A scalable solution recipe for a Ag-based neuromorphic device

Abstract

Integration and scalability have posed significant problems in the advancement of brain-inspired intelligent systems. Here, we report a self-formed Ag device fabricated through a chemical dewetting process using an Ag organic precursor, which offers easy processing, scalability, and flexibility to address the above issues to a certain extent. The conditions of spin coating, precursor dilution, and use of solvents were varied to obtain different dewetted structures (broadly classified as bimodal and nearly unimodal). A microscopic study is performed to obtain insight into the dewetting mechanism. The electrical behavior of selected bimodal and nearly unimodal devices is related to the statistical analysis of their microscopic structures. A capacitance model is proposed to relate the threshold voltage (V_th) obtained electrically to the various microscopic parameters. Synaptic functionalities such as short-term potentiation (STP) and long-term potentiation (LTP) were emulated in a representative nearly unimodal and bimodal device, with the bimodal device showing a better performance. One of the cognitive behaviors, associative learning, was emulated in a bimodal device. Scalability is demonstrated by fabricating more than 1000 devices, with 96% exhibiting switching behavior. A flexible device is also fabricated, demonstrating synaptic functionalities (STP and LTP).

Keywords: Associative learning; Chemical process; Dewetting; Hierarchical structures; Neuromorphic device; Scalability; Self-forming.

Fig. 1

a Schematic demonstrating fabrication protocol of the dewetted Ag synaptic device with in-plane Au electrodes. SEM image of a typical dewetted film consisting of a hierarchical network structure is shown on the right. b 3D plot showing the variation in dewetting parameters (spin coating speed, precursor: solvent ratio, IPA: water ratio in the solvent) explored to obtain the desired dewetting pattern with the corresponding SEM/optical images displayed adjacent to the data point (red box). The optical images refer to the data points 12 to 15 (Scale bar: 0.1 cm, indicated in image 15), while others are SEM images (Scale bar: 1 µm, indicated in image 7). The scale bar applies to all optical and SEM images

Fig. 2

a–c SEM images of samples 2, 10, and 11 exhibiting the bimodal distribution and the histograms (adjacent) corresponding to the particle to particle distance, island to island distance, particle area (blue), and island area (gray) of the dewetted patterns. d–f SEM images of samples 4, 9, and 16 exhibiting nearly unimodal distributions. However, some distribution in the particle size is seen

Fig. 3

a Thermogravimetric analysis of the Ag ink (scan rate: 10 °C/min). b Raman spectra show a decline in organic constituents with increasing annealing time. The bands at 648, 934, 1014, 1168 cm⁻¹ (bands 1–4) correspond to the stretching vibrations of C–S–C, C–O–C, C–C, and C=S, respectively. The other bands at 1297, 1366, 1542 cm⁻¹ (bands 5–7) arise due to the symmetric and anti–symmetric C=O stretching of the carboxylic group, respectively [52, 53]. c Schematic showing the dewetting process where the growth of regions devoid of solvent (called holes) due to evaporation leads to a closely packed structure consisting of Ag particles. d–i SEM images of the Ag film annealed at ~ 190 °C for different time durations to probe the evolution of dewetted structures (Scale bar: 1 µm). j Time-variation of crystallite size and particle size with annealing, as obtained from the XRD pattern and SEM images, respectively. The corresponding XRD patterns are given in k. The peaks match the face-centered cubic (fcc) structure of Ag (JCPDS:04–0783)

Fig. 4

a Lists of samples with dewetted patterns used as active elements in devices (D). I-V characteristics of devices with b nearly unimodal and c bimodal distributions and d those obtained with different annealing times. Current compliance (I_CC) is set at 10 µA in all cases

Fig. 5

a V_th values versus the Ag fill factors in devices D4 to D17. The error bars indicate the deviation of the V_th for three devices in each case. b Schematic of a device-equivalent, parallel plate capacitor model with gray squares indicating the presence of Ag islands (large with $B$ sides) and nanoparticles (small with $a$ sides). The distance between the particles and the Ag islands are $a\prime$ and $B\prime$, respectively. c Calculated elastance versus the Ag fill factor for devices D4 to D17. d Variation in V_th for devices with similar FF of ~ 42% for D1 and D16 and ~ 45% for D3 and D6. e Variation in the elastance values for devices with similar FF. However, D1 and D3 are nearly unimodal (red bars), and D16 and D6 (blue) are bimodal

Fig. 6

a, b (states I and II) Schematic showing the state of the nearly unimodal and the bimodal device before and after pulsing, respectively. The green dotted circle represents the conductive filamentary path possibly formed during pulsing (state II). Short-term potentiation (STP) and long-term potentiation (LTP) were emulated in the devices c D1 at the set I_CC of 600 and 900 µA while exhibiting a retention time (t_r) of 9 and 85 s, respectively, d D6 at the set I_CC of 300 and 600 µA while exhibiting a t_r of 11 and 245 s, respectively. The bottom panel in c and d represents 50 pulses of 5 V with 200 ms width as well as the interval. The reading voltage is 100 mV

Fig. 7

a Associative learning emulated in the device, D6. Food (5 pulses, 2.5 V) and bell (5 pulses, 0.5 V) signals were fed to the device consecutively. The device responded to the food signal but not to the bell signal. During training, the food and bell (50 sets) signals were fed with no gap. After training, the device responds to the bell signal alone with a retention of up to 8 min. All the pulses were of 1 s width and interval. b Photograph of the large area (~ 10 × 10 cm²) dewetted Ag device containing 225 devices each of ~ 600 × 40 µm² area. c I–V characteristics of 75 devices; only 75 out of 225 curves are shown for the sake of clarity. d Color map of switching voltages of the 225 devices. e Histogram showing the variation in the switching voltage for the devices

Fig. 8

a Photograph of the flexible device along with the schematic of the device in the flat and bent states. b–c SEM image and d I-V characteristics of the device in the flat and bent states. e STP and LTP emulated in the flexible device (20 pulses, 2 V: bottom panel) at I_CC of 150 µA (STP) and 250 µA (LTP), in the flat and bent states