Portfolio
AI Hardware / Bioinspired Neuromorphic Nanocomposites
AI Hardware · Advanced Materials

A flexible synapse that learns by association at picowatt power.

Solution-processed CNT/PDMS and ZnO/PVP memristors that switch on 0.14 nW and pair electrical and optical stimuli into a stable association — the same trick a brain uses. A 5×5 hardware crossbar wired into a small ANN recovers +14 pp of accuracy lost in the data-sparse regime.

  • CNT/PDMS
  • ZnO/PVP
  • 5×5 crossbar
  • Photo + electric stimuli
  • Keras ANN
Switching power
0.14 nW
Learning efficiency
108 · record
Endurance
104+ cycles
MNIST @ 90 imgs
90% vs 76%

Matter 7(3), 1230–1244 · 2024 Published

Biological associative learning compared with brain-inspired crossbar architecture under combined electrical and photonic stimulation

Why it matters

Edge AI hits an energy wall: a wearable, in-sensor classifier cannot carry a datacenter power budget, and silicon synaptic elements still train slowly (2–16 min), forget within hours, and are rigid — a poor match for smart skin or roll-to-roll integration.

The brain still beats neuromorphic silicon by orders of magnitude in energy and area, in large part because it learns by association across multi-modal stimuli. This work asks whether a solution-processed nanocomposite can do the same — built from chemistry, not lithography.

How it works

  • Two device stacks. Au / CNT–PDMS / ITO on PET for low-power flexible switching; Ag / ZnO–PVP / ITO for photo-active associative learning.
  • Heterostimuli modulation. Electrical pulses drift Ag+ filaments; UV light triggers photo-reduction inside ZnO. Pairing the two writes a persistent, non-volatile associated state.
  • Solution processing only. Tip-sonication, spin coating, thermal evaporation at 10−5–10−7 Torr. No lithography — roll-to-roll compatible.
  • Crossbar hardware. 5×5 array integrated into a Keras ANN to evaluate in-sensor computing in a deliberately data-sparse regime.
  • Reliability envelope. Cyclic I–V to 104+ pulses, 1 ms pulse-measure, 365 nm UV, bending from flat to R = 5 mm with no on/off-current collapse.
Biological associative learning via volume transmission and chemical neuromodulation mapped onto a machine analogue using photonic broadcast and photo-chemical reduction at an Ag/ZnO/ITO junction
Bio to Machine Volume transmission and chemical neuromodulation in a biological synapse are mirrored by broadcast light + photo-chemical reduction at a wired Ag/ZnO/PVP device.
Three operating regimes of the device: volatile ion drift, training under combined light and bias, non-volatile retention; supporting I-V curves, EDS map of Ag/Zn distribution, and photoresponse currents at varying light intensities
Mechanism Volatile ion drift becomes a persistent state once light and bias co-act — Ag/ZnO EDS map shows the filament that locks the association in place.

What we found

0.14 nW Switching power Record-low for flexible organic neuromorphic devices.
92.5% Photonic classification Digits "7" vs "4" on the 5×5 crossbar from photonic stimuli alone.
90% vs 76% MNIST · 90-image regime ALM-integrated ANN closes the gap that normally needs ~1,000 samples.
32.8× Power efficiency gain And 5.64× area gain when one ANN layer is replaced by the crossbar.
Pavlovian conditioning experiment: bell as conditioned stimulus and food pulse as unconditioned stimulus, response trace showing bell-only firing after training, and scatter of energy consumption vs learning efficiency with this work as the outlier
Pavlovian Light (bell) alone triggers a response only after pairing with electrical pulses (food). The energy × efficiency scatter places this device as an outlier vs. prior synapses.
Brain-inspired associative computing on a physical 5x5 crossbar: hardware integration with electrical and photonic inputs, before/after training probability bars showing successful classification, and photonic-response confusion matrix
Crossbar Photonic input alone is classified into the right output neuron after associative learning — the confusion matrix shows clean diagonal recognition across digits 0–9.
Test accuracy as a function of training-set size: ALM-integrated ANN line consistently above the conventional ANN baseline across the 10 to 100 image regime
Sample efficiency Across the entire data-sparse range (10–100 images), the ALM-integrated ANN sits above a conventional ANN baseline — widest gap at the smallest budgets.
Mechanical bending characterization from flat to R=5mm with on/off current stability across multiple cycles, photograph of the flexible device held between gloved fingers, and scatter plot of switching power vs bend radius placing this work at picowatt levels
Flex Bending from flat to R = 5 mm leaves on/off currents intact across cycles; the device sits orders of magnitude below prior flexible memristors on the power–curvature map.

So what. In-sensor pattern recognition on wearable, chemical, and optical platforms without datacenter-scale training. The recipe is roll-to-roll compatible, so it transfers to smart skin, edge medical sensors, and robotic perception with minimal process change.

Paper & resources

Primary article Matter 7(3), 1230–1244 (2024)
DOI · Matter 10.1016/j.matt.2024.01.008
Companion article Adv. Composites & Hybrid Mater. 6, 14 (2023)
DOI · Adv. Composites 10.1007/s42114-022-00599-9
Author Prashant Dhakal
BibTeX · Matter 2024 
@article{kim2024heterostimuli,
  title   = {Heterostimuli chemo-modulation of neuromorphic nanocomposites for time-, power-, and data-efficient machine learning},
  author  = {Kim, Jae Gwang and Liu, Ruochen and Dhakal, Prashant and Hou, Aolin and Gao, Chongjie and Qiu, Jingjing and Merkel, Cory and Zoran, Mark and Wang, Shiren},
  journal = {Matter}, volume = {7}, number = {3}, pages = {1230--1244},
  year    = {2024}, doi = {10.1016/j.matt.2024.01.008}
}
BibTeX · Adv. Composites 2023 
@article{liu2023neuromorphic,
  title   = {Neuromorphic properties of flexible carbon nanotube/polydimethylsiloxane nanocomposites},
  author  = {Liu, Ruochen and Kim, Jae Gwang and Dhakal, Prashant and Li, Wei and Ma, Jun and Hou, Aolin and Merkel, Cory and Qiu, Jingjing and Zoran, Mark and Wang, Shiren},
  journal = {Advanced Composites and Hybrid Materials},
  volume  = {6}, pages = {14}, year = {2023},
  doi     = {10.1007/s42114-022-00599-9}
}