BACK_TO_FEEDAICRIER_2
Bi2Se3 memristor powers analog neural network
OPEN_SOURCE ↗
REDDIT · REDDIT// 17d agoRESEARCH PAPER

Bi2Se3 memristor powers analog neural network

University of Michigan researchers published an ACS Nano paper on a Bi2Se3 memristor built from 2D bismuth selenide layers that combines long-term retention, analog tuning, and regulator-free operation. In a fully analog reservoir-computing demo, the device controlled a balance lever using about 7 microwatts, pointing to lower-power neuromorphic hardware.

// ANALYSIS

This is the rare memristor paper that feels like a systems advance, not just a materials curiosity. The useful part is that it closes a real feedback-loop problem without forcing the circuit back through digital conversion.

  • Gold filaments grow and retract without bridging the gap, giving smooth analog modulation instead of abrupt switching.
  • The paper reports 10-40% conductance tuning, less than 1% retention loss over 10,000 seconds, about 7 microwatts controller power, and NRMSE 0.094.
  • The gold-assisted, plasma-etching-free PVD crossbar is the commercialization clue: it looks much closer to a manufacturable process than a one-off lab artifact.
  • For developers, the promise is edge control loops where sensing, memory, and actuation stay analog, though array-scale uniformity and endurance still need proof.
// TAGS
bi2se3-memristoredge-aiinferencerobotics

DISCOVERED

17d ago

2026-03-26

PUBLISHED

17d ago

2026-03-25

RELEVANCE

8/ 10

AUTHOR

jferments