A Multimodal Humidity Adaptive Optical Neuron Based on a MoWS2/VOx Heterojunction for Vision and Respiratory Functions

Abstract

Advancements in computing have progressed from near-sensor to in-sensor computing, culminating in the development of multimodal in-memory computing, which enables faster, energy-efficient data processing by performing computations directly within the memory devices. A bio-inspired multimodal in-memory computing system capable of performing real-time low power processing of multisensory signals, lowering data conversion and transmission across several modules in conventional chips is introduced. A novel Cu/MoWS₂/VO_x/Pt based multimodal memristor is characterized by an ON/OFF ratio as high as 10⁸ with consistent and ultralow operating voltages of ±0.2 surpassing conventional single-mode memory functions. Apart from observing electrical synaptic behavior, photonic depression and humidity mediated optical synaptic learning is also demonstrated. The heterojunction with MoWS₂ also enables reconfigurable modulation in both memory and optical synaptic functionalities with changing humidity. This behavior provides tunable conductance modulation capabilities emulating synaptic transmission in biological neurons while showing potential in respiratory detection module for healthcare application. The humidity sensing capability is implemented to demonstrate vision clarity using a convolutional neural network (CNN), with different humidity levels applied as a data augmentation preprocessing method. This proposed multimodal functionality represents a novel platform for developing artificial sensory neurons, with significant implications for non-contact human-computer interaction in intelligent systems.

Keywords: crossbar array; heterojunction; high on/off ratio; humidity; in‐memory computing; memristor; multimodal; vanadium oxide.

Figure 1

a) The analog signals from the conventional sensors are converted by the ADC chip into digital signals and then stored in memory. The data in the memory is loaded and transferred as the output signal by the processing units to the memory. b) Single stimuli sensitive memory computing (monomodal) integrates both sensing and storage functions; data can be sent directly to the processing unit, bypassing unnecessary intermediate steps. c) Multimodal devices possess respond simultaneously to different stimuli such as light, humidity, and show electrical synaptic behavior aside from its non‐volatile memory.

Figure 2

a) Cross sectional overview of the device; HRTEM and STEM – EDS mapping of the layers b) HRXPS spectrum of VO _x thin film c) XRD spectrum of SiO₂ (glass substrate used as a reference), MoWS₂ and VO _x thin films d) The surface morphology of the MoWS₂/VO _x film observed by AFM (e) optical bandgap of MoWS₂.

Figure 3

a) Schematic illustration of device structure Cu/MoWS₂/VO_x/Pt memristor device. b) 10 cycles of Current‐voltage characteristics in single logarithmic form of memristor device under compliance current (CC) of 1.5 mA. c) CC dependent multicyclic current voltage characteristics in single logarithmic form. d) Evaluation of CC based per read energy consumption and on/off ratio. e) Depiction of cyclic variation in the switching parameters of memristor devices as V_SET, V_RESET, I_SET and I_RESET, respectively. f) Double logarithmic plot of current voltage characteristics to reveal the conduction mechanisms of resistive switching. g) Retention of I_on and I_off states measured till 40,000 s and h) cumulative probability of respective the resistance states.

Figure 4

a) Schematic illustration of neurotransmission between biological neurons through synaptic cleft between pre and post neurons and proposed physical device‐based neurotransmission through the fabricated Cu/MoWS₂/VO _x /Pt based memristor device. b) Potentiation and depression characteristics of the memristor devices measured under repetitive pulse train of V_Write = 1 V, V_Erase = −1.2 V, and V_read = 0.1 V with psulse width = 200 ns and pulse interval of = 200 ns, respectively.

Figure 5

a) Current versus time curve for the device during light induced RESET followed by 532 nm, and b) 635 nm of wavelengths, respectively. Proposed schematic representation of c) electric set and d) optical reset mechanism of the devices.

Figure 6

a) I–V curves measured at different RH levels b) I–V curves at RH = 95 and 5% demonstrating reversible switching c) Zoomed in graphs (from Figure 6a) showing the lowering of V_TH and effect on the On/Off ratio under the influence of various RH levels. Plausible RS mechanism under the influence of RH d) in SET and e) RESET mode () schematic illustration of the line stacks in MoWS₂ lattice g) HRSTEM images of MoWS₂ (scale = 1 nm) film showing line defects.

Figure 7

a) The distinguished current response based on the breathing patterns b) Relaxed (healthy) and choked (COPD) bronchi c) V_TH plotted against different RH values d) The circuit diagram designed for the humidity adaptive neuron The SPICE simulation results at the RH levels of e) 0.25 and f) 0.5 g) Spike encoding scheme h) Classification results for the spike neural network (SNN) with a 784‐100‐10 architecture.

Figure 8

Humidity mediated light‐induced PSC using the 465 nm blue light a) in ambient atmosphere (laser current: 900 mA) b) under RH of 20, 50, and 90% c) after ramping the RH from 20% up to 90% and ramping down to 20% d) with pulse and interval width of 0.1s for 10, 30, and 50 pulses at RH = 20% (e based on increasing the laser current from 10 – 108 mW cm⁻² at RH = 90% (pulse and interval width = 0.1s). f) STP to LTP based on increasing the laser power from 10 – 108 mW cm⁻² at RH = 60% (pulse and interval width = 0.1s for 20 pulses). g) based on increasing the pulse width (1s, 3s, and 5s) at RH = 20% h) based on increasing the pulse width (1s, 3s, and 5s) at RH = 90%.

Figure 9

a) Image recognition in the light modulated and humidity mediated neuromorphic visual system demonstrating how humidity affects vision clarity. b) Sample figure visualization with RGB channels and sample visualization at different humidity levels. c) Current response under different RH levels and laser currents. d) Training accuracy V/s epochs for different RH levels. The cat image in (b) is adapted from the Oxford‐IIIT Pet Dataset (https://www.robots.ox.ac.uk/~vgg/data/pets/) under the terms of the CC‐BY Creative Commons Attribution 4.0 International license (https://creativecommons.org/licenses/by/4.0).

Figure 10

a) Schematic representation of the memristor crossbar array, b) an inset displaying the FESEM image of an individual device. c) Current‐voltage (I–V) characteristics of a randomly selected memristor device from the crossbar array, demonstrating consistency with previously studied single devices in terms of memory window and operating voltage. d) Cycle‐to‐cycle endurance of randomly selected devices measured up to 600 cycles, with read operations performed at 0.05 V. e) Cumulative probability distribution of the LRS and HRS during endurance cycles. e) Cycle‐to‐cycle variability in resistive switching voltages (V_SET and V_RESET) analyzed over 100 DC cycles. f) Comparative analysis of the SET/RESET voltages across different studies. g) Comparative analysis of the SET/RESET voltages.^{[ 40 ]} h) Humidity‐responsive I‐V characteristics exhibit comparable trends under crossbar conditions. i) Optically stimulated response under illumination with a 465 nm light source.