal patterns.

(iii) Can critical temperature maps be engineered? Recently, ramp-reversal temperature sweeps have been reported to change the transition temperature in these materials. By optically tracking the fractal patterns, Tc change is shown to occur deep inside both insulating and metallic clusters throughout the entire sample. We find that a theoretical model of the diffusion of point defects for the observed changes and subsequent erasing over the entire sample surface [3]. 

[1] arXiv:2301.04220 (2023)

[2] Phys.Rev.B 107, 205121 (2023)

[3] Adv. Electron. Mater. 2023, 9, 2300085 (2023)