A photovoltaic cell and related photovoltaic reservoir computing system

Abstract

A photovoltaic cell (110) is disclosed, which comprises: a first layer 110-a comprising a substrate arranged to include a transparent conducting material, a second layer 110-b configured as a hole transport layer, a third layer 110-c configured as a semiconductor layer comprising a halide perovskite, a fourth layer 110-d configured as an electron transport layer, and a fifth layer 110-e comprising a conductive contact.

Description

10202401929X filed in the Intellectual Property Office of Singapore on 1 July 2024, the entire contents of which are incorporated herein by reference.TECHNICAL FIELD

[0002] The following relates generally to in-sensor reservoir computing, and more specifically, it relates to a photovoltaic cell and related photovoltaic reservoir computing system configured for in-sensor reservoir computing.BACKGROUND

[0003] For completeness, it is hereby clarified that reference made to the definition of format: “[ref. X]” in any paragraph(s) in the description of the present disclosure is to be construed to refer to the corresponding citation “X” in the “References” section of the present disclosure. For example, [ref. 10] refers to citation

[0010] listed at the “References” section of the present disclosure, while [ref. 4-7] correspondingly refers to citations [4]-[7] mutatis mutandis.

[0004] The Internet of Things (loT) era, driven by the growth of novel sensor technologies, advanced communication techniques and Artificial Intelligence (Al) algorithms, has opened new opportunities in wearable sensing devices [ref. 1], These wearable sensing devices may monitor various physiological parameters such as blood pressure, heart rate, and metabolite levels and track their variations over time with advanced algorithms flagging significant deviations from a baseline level [ref. 2-3], Paradoxically, while the software algorithms based on Deep Neural Networks (DNN) benefit from increased amounts of data, shuttling the data from the sensor to the centralised server (cloud) infrastructure becomes a significant bottleneck for the hardware, due to prohibitive energy costs of data communication [ref. 4], Massive processing of such data in the cloud is set to consume a substantial portion of global electricity [ref. 5], Furthermore, there is an increased threat of private information leakage with this model of cloud computing prevalent in the majority of current loT nodes [ref. 6-7],

[0005] These drawbacks have led to advancement in the idea of dispersing the computing infrastructure through the development of edge computing, i.e., computation in proximity to the data source [ref. 8-9], Typically, these devices embed intelligence in the form of low-power processors to implement AT algorithms near the sensor [ref 10-1 1 ], such that sensitive information does not have to be transmitted over the air. However, these von Neumann processors are extremely inefficient in handling the workloads of state-of-the-art DNNs again due to the shuttling of data, this time between memory and processor. In stark contrast, biological counterparts, such as the mammalian nervous system, present a more efficient framework that co-localises sensing, memory, and computing.

[0006] The raw input from the environment is hierarchically pre-processed in proximity to the sensory receptors. Low-level pre-processing performed by biological receptors extracts salient features from unstructured data to reduce information that must be transmitted over longer distances. Memristive technology has emerged as a new enabler to mimic such brain- inspired computing-in-memory (CIM). However, unlike current ANN training methods, biological synaptic weights are modified over various timescales, from a few seconds by shortterm plasticity to a few hours by long-term plasticity or even days by structural plasticity or synaptic rewiring. Such dynamic processes have been largely unused in the memristive DNN demonstrations so far.

[0007] Further improvements in edge-computing sensory platforms are achieved by extending the concept of in-memory computing to in-sensor computing. While conventional sensors typically merely transduce the input stimuli into electrical outputs, in-sensor computing nodes embed spatiotemporal information of the input in their response characteristics. Consequently, computational operations, such as noise reduction, adaptation, and saliency filtering, may be derived directly from the sensor output. In doing so, computational sensors become context-aware and can reduce the reliance on distant computing, resulting in energy efficiency, reduced data shuttling as well as improved security [ref. 12], Thus, to minimise the shuttling of information, unifying sensing, memory, and computing at the edge could help construct efficient sensor networks [ref. 13-14],

[0008] Additionally, this integration would inherently decrease latency and spatial footprint and improve energy efficiency [ref. 15-16], In turn, this has led to the conception of in-sensor neuromorphic computing [ref. 17]. Computational sensors have demonstrated first- stage image processing, such as image contrast enhancement and noise reduction in anoptoelectronic memristor array [ref. 18], as well as binary classification and edge detection using low-latency ferroelectric sensors [ref. 19], In-sensory tactile processing has also been demonstrated using a mechanically gated transistor, which could perform healthy tissue cl ssification [ref. 20],

[0009] A promising extension of in-sensor computing, by combining with the time-varying dynamics of synapses, emerges by complementing it with the principles of physical reservoir computing (RC). RC is originally a software computation framework derived from recurrent neural networks (RNNs) and uses a fixed, non-linear system called a reservoir 100a - see FIG. 1(a) (top part). The reservoir 100 comprises large number of sparsely connected nodes (being analogous to artificial neurons), which consist of non-linear activation functions. The interconnectedness nature of the reservoir nodes enables feedback from the inputs in the previous timesteps. This equips the reservoir to map input signals into higher dimensional spaces for facile classification 105 (i.e. refer to FIG. lb). RC in software is used for pattern recognition and time-series forecasting in tasks such as weather and financial market forecasting [ref. 21], Physical RC exploits the dynamics of material substrates as the ‘reservoir’, instead of the artificially simulated nodes. Eliminating the dependence on digital processors, it converges neuromorphic and reservoir computing, enabling the processing of time-varying analog stimuli efficiently.

[0010] Among sensory modalities, vision requires the transfer of most information and hence these sensors are the focus for in-sensor computing studies [ref. 22-28], However, prior optoelectronic RC implementations measure change in photoconductivity through read voltage pulses, which consume significant energy. Moreover, low photocurrents are susceptible to noise and require current to voltage conversion for a sensitive readout circuit. Additionally, many of these devices show short-term to long-term memory transition on prolonged stimulation, which require erasing operation for resetting, thereby adding to the energy burden and circuit complexity. Thus, photoactive materials that may play host to non-linear transient phenomena are worth exploring as potential viable candidates for in-sensor optoelectronic RC.

[0011] So, there is a need for a solution that may address at least one of the problems of the prior art, and / or to provide a choice useful in the art.SUMMARY

[0012] The described disclosure herein may relate to a photovoltaic cell and a related photovoltaic reservoir computing system.

[0013] According to a first aspect, there is provided a photovoltaic cell comprising: a first layer comprising a substrate arranged to include a transparent conducting material, a second layer configured as a hole transport layer, a third layer configured as a semiconductor layer comprising a halide perovskite, a fourth layer configured as an electron transport layer, and a fifth layer comprising a conductive contact.

[0014] Preferably, the halide perovskite may be of the formula ABX3, where: A is selected from a monovalent metal cation or an organic cation; B is a metal cation; and X is a halide ion. Additionally or optionally, A may be selected from Cs, CHsNHj" or [HC(NH2)z]~; B may be selected from Pb or Sn; and X may be selected from Cl, Br or I.

[0015] Preferably, the transparent conducting material may comprise one or more of indium tin oxide (ITO), fluorinated tin oxide (FTO), silver nanowires, carbon nanotubes, metal mesh, aluminium zinc oxide, gallium zinc oxide or indium zinc oxide.

[0016] Preferably, the second layer may comprise Spiro-OMeTAD, poly[bis(4- phenyl)(2,4,6-trimethylphenyl)amine]) (PTAA), Poly(3 -hexylthiophene) (P3HT), Tris(4- carbazoyl-9-ylphenyl)amine (TCTA), nickel oxide, copper iodide, copper thiocyanate, molybdenum oxide, vanadium oxide or poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS).

[0017] Preferably, the fourth layer may comprise one or more of titanium oxide, zinc oxide, stannic oxide, tungsten trioxide, 2,2',2"-(l,3,5-benzinetriyl)-tris(l-phenyl-l-H-benzimidazole) (TBPi), C60 and phenyl-C61 -butyric acid methyl ester (PCBM).

[0018] Preferably, the fifth layer may comprise one or more of lithium fluoride, calcium fluoride, titanium oxide, aluminium, silver, copper, titanium or gold.

[0019] Preferably, the first layer may comprise an indium tin oxide (ITO)-coated glass substrate, the second layer may comprise poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS), the third layer may comprise CHsNEfiPbBr? or CEENEEPbE, the fourth layer may comprise phenyl -C61 -butyric acid methyl ester (PCBM), and the fifth layer may comprise lithium fluoride and aluminium.

[0020] Preferably, the third layer may have a thickness of between about 180 nm to about 600 nm.

[0021] According to a second aspect, there is provided a method of preparing a photovoltaic cell, the method comprising the steps of: (a) applying a second layer to a first layer, wherein the first layer comprises a substrate arranged to include a transparent conducting material, and the second layer is configured as a hole transport layer, (b) applying a third layer to the second layer, wherein the third layer comprises a halide perovskite, (c) applying a fourth layer to the third layer, wherein the fourth layer is configured as an electron transport layer, (d) applying a fifth layer to the fourth layer, wherein the fifth layer comprising a metal contact.

[0022] Preferably, step (a) may further comprise, before step (a), the step of (al) cleaning the first layer.

[0023] Preferably, step (al) may further comprise the step of (a2) treating the cleaned first layer using UV-ozone.

[0024] Preferably, step (b) may further comprise solvent annealing after applying the third layer.

[0025] Preferably, the method may further comprise the step of (e) encapsulating the photovoltaic cell with epoxy resin and glass.

[0026] According to a third aspect, there is provided a photovoltaic cell obtained or obtainable from the method as described herein.

[0027] According to a fourth aspect, there is provided a photovoltaic panel array comprising a plurality of photovoltaic cell as described herein.

[0028] According to a fifth aspect, there is disclosed a photovoltaic reservoir computing (PVRC) system, comprising: a receiver configured to receive optical pulse signals associated with spatiotemporal characteristics of an entity, said optical pulse signals characterised by variations at least in intensity due to changes in the spatiotemporal characteristics; and a photodetector configured to provide, by way of in-sensor reservoir computing based upon said received optical pulse signals, corresponding unique photovoltages as output to enable determination of the spatiotemporal characteristics, wherein the unique voltages are respectively stateful representations in time for the spatiotemporal characteristics, wherein thephotodetector includes at least one photovoltaic cell of the first aspect, or the photovoltaic panel array of the fourth aspect.

[0029] Preferably, the PVRC system may be adapted for photoplethysmography (PPG) and the entity may be a human subject Tn this instance, the PVRC system may further comprise: light illuminating means configured to illuminate green light pulses onto the skin tissue of the human subject, wherein the receiver is further configured to receive, based at least in part on the illuminated green light, light pulses reflected off the skin tissue, whereby the reflected light pulses vary temporally in intensity being responsive to changes in blood volume in a microvascular bed under the skin tissue; and wherein the reflected light pulses are the optical pulse signals and the spatiotemporal characteristics is associated with atrial fibrillation (AF).

[0030] Preferably, the PVRC system may be configured as a wearable device

[0031] Preferably, the light illuminating means may be a light emitting diode (LED) configured as an emitter of green light pulses.

[0032] Preferably, the PVRC system may further comprise: output means configured to provide readout of the photovoltages as the output, wherein the photovoltages are open-circuit voltages (70C).

[0033] Preferably, the PVRC system may be configured to operate with power less than 1 fW.

[0034] Preferably, the optical pulse signals may be in the form of light pulse trains.

[0035] According to a sixth aspect, there is disclosed a method of operating the PVRC system of the fifth aspect, comprising: receiving, by the receiver, optical pulse signals associated with spatiotemporal characteristics of an entity, said optical pulse signals characterised by variations at least in intensity due to changes in the spatiotemporal characteristics; and providing, by the photodetector by way of in-sensor reservoir computing based upon said received optical pulse signals, corresponding unique photovoltages as output to enable determination of the spatiotemporal characteristics, wherein the unique voltages are respectively stateful representations in time for the spatiotemporal characteristics.

[0036] Additional benefits and advantages of the disclosed aspects may become apparent from the specification and drawings. The benefits and / or advantages may be individuallyobtained by the various aspects and features of the specification and drawings, which need not all be provided in order to obtain one or more of such benefits and / or advantages.BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying figures, where like reference numerals refer to identical or functionally similar elements throughout the separate views and which together with the detailed description below are incorporated in and form part of the specification, serve to illustrate various aspects and to explain various principles and advantages in accordance with the present disclosure.FIG. 1(a)

[0038] FIG. 1(a) is a schematic representation of two approaches of in-sensor reservoir computing, in which a conventional reservoir comprises several sparsely connected nodes, and an unconventional reservoir is obtained through a single dynamical node with time-delayed feedback, in accordance with aspects of the present disclosure.FIG. 1(b)

[0039] FIG. 1(b) shows inputs belonging to two classes (circles and stars) that are linearly inseparable in a two-dimensional (2D) feature space, in which a reservoir performs non-linear transformation to map the inputs into higher dimensions and render them linearly separable for facile classification, in accordance with aspects of the present disclosure.FIG. 1(c)

[0040] FIG. 1(c) is a perspective view of a p-i-n device stack of a halide perovskite photovoltaic used for photovoltaic reservoir computing, in accordance with aspects of the present disclosure.FIG. 1(d)

[0041] FIG. 1 (d) is a plot depicting transient Vocevolution compared between a halide perovskite photovoltaic and a silicon photodetector with a pulsed light stimulation, in accordance with aspects of the present disclosure.FIG. 1(e)

[0042] FIG. 1(e) is a schematic description of integration of in-sensor learning in the halide perovskite photovoltaic of FIG. 1(c) to sense and compute physiological information from photoplethysmography (PPG), in accordance with aspects of the present disclosureFIG. 1(f)

[0043] FIG. 1 (f) is a schematic diagram of a photovoltaic reservoir computing (PVRC) system arranged to incorporate a halide perovskite photovoltaic, in accordance with aspects of the present disclosure.FIG. 1(g)

[0044] FIG. 1(g) is a flow diagram of a method of operating the PVRC system of FIG. 1(f), in accordance with aspects of the present disclosure.FIG. 1(h)

[0045] FIG. 1(h) is a schematic view of the halide perovskite photovoltaic of FIG. 1(c), in accordance with aspects of the present disclosure.FIG. l(i)

[0046] FIG. l(i) is a flow diagram of a method of preparing the the halide perovskite photovoltaic of FIG. 1(c), in accordance with aspects of the present disclosure.FIG. 2(a)]0047] FIG. 2(a) is a plot depicting temporal correlation of the transient Vocobserved by stimulating the PVRC system with different frequencies of light pulses, in accordance with aspects of the present disclosure.FIG. 2(b)

[0048] FIG. 2(b) is a plot depicting the evolution of the transient Vocwith light pulses of different intensities, in accordance with aspects of the present disclosure.FIG. 2(c)

[0049] FIG. 2(c) is a plot depicting state-dependent response to light pulses shown through paired-pulse facilitation (PPF), in accordance with aspects of the present disclosure.FIG. 2(d)

[0050] FIG. 2(d) is a plot depicting PPF measured for varying At, in accordance with aspects of the present disclosure.FIG. 2(e)

[0051] FIG. 2(e) is a plot depicting response to a single light pulse shown for a PVRC system configured with different perovskite layer thicknesses, in accordance with aspects of the present disclosure.FIG. 2(f)

[0052] FIG. 2(f) is a plot depicting PPF index of the photosensor of three different thicknesses shown as a function of pulse timing, in accordance with aspects of the present disclosure.FIG. 3(a)

[0053] FIG. 3(a) is a plot depicting UV-Vis absorption spectrum of spin-coated MAPbBr3thin films of three different thicknesses and concentrations, in accordance with aspects of the present disclosure.FIG. 3(b)

[0054] FIG. 3(b) is a corresponding Tauc plot of spin-coated MAPbBr3thin films of three different thicknesses and concentrations, in accordance with aspects of the present disclosure.FIG. 4(a)

[0055] FIG. 4(a) is a plot depicting minimal cycle-to-cycle variation in the PPF indices in 1000 cycles, in accordance with aspects of the present disclosure.FIG. 4(b)

[0056] FIG. 4(b) depicts observation of a standard deviation of 0.24 % in 1000 cycles, in accordance with aspects of the present disclosure.FIG. 5(a)

[0057] FIG. 5(a) is a plot for PPF index measured at a timescale with light pulses of duration at 0. 1 ms, in accordance with aspects of the present disclosure.FIG. 5(b)

[0058] FIG. 5(b) is a plot for PPF index measured at a timescale with light pulses of duration at 1 ms, in accordance with aspects of the present disclosure.FIG. 5(c)

[0059] FIG. 5(c) is a plot for PPF index measured at a timescale with light pulses of duration at 10 ms, in accordance with aspects of the present disclosure.FIG. 6(a)

[0060] FIG. 6(a) is a plot mapping various combinations of a 4-bit input into different reservoir states, whereby 16 unique reservoir states are obtained at the end of the 4-bit pulse train, in accordance with aspects of the present disclosure.FIG. 6(b)

[0061] FIG. 6(b) shows a scheme of using 4-bit reservoir states to pre-process a binarized MNIST image, in accordance with aspects of the present disclosure.FIG. 6(c)

[0062] FIG. 6(c) shows a neural network architecture to compare the effect of PVRC preprocessing against a baseline artificial neural network, in accordance with aspects of the present disclosure.FIG. 6(d)

[0063] FIG. 6(d) is a plot relating to recognition accuracy of a PVRC pre-processed neural network and a typical artificial neural network, in accordance with aspects of the present disclosure.FIG. 7(a)

[0064] FIG. 7(a) is a plot of voltage versus time relating to a pulsing scheme of 4-bit light pulses represented for a 4-bit input of ‘1010’, in accordance with aspects of the present disclosure.FIG. 7(b)

[0065] FIG. 7(b) is a plot of voltage versus time relating to a pulsing scheme of 4-bit light pulses represented for a 4-bit input of ‘ 1011’, in accordance with aspects of the present disclosure.FIG. 8(a)

[0066] FIG. 8(a) depicts a box plot of 20 values obtained for the same input, relating to low cycle-to-cycle variation in the mapping of 4-bit light pulse inputs over 20 cycles with a maximum standard deviation of 1.8%, in accordance with aspects of the present disclosure.FIG. 8(b)

[0067] FIG. 8(b) depicts a collection of plots pertaining to 20 transient evolutions of Vocoverlapped for 15 unique inputs, relating to low cycle-to-cycle variation in the mapping of 4- bit light pulse inputs over 20 cycles with a maximum standard deviation of 1.8%, in accordance with aspects of the present disclosure.FIG. 9

[0068] FIG. 9 shows significant device-to-device variation in the mapping of 4-bit inputs with a minimum standard deviation of 29.7% in the voltage level for the 4 -bits inputs among 10 devices, in accordance with aspects of the present disclosure.FIG. 10(a)

[0069] FIG. 10(a) is a schematic workflow of a NARMA2 task, in accordance with aspects of the present disclosure.FIG. 10(b)

[0070] FIG. 10(b) is a plot pertaining to reconstruction of the second-order non-linear function during the training phase, in accordance with aspects of the present disclosure.FIG. 10(c)

[0071] FIG. 10(c) is a plot pertaining to prediction of the function using the trained model, in accordance with aspects of the present disclosure.FIG. 11(a)

[0072] FIG. 11(a) is a plot depicting the transient evolution of measured voltage between timesteps, as randomly generated inputs are provided as a light pulse train to the PVRC system, in accordance with aspects of the present disclosure.FIG. 11(b)

[0073] FIG. 1 1 (b) is a plot of input values (u) in NARMA2 encoded into corresponding light pulse intensities, in accordance with aspects of the present disclosure.FIG. 12

[0074] FIG. 12 is a plot depicting voltage outputs from 6 devices relative to timesteps, in accordance with aspects of the present disclosure.FIG. 13(a)(i)

[0075] FIG. 13 (a)(i) shows a schematic representation of generation of PPG signals by the reflection of green light from an artery, in accordance with aspects of the present disclosure.FIG. 13(a)(ii)

[0076] FIG. 1 (a)(ii) shows comparison of the PPG signals generated from a healthy heart and a diseased heart, in accordance with aspects of the present disclosure.FIG. 13(a)(iii)

[0077] FIG. 13(a)(iii) shows PPG signals taken from the MIMIC PERform dataset and provided to the PVRC system for pre-processing, in accordance with aspects of the present disclosure.FIG. 13(b)

[0078] FIG. 13(b) depicts a sampling protocol to obtain the reservoir state as an N* 5 input vector for classification using a feed-forward neural network, in accordance with aspects of the present disclosure.FIG. 13(c)

[0079] FIG. 13(c) depicts receiver operating characteristics showing respective areas under a receiver operating characteristic (ROC) curve vis-a-vis scenarios with and without using the PVRC system for a classifier, in accordance with aspects of the present disclosure.FIG. 13(d)

[0080] FIG. 13(d) depicts confusion matrices showing classification performance for 1 1 test cases, in accordance with aspects of the present disclosure.FIG. 14(a)

[0081] FIG. 14(a) depict an example plot of PPG signals present in the MIMIC PERform database, in accordance with aspects of the present disclosure.FIG. 14(b)

[0082] FIG. 1 (b) depict an example plot of PPG signals present in the MIMIC PERform database, in accordance with aspects of the present disclosure.FIG. 15(a)

[0083] FIG. 15(a) depicts an example plot of a low dimensional representation of the data used for the classification, in accordance with aspects of the present disclosure.FIG. 15(b)

[0084] FIG. 15(b) depicts an example plot of a low dimensional representation of the data used for the classification, in accordance with aspects of the present disclosure.FIG. 16(a)

[0085] FIG. 16(a) is a schematic showing a lateral device with symmetrical gold electrodes on SiCh / Si substrate, in accordance with aspects of the present disclosure.FIG. 16(b)

[0086] FIG. 16(b) is a schematic showing a SEM image of the device’s active area, in accordance with aspects of the present disclosure. A poling electric field of 2.5 V / pm is applied across a gap of 15 pm. Subsequently, an EDX elemental line scan was performed to obtain the halide concentration profile.FIG. 16(c)

[0087] FIG. 16(c) is a graph showing the bromine concentration profile before and after poling the CH3NH3PbBr3sample, in accordance with aspects of the present disclosure. The graph shows that a significant concentration gradient developed due to the migration of bromide ions towards the positive terminal after poling.FIG. 16(d)

[0088] FIG. 16(d) is a graph showing the iodine concentration profile before and after poling the CH3NH3PbBr3sample, in accordance with aspects of the present disclosure. The graph shows that a minor concentration gradient formed as compared to the CH3NH3PbI3sample.FIG. 16(e)

[0089] FIG. 16(e) is a graph showing that the switchable photovoltaic effect produces a greater open-circuit voltage ( Voc) in the lateral configuration of CH3NH3PbBr3than CH3NH3PbI3, in accordance with aspects of the present disclosure.FIG. 16(f)

[0090] FIG. 16(f) is a graph showing the transient Vocresponse of identical photovoltaic devices of CH3NH3PbBr3and CH3N H3PbI3to a light pulse train, in accordance with aspects of the present disclosure. The results indicate that CH3NH3PbBr3demonstrates short-term memory characteristics.FIG. 17(a)

[0091] FIG. 17(a) is a schematic showing kinetic charge heterogeneity of the electronic and ionic carriers in the duration of an optical pulse illumination, including states: I. Equilibrium under dark; II Initial fast rise; III. Equilibrium under light; and IV. Decay under dark, in accordance with aspects of the present disclosure.FIG. 17(b)

[0092] FIG. 17(b) is diagram showing that in the dark equilibrium state, an ionic concentration gradient is formed with excess cations and anions accumulating near the holetransport layer (HTL) and electron-transport layer (ETL), respectively, in accordance with aspects of the present disclosure.FIG. 17(c)

[0093] FIG. 17(c) is graph showing that when an illumination light is introduced, the ionic accumulation layer acts as a screening layer on photogenerated electrons and holes and accelerates interfacial recombination, in accordance with aspects of the present disclosure. A characteristic energy valley in the conduction band can be observed.FIG. 17(d)

[0094] FIG. 17(d) is graph showing that under open-circuit conditions, the ionic accumulation layer is partially dispersed during the illumination, which could be observed to slowly recover towards equilibrium in the dark, in accordance with aspects of the present disclosure.FIG. 17(e)

[0095] FIG. 17(e) is graph showing that the slow rise and decay in the measured Vocfollows a simulated cell in the same operating conditions, in accordance with aspects of the present disclosure.FIG. 18

[0096] FIG. 18 is a series of images showing EDX elemental mapping of Au / CH3NH3PbBr3 / Au device comparing pristine and poled conditions, in accordance with aspects of the present disclosure, scale at 5 pm. The poled device shows reduced bromide concentration near the cathode which indicates migration of bromide ions.FIG. 19(a)

[0097] FIG. 19(a) is a graph showing temporal evolutions of Vocunder illuminated and dark conditions for a CH3NH3PbBr3photovoltaic device, in accordance with aspects of the present disclosure.FIG. 19(b)

[0098] FIG. 19(b) is a graph showing temporal evolutions of Vocunder illuminated and dark conditions for a CH3NH3PbI3photovoltaic device, in accordance with aspects of the present disclosure. Comparing FIG. 19(a) with FIG. 19(b) shows that the time scale of equilibration for CH3NH3PbBr3is much slower than that of CH3NH3PbI3.FIG. 20

[0099] FIG. 20 is a schematic depiction of the Driftfusion simulation model of the device, in accordance with aspects of the present disclosure. The optoelectronic material parameters are adopted from literature.DETAILED DESCRIPTION

[0100] Aspects according to the present disclosure will be described, by way of example only, with reference to the drawings. Like reference numerals and characters in the drawings refer to like elements or equivalents.

[0101] Aspects of the present disclosure set out a photovoltaic cell and a related photovoltaic reservoir computing (PVRC) system 150 (i.e. see FIG. 1(f)) for in-sensor reservoir computing. More specifically, based on aspects of the present disclosure, the photovoltaic cell may be a halide perovskite based solar cell 110 that is proposed to be incorporated as part of the PVRC system 150. In this regard, FIG. 1(c) is a perspective view of a p-i-n device stack of the halide perovskite solar cell 110, in which an example cubic crystal structure of CH3NH3PbBr3is depicted, used for photovoltaic reservoir computing (RC), in accordance with aspects of the present disclosure. For sake of brevity, the halide perovskite solar cell 110 is hereafter referred to as: “perovskite photovoltaic 110”.

[0102] The following description provides examples of the perovskite photovoltaic 110 and the PVRC system 150, but they are not limiting on the scope, applicability, or examples set forth in the claims. Changes may be made in the function and arrangement of elements discussed without departing from the scope of the disclosure. Various examples may omit, substitute, or add various procedures or components as appropriate. For instance, the methods described may be performed in an order different from that described, and various steps may be added, omitted, or combined. Also, features described with respect to some examples may be combined in some other examples. For example, an apparatus may be implemented or a method may be practiced using any number of the aspects set forth herein. In addition, the scope of the disclosure is intended to cover such an apparatus or method which is practiced using other structure, functionality, or structure and functionality in addition to, or other than, the various aspects of the disclosure set forth herein. It should be understood that any aspect of the disclosure disclosed herein may be embodied by one or more elements of a claim. The word “exemplary” is used herein to mean “serving as an example, instance, or illustration”.Any aspect described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other aspects.

[0103] Regarding the perovskite photovoltaic 110, excellent optoelectronic properties (i.e. high optical absorption coefficients, long electronic carrier diffusion lengths [ref. 29]) are observed in low temperatures solution based methods, due to its highly ionic nature and low formation energies that ease the crystallization. However, the soft ionic bond also results in high ionic activity [ref. 30], be it at the bulk or interface, that trigger hysteresis responses. Typically, solar cells are characterised by fast measurements such as voltage sweeps, which yields key photovoltaic parameters such as short-circuit current (Jsc) and open-circuit voltage (Foc). However, measuring the solar cell parameters as a function of time reveal dynamics that may be utilised for unconventional computing.

[0104] It is to be appreciated that the disclosed perovskite photovoltaic 110 may compute with transient, stateful Vocand thus, are well-suited to implement an in-sensor low-power, optoelectronic RC without need for additional peripheral circuits. Conventional renditions of optoelectronic RC employ photo-memristors having readout that requires energy-consuming external read voltages. The perovskite photovoltaic 100, on the other hand, may respond to optical stimulation by innately generating photovoltages and photocurrents which can directly be read out, without any external bias. Such self-powered computational sensors further reduce the operational energy budget. Conventional (silicon-based) photodetectors are designed to produce a fast response to stimuli, there is no latency which makes them state independent. However, delayed response beneficially imparts a memory function to photodetectors (e.g. made of perovskite photovoltaics).

[0105] The spatiotemporal dynamics of the Vocare studied by stimulating with light pulses (e.g. wavelength: 525 nm; intensity: 4.3 mWcm’2; pulse width: 1 ms; and duty cycle: 50%). A light pulse train at 1000 Hz frequency is transmitted (i.e. illuminated on) to the perovskite photovoltaic 110 and a conventional silicon photodetector. The Vocof the silicon photodetector is identical at the end of each pulse, because of an instantaneous response to stimulation. On the other hand, in the perovskite photovoltaic 110, the Vocsequentially rises with each light pulse (i.e. see FIG. 1(d), which is a plot 115 depicting transient Vocevolution compared between the perovskite photovoltaic 110 and the silicon photodetector with a pulsed light stimulation). This increment is analogous to the strengthening of biological synapses with repetitive activity known as potentiation [ref. 31], emphasizing the essence of the slowdynamics. The facilitation characteristic in Vocmay enable the mapping of unique light pulse trains into distinct states of the state reservoir.

[0106] From the proposed PVRC system 150, unique time series inputs are shown to be transformed into a higher dimensional feature space enabling easy extraction of higher order information. This is corroborated by benchmarking the performance of the proposed PVRC system 150 in image classification and time-series prediction tasks. An example practical utility of the proposed PVRC system 150 is highlighted in integration with a widely used biomedical analysis technique for monitoring cardiac health-Photoplethysmography (PPG) (i.e. refer to FIG. 1(e)).|0107| FIG. 1(f) is a schematic diagram of the PVRC system 150 that includes the perovskite photovoltaic 110, in accordance with aspects of the present disclosure. Broadly, the PVRC system 150 may comprise: a receiver 150-a configured to receive optical pulse signals 160 associated with spatiotemporal characteristics of an entity (not shown), said optical pulse signals 160 characterised by variations at least in intensity due to changes in the spatiotemporal characteristics; and a photodetector 150-b (or otherwise a photosensor) configured to provide, by way of in-sensor reservoir computing based upon said received optical pulse signals 160, corresponding unique photovoltages as output to enable determination of the spatiotemporal characteristics. The unique voltages are respectively stateful representations in time for the spatiotemporal characteristics. Particularly, the in-sensor reservoir computing converts or transforms the received optical pulse signals 160 into the corresponding unique photovoltages that are provided as the output. The photodetector 150-b includes at least one perovskite photovoltaic 110, or a photovoltaic panel array (not shown) being an arrangement of a plurality of perovskite photovoltaics 110. The optical pulse signals 160 may be received as light pulse trains.

[0108] Additionally or optionally, the PVRC system 150 may further comprise output means 150-c configured to provide readout of the photovoltages as the output, in which the photovoltages are (transient) open-circuit voltages (VQC). In some examples, the PVRC system 150 may be operated with a power less than 1 fW.

[0109] In an example (as a practical application), the PVRC system 150 is adapted for photoplethysmography (PPG), and the entity is a human subject. The PVRC system 150 may be configured as a wearable device to be worn by the human subject (e g. as a fitness tracker for personal health monitoring). In this case, the PVRC system 150 may further, additionallyor optionally, comprise a processor 150-d and light illuminating means 150-e configured to illuminate green light pulses 170 onto the skin tissue of the human subject. The light illuminating means 150-e may be a light emitting diode (LED) configured as an emitter of green light pulses 170. The processor 150-d, which is coupled to the light illuminating means 150-e, may be configured to receive control signals (or information) to enable programming and controlling of the light illuminating means 150-e for emitting the green light pulses 170.

[0110] It is to be appreciated that the processor 150-d may be a general-purpose processor, a DSP, an ASIC, a CPU, an FPGA or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the described functions. A general-purpose processor may be a microprocessor, although alternatively, the processor may be any processor, controller, microcontroller, or state machine.[OHl] Information and signals described herein may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips that may be referenced throughout the description may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

[0112] Under this example (for the PPG application), the receiver 150-a is further configured to receive, based at least in part on the illuminated green light 170, light pulses reflected off the skin tissue, whereby the reflected light pulses vary temporally in intensity being responsive to changes in blood volume in a microvascular bed under the skin tissue. The reflected light pulses are the optical pulse signals 160 and the spatiotemporal characteristics is associated with atrial fibrillation (AF). This example application is to be discussed in further details in the section “Atrial Fibrillation Detection Through PPG” of the present disclosure.

[0113] Notwithstanding, it is to be appreciated that the scope of potential applications for the PVRC system 150 is not limited to PPG. Indeed, the PVRC system 150 may also be deployed in applications, where the input may be converted (or transformed) into light pulses, such as including gesture recognition, computer vision, communication, encryption, LiDAR (i.e. Light Detection and Ranging), biosignal processing or the like.

[0114] In connection, FIG. 1(g) is a flow diagram of a method 180 of operating the PVRC system 150 of FIG. 1(f), in accordance with aspects of the present disclosure. At 180-a, the method 180 may comprise: receiving, by the receiver 150-a, optical pulse signals 160associated with spatiotemporal characteristics of an entity, said optical pulse signals 160 characterised by variations at least in intensity due to changes in the spatiotemporal characteristics.

[0115] At 180-b, the method 180 may further comprise: providing, by the photodetector 150-b by way of in-sensor reservoir computing based upon said received optical pulse signals 160, corresponding unique photovoltages as output to enable determination of the spatiotemporal characteristics, wherein the unique voltages are respectively stateful representations in time for the spatiotemporal characteristics.

[0116] An advantage worth noting is the read-out variable in the proposed PVRC system 150 is in the form of photovoltage, as opposed to the conventional photocurrent. This may immensely facilitate integration with peripheral circuits without needing a current-to-voltage conversion circuits that dissipate a large fraction of energy in current memristive Al hardware [ref. 32-33], Additionally, as discussed, the read operation in the proposed PVRC system 150 consumes minimal (e.g. < 1 fW) power, since the voltage (e.g. Voc< 1 V) is measured in an open circuit condition (e.g. leakage current ~ 100 fA) This originates from the self-powered sensing and computing of optical inputs enabled by the photovoltaic device architecture.• DEVICE STACK OF PROPOSED PHOTOVOLTAIC CELL

[0117] Exemplary, non-limiting embodiments of a photovoltaic cell are set out below. In some aspects, the photovoltaic cell may be the perovskite photovoltaic 110 as afore discussed.

[0118] FIG. 1(h) is a schematic view of the perovskite photovoltaic 110 of FIG. 1(c), in accordance with aspects of the present disclosure The perovskite photovoltaic 110 (has a device stack that) comprises: a first layer 110-a comprising a substrate arranged to include a transparent conducting material, a second layer 110-b configured as a hole transport layer, a third layer 110-c configured as a semiconductor layer comprising a halide perovskite, a fourth layer 110-d configured as an electron transport layer, and a fifth layer 110-e comprising a conductive contact.

[0119] The transparent conductive material of the first layer 110-a may comprise a bottom electric contact configured to connect to one or more peripheral circuits. The transparent conductive material may be able to transmit light. The transparent conductive material may have high conductivity. The transparent conductive material may comprise one or more ofindium tin oxide (ITO), fluorinated tin oxide (FTO), silver nanowires, carbon nanotubes, metal mesh, aluminium zinc oxide, gallium zinc oxide or indium zinc oxide.

[0120] The substrate may be a silicon wafer, glass, stainless steel or polymer. In an example, the transparent conductive material may be contacted or coated onto (a top surface of) the substrate. The substrate may be an ITO-coated glass or a FTO-coated glass. That is to say, the transparent conducting material coated on the substrate may include ITO, FTO or the like, as above set out. This is not to be construed as limiting; in other examples, the substrate may instead be formed with a portion thereof to include the transparent conductive material (e g. the transparent conductive material may partially be embedded in the substrate in one implementation).

[0121] The hole transport layer of the second layer 110-b may be an electrically conductive layer. This layer 110-b may have selectivity for holes. The second layer 110-b may comprise Spiro-OMeTAD, poly[bis(4-phenyl)(2,4,6-trimethylphenyl)amine]) (PTAA), Poly(3- hexylthiophene) (P3HT), Tris(4-carbazoyl-9-ylphenyl)amine (TCTA), nickel oxide, copper iodide, copper thiocyanate, molybdenum oxide, vanadium oxide or poly(3,4- ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS).

[0122] The semiconductor layer of the third layer 110-c may be light sensitive (electrically measurable response to light). The semiconductor layer may have short-term memory. The mechanism to impart memory may be through mobile ions, charge trapping, mobile vacancies or persistent photoconductivity. The semiconductor may be an organic semiconductor. The semiconductor may be a dye-sensitive semiconductor. The semiconductor may comprise one or more of light-absorbing dyes, 2D materials, metal oxides, transition metal dichalcogenides, quantum dots, antiperovskites and perovskites The light-absorbing dyes may be organic, metalloorganic or biologically-derived.

[0123] The perovskite may be a halide perovskite, double halide perovskite, or an inverse perovskite. The halide perovskite may have the formula ABX3, where: “A” is selected from a monovalent metal cation or an organic cation; “B” is a metal cation; and “X” is a halide ion. “A” may be selected from Cs, CH3NH3+or [HC(NHz)2]_, “B” may be selected from Pb or Sn, and “X” may be a halide, such as Cl, Br or I.

[0124] The semiconductor layer may have a thickness of between about 150 nm to about 650 nm. The thickness may be between about 180 nm to about 600 nm, or about 180 nm toabout 200 nm, or about 200 nm to about 220 nm, or about 220 nrn to about 240 nm, or about 240 nm to about 260 nm, or about 260 nm to about 280 nm, or about 280 nm to about 300 nm, or about 300 nm to about 350 nm, or about 350 nm to about 400 nm, or about 400 nm to about 450 nm, or about 450 nm to about 500 nm, or about 550 nm to about 600 nm, or about 180 nm to about 600 nm, or about 200 nm to about 550 nm, or about 220 nm to about 500 nm, or about 240 nm to about 500 nm, or about 260 nm to about 450 nm, or about 280 nm to about 400 nm, or about 300 nm to about 350 nm.

[0125] The electron transport layer of the fourth layer 110-d may have a selectivity for electrons. The fourth layer 110-d may comprise one or more of titanium oxide, zinc oxide, stannic oxide, tungsten trioxide, 2,2',2"-(l,3,5-benzinetriyl)-tris(l-phenyl-l-H-benzimidazole) (TBPi), C60 and ph enyl-C61 -butyric acid methyl ester (PCBM)

[0126] The conductive contact of the fifth layer 110-e may comprise a top electric contact configured to connect to one or more peripheral circuits. The conductive contact may have high conductivity. The fifth layer 110-e may comprise one or more of lithium fluoride (LiF), calcium fluoride (CaF2), titanium oxide (TiCh), aluminium (Al), silver (Ag), copper (Cu), titanium (Ti), or gold (Au). The fifth layer 110-e may comprise a dual layer of lithium fluoride and aluminium, wherein the lithium fluoride is between the fourth layer and the aluminium.

[0127] The lithium fluoride layer may have a thickness of between about 0.1 nm to about 10 nm, or about 0.2 nm to about 8 nm, or about 0.3 nm to about 5 nm, or about 0.5 nm to about 2 nm, to about 0.6 nm to about 1 nm, or about 0.7 nm to about 0.9 nm. The aluminium layer may have a thickness of between about 10 nm to about 1000 nm, or about 20 nm to about 800 nm, or about 30 nm to about 500 nm, or about 50 nm to about 200 nm, or about 70 nm to about 120 nm, or about 90 nm to about 110 nm.

[0128] The photovoltaic cell may comprise: a first layer comprising an indium tin oxide (ITO)-coated glass substrate, a second layer comprising poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS), a third layer comprising CH^NHsPbBrS or CHsNEEPbE, a fourth layer comprising phenyl-C61 -butyric acid methyl ester (PCBM), and a fifth layer comprising lithium fluoride and aluminium.

[0129] Exemplary, non-limiting embodiments of a method of preparing the photovoltaic cell are set out below.

[0130] In this regard, FIG. l(i) is a flow diagram of a method 190 of preparing a photovoltaic cell (which may be the perovskite photovoltaic 110), in accordance with aspects of the present disclosure. The method 190 may comprise the following steps of:• at 190-a, (a) applying a second layer 110-b to a first layer 110-a, wherein the first layer 110-a comprises a substrate arranged to include a transparent conducting material and the second layer 110-b is configured as a hole transport layer;• at 190-b, (b) applying a third layer 110-c to the second layer 110-b, wherein the third layer 110-c comprises a halide perovskite;• at 190-c, (c) applying a fourth layer 110-d to the third layer 110-c, wherein the fourth layer 110-d is configured as an electron transport layer; and• at 190-d, (d) applying a fifth layer 110-e to the fourth layer 110-d, wherein the fifth layer 110-e comprises a conductive contact.

[0131] Step (a) may further comprise, before step (a), the step of: (al) cleaning the first layer 110-a. The cleaning step may include cleaning with one or more of detergent solution, alcohol and water, any of which may be combined with ultrasonication. Step (al) may further comprise the step of: (a2) treating the cleaned first layer 1 10-a using UV-ozone.

[0132] Step (b) may further comprise one or both of solvent annealing and heat treatment in an inert environment after applying the third layer 110-c. The solvent annealing step may use DMF. Step (b) may further comprise solvent annealing after applying the third layer 110- c. The method 190 may further comprise the step of: (e) encapsulating the photovoltaic cell with epoxy resin and glass. The method 190 may further comprise the step of: (e) encapsulating the photovoltaic cell with epoxy resin, glass or a combination thereof.

[0133] Additional technical details regarding various aspects vis-a-vis the method 190 of preparing the photovoltaic cell are set out in the “Methods” section of the present disclosure.• RESULTS10134] Neuromorphic devices with reservoir properties possess transient memory and nonlinearity. These dynamics may be conceived in the slow open circuit photovoltage as a synaptic memory trace. In an aspect, the perovskite photovoltaic 110 may be fabricated in an inverted p-i-n architecture configured with at least five layers - refer to FIG 1 (c), FIG. 1 (h), and FIG.l(i). As explained, the first layer 110-a may be an Indium Tin Oxide (ITO)-coated glass substrate, the second layer 110-b may be a (spin-coated) hole transport layer formed of Poly (3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS), the third layer 110-c may be a semiconductor (absorber) layer formed of CH3NH3PbBr3(in which a wide bandgap halide perovskite is utilised), the fourth layer 110-d may be a (spin-coated) electron transport layer formed of ph enyl-C61 -butyric acid methyl ester (PCBM), and the fifth layer 1 10-e may comprise at least one conductive contact (e g. formed of LiF and / or Al). The conductive contact per se is an electrode. It is to be appreciate that LiF is commonly used to protect the underlying layers from Al evaporation, and to also improve carrier transport.|0135| In an example, it is to be appreciated that if the first layer 110-a is viewed as the bottommost layer, and the fifth layer 110-e as the topmost layer, the second to the fourth layers 110-b, 110-c, 110-d are then arranged sequentially and intermediate between the first and fifth layers 1 10-a, 1 10-e, whereby the second layer 1 10-b is formed on the first layer 1 10-a, and the third layer 110-c is arranged sandwiched between the second and fourth layers 110-b, 110-d. The fourth layer 1 10-d is arranged intermediate the third layer 1 10-c and the fifth layer 1 10-e. So, if viewed in this envisaged order, the conductive contact (at the fifth layer 110-e) may be considered as a top electrode.

[0136] The band gap of the CH3NH3PbBr3thin film (arranged in the third layer 110-c) is calculated to be approximately 2.32 eV (i.e. refer to FIG. 3(b)). FIG. 3(b) is a (Tauc) plot 305 of spin-coated MAPbBr3thin films of three different thicknesses (e.g. being of thicknesses 230 nm, 360 nm, and 560 nm respectively, under three concentrations of IM, 1.4M, and 1.8M). For avoidance of doubt, MAPbBr3is methylammonium bromide-based.

[0137] Voltage sweep is a standard method of capturing the relevant solar cell parameters. However, halide perovskites are reported to have intricate physical phenomena in timescales ranging from femtoseconds to hundreds of seconds [ref. 34-35], These underlying processes which are manifested in phenomena such as light soaking effects, voltage sweep hysteresis, and slow open circuit voltage decay (OCVD) may be utilised to establish the paradigm for their integration in neuromorphic computing [ref. 36], It is hereby disclosed that the transient Vocis illustrated as one of the measurable physical quantities that may be used to enable computation of spatiotemporal information.- Short-Term Plasticity in Transient Voc

[0138] The spatiotemporal dynamics of the Vocare studied by stimulating with light pulses(e.g. wavelength: 525 nm; intensity: 4.3 mWcm’2; pulse width: 1 ms; duty cycle: 50 %). Photosensors encounter optical signals of various frequencies and intensities. Therefore, the influence of varying pulsing parameters is studied. Changing the pulsing frequency modulates the rate of potentiation and saturation, thus capturing the temporal information of the stimuli (i.e. refer to FIG. 2(a)). FIG. 2(a) is a plot 200 depicting temporal correlation of the Vocobserved by stimulating the PVRC system 150 with different frequencies of light pulses.

[0139] The evolution of Vocwith consecutive light pulses (e.g. frequency: 1 kHz) may also depend on light intensity, as shown in FIG. 2(b). FIG. 2(b) is a plot 205 depicting the evolution of the Pocwith light pulses of different intensities. This implies that light pulses with high frequency and high intensity may seem to produce the greatest influence on the Voc. Specifically, high frequency or intensity results in a larger photon flux being absorbed, hereby increasing the carrier generation rate and producing a faster build-up of the Voc. This suggests that apart from the temporal, the PVRC system 150 may also respond to analog parameters of optical stimuli.

[0140] The Vocshows short-term plasticity through paired-pulse facilitation (PPF), i.e. two consecutive identical light pulses (e.g. wavelength: 525 nm; intensity: 4.3 mWcm'2; pulse width: 1 ms; duty cycle: 50 %) closely spaced in time result in a higher peak Vocthan the corresponding generated by a single light pulse (i.e. refer to FIG. 2(c)). FIG. 2(c) is a plot 210 depicting state-dependent response to light pulses shown through PPF In the plot 210, it is highlighted that the peak response reached with the second light pulse is higher than that of the first light pulse, and that is indicative of remnant history (short-term plasticity) in the (transient) Voc. After the second pulse, it is observed that the Vocstarts decaying towards zero (i.e. represented by a dark horizontal equilibrium line in plot 210), with no permanent change. As discussed, this is an indication of short-term memory, implying that the response to a light pulse is temporarily state and history-dependent.

[0141] It is to be appreciated that a PPF index may be calculated by taking the ratio of the response to the second light pulse and the first light pulse. The PPF index is computed to be at 120.3 % for a 5 ms time difference between the light pulses, and it decreases exponentially with increasing pulse timing (At) (i.e. refer to FIG. 2(d)). FIG. 2(d) is a plot 215 depicting PPFmeasured for varying At. In the plot 215, the data points are fitted to a double-exponential decay curve to calculate the time constants. Such a decay also shows a non-linear dependence on the pulse timing. This suggests that more recently received information may produce a more significant effect on the internal state of the PVRC system 150. The robustness of the PVRC system 150 is verified by supplying two consecutive light pulses to the PVRC system 150 for 1000 cycles, and the evaluation results are shown in FIGs. 4(a)-4(b). FIGs. 4(a)-4(b) are respective plots 400, 405 depicting minimal cycle-to-cycle variation in the PPF indices, and observation of a standard deviation of 0.24 % in 1000 cycles, in accordance with aspects of the present disclosure.

[0142] Natural input streams embed valuable information in different timescales. For applications such as in-sensor motion perception, the temporal resolution of the sensing and computing is to be suitably adjusted to optimise the recognition accuracy vs computation efficiency trade-off [ref. 37], A general time series may be viewed as a combination of signals and noise of different frequencies; the ideal temporal resolution to process is preferably identified based on the task, i.e., the salient frequency range of the time-series signal. At the device level, being able to engineer the timescale of dynamics may help design applicationspecific in-sensor computational units. For example, a system designed to perceive moving vehicles (e.g. > 10 m / s) in a visual field may benefit from in-sensor filtering of slower-moving objects such as pedestrians (e g. moving at 1-2 m / s). The PVRC system 150, in this regard, provides straightforward tunability in kinetics by varying a thickness of the perovskite layer (i.e. refer to FIG. 2(f)). FIG. 2(f) is a (semi -logarithmic) plot 225 depicting PPF index of the photodetector 150-b (that includes the perovskite photovoltaic 110) arranged to be of three different thicknesses shown as a function of pulse timing. In particular, the plot 225 highlights that due to the stark difference in the volatility, and only a particular thickness may be considered as suitable for sampling signals in the different time scales.

[0143] Halide perovskite films of different thicknesses are obtained by changing the precursor concentration. In an example, for evaluation purposes, spin-coating three concentrations of a precursor solution - IM, 1.4M, and 1.8M - is performed to obtain respective thicknesses of 230 nm, 360 nm, and 560 nm for the perovskite layer in the perovskite photovoltaic 110 utilised by the PVRC system 150 (i.e. refer to FIG. (2e)). FIG. 2(e) is a plot 220 depicting response to a single light pulse shown for the PVRC system 150 configured with said different thicknesses for the perovskite layer. Comparing the PPF index in differenttimescales, an assessment is then performed on the suitability of the different thicknesses of the perovskite layer for a certain temporal window

[0144] FIGs. 5(a)-5(c) are respective plots 500, 505, 510 relating to PPF index measured at three different timescales with light pulses of durations at 0.1 ms, 1 ms, and 10 ms, in accordance with aspects of the present disclosure. It is to be appreciated that the At between the pulses is varied. A steeper change in PPF indices with varying At indicates better expression and resolution of temporal features in that timescale (FIG. 2(f)) . In the order of 10"4s, it is observed the PVRC system 150, when configured with the perovskite layer of thickness 230 nm (i.e. obtained based on the precursor solution of concentration of IM - hereafter termed “IM device”), possesses a stronger correlation with pulse timing, whereas the PVRC system 150, when configured with perovskite layers of 360 nm and 560 nm respectively (i.e. obtained based on the precursor solution of respective concentrations of 1.4M and 1.8M - hereafter termed “1.4M device” and “1.8M device” respectively), changes relatively little. This implies the efficacy of processing the spatiotemporal information in this timescale is the highest for the IM device, amongst the three devices.

[0145] On the other hand, for 10'3s, the IM and 1.8M devices have a monotonous PPF index dependence on pulse timing In comparison, the 1 ,4M device captures the information effectively in this timescale. For the longer 10'2s timescales, the 1.8M device appears to be the best suited, as the other two devices lose pertinent information completely. It is highlighted that such a comparison of the temporal resolution for the perovskite layer of different thicknesses may help in application specific engineering by exploiting the complete dynamic range of the material substrate.- MNIST Handwritten Digit Recognition

[0146] Based on the strategy disclosed by [ref. 38] on using a single non-linear dynamical node with time-delayed feedback to fulfil the prerequisites of a physical reservoir 100b (i.e. refer to FIG. 1(a) (bottom part)), several two-terminal and three-terminal memristors are used to realise physical reservoir computing (RC) [ref. 39-47], Implementing in-sensor computing using physical reservoirs may produce even more efficient systems with tremendous potential for deployment in low-power electronics. In reservoir computing, a time-varying input is to be reliably transformed into a unique higher-dimensional state for classification. In this instance, the rich dynamics present in the PVRC system 150 is highlighted by mapping binaryinput pulse streams of 4-bits into distinct reservoir states - i.e. into 16 combinations of l’s and 0’s (where ‘ 1’ represents an ON state and ‘0’ represents an OFF state) ranging from ‘0000’ to ‘ 1111’.

[0147] In connection, FIGs. 7(a)-7(b) are respective plots 700, 705 of voltage versus time relating to a pulsing scheme of 4-bit light pulses (e.g. at 525 nm and 4.3 mWcm'2) represented for a 4-bit input of ‘1010’ and a 4-bit input of ‘ 1011’. Each bit is a light pulse with a period of 1 ms and a duty cycle of 50%. The voltage for each combination is fetched at the end of 4 light pulses (which collectively represent a 4-bit combination) at t = 4 ms. The voltage readout is labelled as Vn, where ‘n ’ is the decimal value obtained by converting a 4-bit binary value.|0148] In an aspect, the IM device may be used to demonstrate the following features of physical RC in the perovskite photovoltaic 110. A viable reservoir computing system is preferably configured to resolve unique inputs into different Vocreadout values with minor cycle-to-cycle variations. The 16 combinations of the 4-bits as light pulses form 16 different time-varying inputs. The voltage rises, when the IM device receives a ‘ 1’ and falls upon receiving a ‘O’. The trajectory of voltage with number of bits shows the separability of the inputs as consequence of unique patterns of rise and fall periods, as depicted in FIG 6(a). FIG. 6(a) is a plot 600 mapping various combinations of the 4-bit input (i .e. ‘0000’ to ‘ 1 1 1 1 ’) into different reservoir states, in which 16 unique reservoir states are obtained at the end of the 4- bit pulse train, in accordance with aspects of the present disclosure.

[0149] At the end of the-4 bits of input, the Vocis sampled and fetched as the readout for that combination. 16 different reservoir states are observed for the different combinations of input pulse trains (Vo, VltV2, .... F15). The cycle-to-cycle variation of the obtained states is low over 20 cycles with a maximum standard deviation of 1.8 %, as shown by FIGs. 8(a)-8(b). Specifically, FIG. 8(a) depicts a box plot 800 of 20 values obtained for the same input, and FIG. 8(b) depicts a collection 805 of plots pertaining to 20 transient evolutions of Vocoverlapped for 15 unique inputs. Accordingly, the results are a testament to the reliability and robustness of the PVRC system 150 being capable of high-dimensional mapping of temporal inputs with fidelity. The above procedure is repeated for 10 example devices (of the IM device) showing high device-to-device variation in the mapping of 4-bit inputs with a minimum standard deviation of 29.7 % in the voltage level for the 4-bit inputs among said 10 devices (i.e. see FIG. 9, where the x-axis labels represent the 4-bit binary inputs converted todecimal format). However, the variation may be exploited to expand the reservoir state by supplying the same input to different devices [ref. 48],

[0150] In an example, with reference to a scheme 605 illustrated in FIG. 6(b), by using the 4-bit mapping as a pre-processing scheme, the proposed PVRC system 150 is employed in image classification to further evaluate the computational viability, by compressing the input image and yet retaining the recognition performance. For this purpose, images from the MNIST (i.e. Modified National Institute of Standards and Technology) dataset for handwritten digit classification are first transformed to exploit the temporal dynamics of the photovoltaic reservoir. The images containing 28*28 pixels are first binarized. Next, the binarized images of the 28*28 pixel array are segmented into 28 rows, each row of 28 pixels. Each row is further divided into 7 sets of 4 pixels, where each set belongs to one of the 16 possible combinations of 4-bits.

[0151] Using the PVRC mapping of these combinations, each set is assigned a corresponding voltage valueInthis manner, 4 pixels are reduced to one value. Hence, an image represented as a 28*28 matrix (i.e. 784 elements) may be compressed to a 1* 196 matrix (i.e. 196 elements) - refer to FIG. 6(b) which shows the described scheme 605 of using the 4-bit reservoir states to pre-process a binarized MNIST image. As a result of this pre-processing, the complexity of the two-dimensional input vector is reduced to a onedimensional vector with minimal loss in information. Reduction in the input size implies fewer parameters are required in the neural network, and thus training is performed at a lower expense of computational resources. This readout is used as the input to a basic single-layer feedforward neural network (FNN) for supervised learning to allow the reservoir preprocessing to bear most of the load of the recognition performance. The discussion of the NN is set out in the “Methods" section below.

[0152] Further, the performance of the PVRC system 150 is compared to an optimised artificial neural network (ANN) architecture 610 for 20 epochs - refer to FIG. 6(c). FIG. 6(d) is a plot 615 relating to recognition accuracy of a PVRC pre-processed neural network and a typical artificial neural network, in accordance with aspects of the present disclosure. The PVRC pre-processed neural network may achieve comparable accuracies as the ANN (e.g. 96.35% vs 98%), while utilising far fewer parameters (e.g. 192,080 vs 1,003,520). Hence, the image recognition task highlights the computational efficacy of the PVRC system 150 through a robust nonlinear mapping of various temporal inputs into unique reservoir states.- Second-Order Non-Linear Dynamic Equation Prediction

[0153] As discussed in the preceding section, static images are pre-processed into pseudo time-series for the task. To further demonstrate the essence of the PVRC system 150 for true time-series analysis, the Nonlinear Autoregressive Moving Average (NARMA2) task is adopted for evaluation [ref. 49], NARMA2 task is a commonly used benchmark in reservoir computing.

[0154] FIG. 10(a), in its entirety, depicts a schematic workflow 1000 of a NARMA2 task. The task involves predicting the future value of a time-series based on its past values. The volatile and nonlinear features may be exploited to solve, and predict a second order nonlinear dynamic equation. The target values y(k) of the function are set out in equation (1):

[0155] y k + 1) = 0.4y(fc) + 0.4y(k)y(k — 1) + 0.6u3(k) + 0.1 (1)]0156] wherein k represents the time step (e.g. in a range of: 0 to 500), u(k) represents the input at time step k (e.g. in a range of: 0.0 to 0.5), and y(k represents the output at time step k.

[0157] Equation (1) suggests the current target value may also depend on the previous two timesteps. Although the relationship between the inputs and outputs is stated explicitly in this case, such a time-series, where the function is implicit but shows a similar state dependence, may also be found in electrical, mechanical, control, and other engineering fields [ref. 50], The short-term memory of the PVRC system 150 contributes to learning and predicting such functions. A random set of 500 input values u(k) ranging between 0.0 and 0.5 is generated (i.e. see FIG. 10(a)). These generated values are then transmitted to the PVRC system 150 in the form of a light pulse train, where the analog values are encoded into light pulse intensity of the light pulse train (see FIG. 10(a)), and the Vocis measured against time (see FIG. 11(a)). FIG. 11(a) is a plot 1100 depicting the transient evolution of measured voltage, between timesteps of 350 to 400, as the randomly generated inputs are supplied as the light pulse train to the PVRC system 150, in accordance with aspects of the present disclosure.

[0158] The encoding of inputs into intensity levels is performed using a linear mapping function (see FIG. 11(b)). FIG. 11(b) is a plot 1110 of input values (u) in NARMA2 encoded into corresponding light pulse intensities. The linear mapping is crucial to ensure that the nonlinearity in the reservoir output originates only from the device response.

[0159] To generate the reservoir output, the input is transmitted to 6 different example devices of the PVRC system 150 The voltage is sampled 5 times per timestep to expand the reservoir by adding virtual nodes (see FIG. 10(a), where the voltage outputs from 2 devices are shown). The 5 samples from each device make a row of the reservoir state From the 6 devices, a 6x5 vector is obtained to define the reservoir state of that timestep. The voltage outputs from the 6 devices are shown for 500 timesteps. This results in a reservoir size of 30. The device- to-device variation in the response to these pulse trains inherently enhances the separability of the reservoir states by expansion of the reservoir size (i.e. see plot 1200 in FIG. 12).

[0160] The 500 timesteps are then divided into 450 training and 50 testing input values. The description of the linear regression model is provided in the “Methods” section. The target values for the training phasetesting phaseare obtained by plugging in the random inputs (u) into equation (1). For the first 450 timesteps, the reservoir experimentally solves and reconstructs the function during the training phase (i.e. refer to plot 1010 in FIG. 10(b)). To show that the PVRC system 150 may predict the function upon learning, the independently generated 50 input values are solved with a normalized mean squared error (NMSE) of 0.0281 (i.e. refer to plot 1020 in FIG. 10(c)). It is to be appreciated that this evaluation discloses the prediction of second-order non-linear equation through insensor optoelectronic RC. With reference to the literature of physical reservoirs demonstrating the NARMA2 task, the PVRC system 150 may achieve comparable NMSEs as that of a 3- terminal ion-gating reservoir, which utilises a reservoir size that is approximately 5 times larger [ref. 45],- Atrial Fibrillation Detection through PPG

[0161] As discussed, the benchmarking tests show the PVRC system 150 is a robust and viable system to process time series data. However, the superiority of in-sensor learning is from the ability to sense and compute optical inputs in a single step. In an example, the PVRC system 150 may be utilised for biomedical applications by integrating said system 150 into a physiological diagnostic method known as photoplethysmography (PPG). PPG signals are initially obtained by pulsing (via the light illuminating means 150-e of the PVRC system 150) green light onto (the skin tissues of) a human wrist or fingertip The green light interacts with the skin tissues and gets reflected (which are received by the receiver 150-a of the PVRC system 150).

[0162] The intensity of the reflected light varies and depends on the blood volume in a microvascular bed under the skin tissues As the heart pumps blood, the volume of the vessels pulsates with the same frequency. This pulsation is imprinted in the intensity of the reflected light, which is measured using the photodetector 150-b of the PVRC system 150 The PPG signal may thus be used to assess the various stages of the heartbeat. Thus, the normal sinus rhythm gives a specific waveform to the PPG signal, which is relatively uniform. Any deviation from normalcy in the PPG signal may indicate anomalous heart conditions, thus establishing PPG as a non-invasive, peripheral method to detect chronic heart diseases.

[0163] PPG is commonly incorporated into modern wearables such as smartwatches or fitness trackers, but since the computational and power resources in wearables are typically limited or constraint, the PPG signals are normally processed remotely in the cloud, or by a connected local device such as a smartphone [ref. 51], This wireless communication however leads to significant power consumption and latency. Performing computations in-sensor may thus help reduce reliance on remote processing, thereby enabling efficient health monitoring systems [ref. 4],

[0164] In this example, the PPG signals are pre-processed to detect atrial fibrillation (AF), which is a common cardiac arrythmia characterised by sporadically irregular heart rate and is associated with an increased risk of cardiac arrest [ref 52], However, AF may remain asymptomatic or undetected in intermittent clinical check-ups [ref. 53], PPG signals are considered a time series, in which AF may be identified in varying peak-to-peak intervals and pulse contours. As discussed, the PVRC system 150 is suitable to process such time-varying optical signals. Unlike a typical photosensor, the state-dependent response of the PVRC system 150 helps to capture the history of the optical pulses received (i.e. the reflected light). Any variability in the pulse rate is encoded in the reservoir output. Hence, to supplement the intensity variation of the reflected light, the response of the PVRC system 150 is also used to embed additional features of the PPG signals.

[0165] It is to be appreciated that in the context of the PVRC system 150, a light pulse of a specified intensity received (by the PVRC system 150) at different time, and C (wherein t2 > ti) causes different responses to be output by the PVRC system 150 respectively at ti and C, because (the history of) the internal state of the PVRC system 150 has since changed, due to, assuming the PVRC system 150 has already received other light pulses in the periodintervening ti and i; (thereby altering the internal state of the PVRC system 150 in between said period spanning ti and A).

[0166] Derivative information, e.g. heart rate variation, may easily be extracted by employing a simple feedforward neural network. Conventionally, deep neural network architectures, e.g. CNN, RNN or the like, are required to detect AF in PPG signals [ref. 53], These conventional methods demand high computational resources, which is scarce in edge devices, such as the wearables. The in-sensor transformation performed by the PVRC system 150 resolves the PPG signals from which the AF incidences may be facilely identified.

[0167] 15-second windows of PPG signals from an AF labelled dataset [ref. 54-55] are considered - refer to FIGs. 14(a)-14(b) which depict respective example plots 1400, 1405 of PPG signals present in the MIMIC PERform database. The intensity variation is normalized to 0-140 mWcm'2. The raw PPG signals in the dataset are a time series of green light intensity reflected from patients’ blood vessels sampled at 125 Hz. This light intensity is generated by a programmable LED source and illuminated onto the PVRC system 150, and the Vocis measured against time (see FIG 13(a)).

[0168] In particular, FIG. 13(a)(i) shows a schematic representation of generation of the PPG signals by the reflection of green light from an artery, FIG. 13(a)(ii) shows comparison of the PPG signals generated from a healthy heart (NSR) and a diseased heart (AF), and FIG. 13(a)(iii) shows PPG signals taken from the MIMIC PERform dataset and provided to the PVRC system 150 for pre-processing, in accordance with aspects of the present disclosure.

[0169] As performed under the NARMA2 task, for each timestep (e.g. 8 ms long, since the sampling frequency is 125 Hz), the voltage is sampled 5 times. The reservoir readout is represented by an Nx5 vector, where N is the total number of timesteps in the signal (e.g. N = 1825 in this case) (refer to FIG. 13(b)). FIG. 13(b) depicts a sampling protocol to obtain the reservoir state as the N><5 input vector for classification using a feed-forward neural network (FNN). To highlight the value added by the in-sensor reservoir preprocessing, classification efficacy from the output of the PVRC system 150 and the original raw data (without PVRC processing) are compared. The FNN is used to circumvent the need for complex statistical techniques and deep neural networks that are usually applied to process PPG signals. The description of the classification algorithm is set out below in the “Methods” section.

[0170] A receiver operating characteristic (ROC) curve is used to depict the performance of binary (two possible outcomes) classifier. The curve is obtained by plotting the true positive rates and false positive rates for different decision thresholds. Specifically, the area under the (ROC) curve (AUC) is considered the key metric. It is to be appreciated that AUC is the ratio of True Positives to False Positives, and it conveys how well a classifier may separate the signals in the data from noise. A classifier with AUC equal to 1 may perfectly separate positive and negative classes, and if AUC is equal to 0, the classifier may predict all positives as negatives and vice-versa, whereas if AUC is equal to 0.5, the classifier may not distinguish the classes and predicts a class at random.

[0171] FIG. 13(c) depicts ROC 1300 that show respective areas under a ROC curve vis-a- vis scenarios with, and without using the PVRC system 1500 for the classifier. From the ROC for classification of AF signals in the dataset, it is observed that the AUC is larger for a classifier that involves use of the PVRC system 150 (i.e. at 0.833) compared to not using the PVRC system 150 (i.e. at 0.733). Although the AUC for the classifier, based on the case of using the PVRC system 150, is only 0.833, it is to be appreciated that the emphasis is on the improvement over the AUC of the alternative that does not involve use of the PVRC system 150. The AUC covers a larger area, given the limited number of training and test samples, signifying that the model based on the PVRC system 150 may distinguish between the AF and NSR classes better.|0172| The classification performance for the test dataset may also be represented in confusion matrices. The confusion matrix summarizes the classifier's performance, especially when the number of training samples per class is unbalanced. The goal is to minimize the counts outside the main diagonal FIG. 13(d) depicts confusion matrices 1 10a, 1310b that show classification performance for 11 test cases, in accordance with aspects of the present disclosure. From the confusion matrices 1310a, 1310b, it is evident the PVRC system 150 beneficially helps to increase the true positive and true negative classification.

[0173] In an aspect, to further emphasise the superior performance of the PVRC system 150, principal component analysis (PC A) is performed on the pre-processed and raw inputs to find the low dimensional representations of the data used for the classification. It may be shown that the high dimensional distribution of data is converted to a lower dimension by the PVRC system 150, i.e. refer to FIGs. 15(a)-15(b).

[0174] FIGs. 15(a)-15(b) depict respective example plots 1500, 1505 of low dimensional representations of the data used for the classification It is to be appreciated that the principal components using the reservoir are much smaller, suggesting that the reservoir / material may be able to compress the patterns and create smaller, uncorrelated features. With smaller features, the training time may also be decreased. Advantageously, this dimensionality reduction makes the data less sparse and more manageable for a model.- Influence of Ionic Migration in Halide Perovskites on the Short-Term Memory

[0175] The coupled ionic-electronic dynamics are the linchpin of the state-dependent response and memory in the transient Voc[ref. 56], To empirically visualise this principle, the CH3NH3PbBr3perovskite is compared against identically processed CH3NH3PbI3in the same device architecture. First, the degree of ion migration between the two systems is observed. For this, a lateral device with symmetrical electrodes is fabricated. Gold electrodes with a 15 pm gap are patterned on SiC>2 / Si substrates using photolithography, as shown in FIG. 16(a), onto which the two perovskite samples are spin coated. To trigger ionic migration, an electric field of 2.5 V / pm is applied for 600s across the lateral Au / perovskite / Au device, as shown in FIG. 16(b). The magnitude is selected to mimic the electric field generated by a Vocof 0.5 V of a solar cell with 200 nm thick perovskite layer.

[0176] To quantify the extent of ion migration, the concentration of the halide species, a key mobile ion, is measured across the width of the device using energy dispersive X-ray (EDX) line scan, as shown in FIG. 18. Under this electric poling condition, the negatively charged halide ions drift towards the positive terminal to form a concentration gradient across the profile. A steeper concentration gradient indicates greater ion migration. Whereas, in a pristine unpoled device, the halide ions are uniformly dispersed, and no concentration gradient is observed. Between the two samples in consideration, the concentration gradient of the halide species is steeper for CH3NH3PbBr3, as shown in FIG. 16(c), than CH3NH3PbI3as shown in FIG. 16(d). Hence, under the typical electric field applied by the Voc, CH3NH3PbBr3shows more significant ion migration.

[0177] This difference between CH3NH3PbBr3and CH3NH3PbI3is further highlighted by measuring the switchable photovoltaic effect in lateral devices. Switchable photovoltaic effect is a phenomenon attributed to the self-doping of halide perovskites due to the bulk migration of ions under electric poling [ref. 57], Obtaining a Vocin a symmetrical (withoutcarrier selective transport layers) device is an experimental demonstration of ionic migration in halide perovskites. Here, the two perovskites are put under the same poling conditions (2.5 V / pm for 600s). The Vocof the pole devices is extracted by a voltage sweep under 10 mWcm"2illumination (e.g. wavelength: 525 nm for CH3NH3PbBr3and 623 nm for CH3NH3PbI3). The Vocgenerated in CH3NH3PbBr3is more than twice that of CH3NH3PbI3indicating greater ionic activity in the former, as shown in FIG. 16(e). To extrapolate this observation to photovoltaic devices, the two perovskites are compared in an identical vertical configuration (ITO / PEDOT:PSS / perovskite (1.8M) / PCBM / LiF / Al). Upon stimulating with a light pulse train, only the transient Vocof CH3NH3PbBr3device shows short-term potentiation, as shown in FIG. 16(f). The presence of short-term memory in CH3NH3PbBr3may further be highlighted by showing the gradual response of the transient Vocto illumination and dark conditions, as shown in FIG. 19(a). This corroborates the relation established between coupled ionic-electronic dynamics and short-term memory in halide perovskite photovoltai cs.- Coupled Ionic-Electronic Dynamics in the Temporal Evolution ofV0C

[0178] The temporal resolution achieved with the different device formulations may be understood by the coupled ionic-electronic processes (including drift and diffusion of charged species and recombination of electrons and holes) across the thickness of perovskite layer. Nonlinearity and volatility are the keystones of implementation in the Reservoir computing (RC) framework. Halide perovskite materials are rich in dynamics in that manner. Due to the mobile ionic defects / species (commonly reported with a range 1015to 1019 / cm3), halide perovskite thin films may be understood as a semiconductor that is not homogeneous and isotropic under standard operating conditions. Consequently, the electric field in perovskite materials is highly dependent on the ionic charge distribution and dominated by high densities of relatively slow-moving mobile ionic defects. These devices tend to generate high-order memory or hysteretic effects in timescales from microseconds to seconds in the optoelectronic response. A complex interplay of ion migration and carrier generation-recombination dynamics may result in the unique phenomena highlighted in the present disclosure.

[0179] The theoretical Vocunder illumination is closely related to the difference between the Fermi energy levels of electrons (EPn) and holes (EFP), known as the quasi -Fermi level splitting (QFLS = EPn- EPp). The splitting occurs due to the photoinduced population of the electrons in the conduction band and the holes they leave behind in the valence band. Consequently, an increased non-radiative recombination pathway in the bulk or interfacewould limit the maximum QFLS, and hence the Voc. The measured value of Vocis lower than the QFLS depending on the carrier selectivity of the transport layers and energy level alignment of the transport layers [ref. 58],

[0180] To illustrate this higher-order effect in the transient open-circuit voltage in the PVRC system 150 of the present disclosure, simulated results showing the dynamical effects are provided using a numerical modelling tool known as Driftfusion [ref. 59], which considers mobile ionic species in addition to electronic carriers (i.e. refer to FIG. 20).

[0181] In the dark equilibrium state, the built-in voltage in the halide perovskite p-i-n device generate an electric field that compels the mobile ionic species to drift accordingly, as shown in FIG. 17(a), state 1. Accordingly, the positive charged species (bromide vacancies, methylammonium, etc) accumulate at the HTL / perovskite interface, as shown in FIG. 17(b). The accumulation of ions actively screens and reduces the effective built-in field. Under illumination, electron-hole pairs are generated and separated by the built-in field, as shown in FIG. 17(a), state 2. Since the ions are relatively less mobile than the electrons and holes, the electronic transport is hindered, and valley-like energy bands could be formed near the HTL- perovskite interface, as shown in FIG. 1 (c). These valleys are conducive to recombination of photogenerated carriers, thus the instantaneous Vocafter illumination is low, as shown in FIG. 17(d).

[0182] The splitting of the quasifermi levels means that under the reduced built-in field, there will be diffusion of the ionic accumulation at the interface, as shown in FIG. 17(e). The inhomogeneity in charge distribution is reduced, resulting in Vocequilibration. The combination of the reduced field due to ionic screening and accelerated recombination due to ionic accumulation leads to rich dynamics in the slow equilibrating of Voc. After the illumination is removed, charges stored across the transport layer interfaces recombine. Due to the relatively slower ions, charge accumulation occurs at a longer timescale. Interestingly, when a second light pulse arrives before the charge is fully accumulated, the electron hole experiences a reduced ionic screening effect and recombination rate, thus leading to a higher / lower photovoltage gain. Consequently, non-linear gain may be generated in this manner.- Summary

[0183] The present disclosure sets out a distinctive approach of realising an optoelectronic self-powered in-sensor physical reservoir computing (RC). Through a functional interpretation of the transient Vocdecay, suitable dynamics for RC are identified and located in the perovskite photovoltaic 110. Utilising the Vocas the internal state variable is advantageous regarding integration and power consumption in output readouts, since the need for power-hungry current-voltage converters, as well as read voltage pulses may be eliminated. The short-term synaptic behaviour is realised in the measured voltage, in response to example light pulses of 525 nm. The volatility of the memory is modulated by changing the thickness of the perovskite layers and explained by studying the carrier recombination dynamics at the interface and the bulk. The engineering of transient dynamics is essential for task specificity. This engineering control renders the PVRC system 150 versatile for applications in a range of timescales.

[0184] Accordingly, combining RC with in-sensor computing enables efficient edge computing systems. The unique properties of halide perovskite create rich spatiotemporal dynamics for in-sensor RC of optical stimuli. By harnessing energy harvesting and coupled ionic-electronic dynamics of halide perovskite, the transient open circuit voltage of a methyl ammonium bromide-based photovoltaic may serve as a self-powered short-term memory for optoelectronic RC.

[0185] The example of the image recognition task demonstrates the computing prowess of the proposed PVRC system 150. The PVRC pre-processing performs like an optimised ANN, but with fewer trainable parameters than the latter. As reservoir computing is intended to process time series inputs, its efficacy in solving second-order nonlinear dynamic equations is benchmarked. In this regard, NMSE of 0.0028 is achieved, which is comparable to the current state-of-the-art solutions, even with a smaller reservoir size.

[0186] Additionally, a practical application which unifies sensing and computing capabilities is demonstrated by deploying the PVRC system 150 for PPG to detect anomalous heartbeat signatures. An improvement in the recognition of AF signals is achieved by using a facile neural network architecture. By performing ionic drift-diffusion simulations, the high- order dynamics, conducive for physical RC, are attributed to co-dependent kinetics of interfacial charge accumulation and carrier recombination. The influence of ion migration in the origin of short-term memory in the transient Vocis highlighted by comparing bromide and iodide systems.EXAMPLES

[0187] The following discussions set out some technical details regarding various aspects of the present disclosure, but it is to be understood they are merely examples to enable improved understanding of the subject matter, and hence they are not to be construed as limiting on the protection scope of the subject matter.- Photovoltaic Device Fabrication

[0188] To fabricate the perovskite photovoltaic 110, ITO-deposited glass substrates are first cleaned by ultrasonication in Decon™ 90 soap solution, deionized water, and isopropyl alcohol. Cleaned and dried substrates are then UV-ozone surface treated for 20 minutes. A suspension of PEDOT:PSS (Al 4083) (Clevios; obtained from Heraeus Epurio of Leverkusen, Germany) in water is used to spin coat a thin film, which was then annealed at 200 °C for 5 minutes. Subsequently, IM, 1.4M, and 1.8M solutions of CH3NH3PbBr3are prepared by adding equimolar amounts of methylammonium bromide (CH3NH3Br) and lead bromide (PbBr2) into a 3: 1 mixture of dimethyl formamide (DMF) and dimethyl sulfoxide (DMSO) to form the precursors. The precursors are spin-coated onto the substrate in an Ar glovebox using toluene as the antisolvent. The resultant films are then solvent-annealed for 10 minutes in a DMF environment, and then heated in the Ar glovebox at 100 °C for 10 minutes. A 20 mg / mL solution of PCBM in chlorobenzene was spin-coated and annealed at 100 °C for 10 minutes. Subsequently, the prepared substrates are loaded into a thermal evaporator to deposit 0.8 nm of LiF and 100 nm of Al. The devices are then encapsulated using UV-cured epoxy resin and glass. The same procedure is followed to fabricate CH3NH3PbI3devices, replacing CH3NH3Br and PbBr2with CH3NH3I and Pbl2, respectively.- Lateral Electrodes Fabrication

[0189] Photoresist (AZ 5214E) (obtained from Merck of Darmstadt, Germany) is spin- coated on a clean substrate (Si / Si O2), followed by photolithography using a mask aligner (MJB 4). Subsequently, the substrate is immersed in a developer solution (e.g. AZ developer: DI water = 1: 1 (obtained from Merck of Darmstadt, Germany)) for one minute to obtain the electrode pattern. Next, 5 nm of Cr and 50 nm of Au are deposited by e-beam and thermal evaporation, respectively. After the deposition, the electrodes are obtained by a lift-off process in acetone solution for 2 minutes.- Film Characterization

[0190] For characterization of the film, absorbance spectra are obtained using Shimadzu™ UV-3600 UV-vis-NIR spectrophotometer. Cross-sectional images are obtained using fieldeffect scanning electron microscopy (FE-SEM, JEOL J7600F) EDX elemental line scan and mapping are performed using FE-SEM, JEOL JSM-7800F Prime with an acceleration voltage of 20 kV.- Device Characterization

[0191] To characterise the PVRC system 150, all the measurements are made in ambient conditions in a Semishare™ cryogenic probe station. The transient Vocdecay measurements are made using an Agilent™ Digital Oscilloscope DSOX-3034. Pulsed light measurements are performed using a Keysight™ B2912 source measurement unit (SMU), and a Thorlabs™ DC2200 programmable 525 nm LED as the light source. The open circuit condition is maintained by the input impedance of the SMU (e g. greater than 10 GQ). The LED intensity is calibrated using a Thorlabs™ power meter.- MNIST Handwritten Digit

[0192] The reservoir's readout layer of 10 neurons, corresponding to 10 classes, is trained with varying numbers of neurons in a preceding hidden layer. The variation in validation accuracy is monitored to arrive at the best performing neural network architecture. The neural network (NN) model is initialized with a uniform kernel, and hyperparameter tuning further then improved the best achievable accuracy. It is observed that by keeping the number of neurons in the hidden layer as a multiple of the resized reservoir output (e.g. 784 input pixels grouped 4 at a time to form an input of length 196), the artificial neural network (ANN) may capture the dynamics of the data reasonably well.- NARMA2 Task

[0193] A Linear Regression model is trained with the coefficient of determination, R2, as the regression score function The metric used is NMSE, and a 10-fold cross validation is performed to avoid overfitting on the small dataset used. The reservoir may model non-linear transformations relatively well enough that a linear readout layer is sufficient to predict the second-order variation.-PPG

[0194] The neural network (NN) model training is devised to use the Adam optimizer with a learning rate of IE-3, and a batch size of 6 to balance performance and speed. The architecture of the best performing NN readout layer for this task is designed via a simplistic Neural Architecture Search exploration to be one dense layer of 500 neurons followed by a dropout layer with the dropout rate set to 0.2.

[0195] All of the methods described herein describe possible implementations, and that the operations and the steps may be rearranged or otherwise modified and that other implementations are possible. Further, aspects from two or more of the methods, if applicable, may be combined.

[0196] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Other examples and implementations are within the scope of the disclosure and appended claims. For example, due to the nature of software, functions described herein may be implemented using software executed by a processor, hardware, firmware, hardwiring, or combinations of any of these. Features implementing functions may also be physically located at various positions, including being distributed such that portions of functions are implemented at different physical locations.

[0197] As used herein, including in the claims, “or” as used in a list of items (for example, a list of items prefaced by a phrase such as “at least one of’ or “one or more of’) indicates an inclusive list such that, for example, a list of at least one of A, B, or C means A or B or C or AB or AC or BC or ABC (such as, A and B and C) Also, as used herein, the phrase “based on” shall not be construed as a reference to a closed set of conditions. For example, an example step that is described as “based on condition A” may be based on both a condition A and a condition B without departing from the scope of the present disclosure. In other words, as used herein, the phrase “based on” shall be construed in the same manner as the phrase “based at least in part on”.

[0198] In the appended figures, similar components or features may have the same reference label. Further, various components of the same type may be distinguished by following the reference label by a dash and a second label that distinguishes among the similarcomponents. If just the first reference label is used in the specification, the description is applicable to any one of the similar components having the same first reference label irrespective of the second reference label, or other subsequent reference label.

[0199] The description set forth herein, in connection with the appended drawings, describes example configurations and does not represent all the examples that may be implemented or that are within the scope of the claims. The term “example” used herein means “serving as an example, instance, or illustration,” and not “preferred” or “advantageous over other examples”. The detailed description includes specific details for the purpose of providing an understanding of the described techniques. These techniques, however, may be practiced without these specific details. In some instances, known structures and devices are shown in block diagram form in order to avoid obscuring the concepts of the described examples.INDUSTRIAL APPLICABILITY

[0200] The perovskite photovoltaic 110 and the PVRC system 150 as disclosed herein may find applications in areas such as: gesture recognition, computer vision, communication, encryption, LiDAR, biosignal processing or the like. It will be apparent that various other modifications and adaptations of the present disclosure will be apparent to the person skilled in the art after reading the foregoing disclosure without departing from the spirit and scope of the present disclosure, and it is intended that all such modifications and adaptations come within the scope of the appended claims.REFERENCES[1] Ates, H. C. et al. End-to-end design of wearable sensors. Nat. Rev. Mater. 7, 887-907 (2022).[2] Min, J. et al. An autonomous wearable biosensor powered by a perovskite solar cell. Nat. Electron. 6, 630-641 (2023).[3] Van Doremaele, E. R. W., Ji, X , Rivnay, J & Van De Burgt, Y. A retrainable neuromorphic biosensor for on-chip learning and classification. Nat. Electron. 6, 765-770 (2023).[4] Chatterjee, B., Mohseni, P. & Sen, S. Bioelectronic Sensor Nodes for the Internet of Bodies. Annu. Rev. Biomed. Eng. 25, 101-129 (2023).[5] Koot, M. & Wijnhoven, F. Usage impact on data center electricity needs: A system dynamic forecasting model. Appl. Energy 291, 116798 (2021).[6] Dodda, A., Trainor, N., Redwing, Joan. M. & Das, S All-in-one, bio-inspired, and low- power crypto engines for near-sensor security based on two-dimensional memtransistors. Nat. Commun. 13, 3587 (2022).[7] Liu, B. et al. When Machine Learning Meets Privacy: A Survey and Outlook. ACM Comput. Surv. 54, 1-36 (2022).[8] Shi, W ., Cao, J , Zhang, Q., Li, Y. & Xu, L. Edge Computing: Vision and Challenges. IEEE Internet Things J. 3, 637-646 (2016).[9] Take it to the edge. Nat. Electron. 2, 1-1 (2019).

[0010] Pettersen, R., Johansen, H. D. & Johansen, D. Secure Edge Computing with ARM TrustZone: in Proceedings of the 2nd International Conference on Internet of Things, Big Data and Security’ 102 - 109 (SCITEPRESS - Science and Technology Publications, 2017). doi : 10.5220 / 0006308601020109.

[0011] Lee, K. H. & Verma, N. A Low-Power Processor With Configurable Embedded Machine-Learning Accelerators for High-Order and Adaptive Analysis of Medical-Sensor Signals. IEEE J. Solid-State Circuits 48, 1625-1637 (2013).

[0012] Burg, A., Chattopadhyay, A. & Lam, K.-Y. Wireless Communication and Security Issues for Cyber-Physical Systems and the Internet-of-Things. Proc. IEEE 106, 38-60 (2018).

[0013] A. Rubino, C. Livanelioglu, N. Qiao, M. Payvand, & G. Indiveri. Ultra-Low-Power FDSOI Neural Circuits for Extreme-Edge Neuromorphic Intelligence. IEEE Trans. Circuits Sy st. Regul. Pap. 68, 45-56 (2021).

[0014] Near-sensor and in-sensor computing. (Springer, 2022).

[0015] Shulman, R. G., Rothman, D. L., Behar, K. L. & Hyder, F. Energetic basis of brain activity: implications for neuroimaging. Trend Neurosci. 27, 489-495 (2004).

[0016] Jaeger, H. Towards a generalized theory comprising digital, neuromorphic and unconventional computing. Neuromorphic Comput. Eng. 1, 012002 (2021).

[0017] Zhou, F. & Chai, Y. Near-sensor and in-sensor computing. Nat. Electron. 3, 664-671 (2020).

[0018] Zhou, F. et al. Optoelectronic resistive random access memory for neuromorphic vision sensors. Nat. Nanotechnol. 14, 776-782 (2019).

[0019] Cui, B. et al. Ferroelectric photosensor network: an advanced hardware solution to realtime machine vision. Nat. Commun. 13, 1707 (2022).

[0020] Chen, S. et al. In-Memory Tactile Sensor with Tunable Steep-Slope Region for Low- Artifact and Real-Time Perception of Mechanical Signals. ACS Nano 17, 2134-2147 (2023).

[0021] Verstraeten, D., Schrauwen, B., D’Haene, M. & Stroobandt, D. An experimental unification of reservoir computing methods. Neural Netw. 20, 391-403 (2007).

[0022] Sun, L. et al. In-sensor reservoir computing for language learning via two-dimensional memristors. Sci. Adv. 7, eabg1455 (2021).

[0023] Zhang, Z. et al. In-sensor reservoir computing system for latent fingerprint recognition with deep ultraviolet photo-synapses and memristor array. Nat. Commun. 13, 6590 (2022).

[0024] Lao, J. et al. Ultralow-Power Machine Vision with Self-Powered Sensor Reservoir. Adv. Sci. 9, 2106092 (2022).

[0025] Wu, X. et al. Wearable in-sensor reservoir computing using optoelectronic polymers with through-space charge-transport characteristics for multi-task learning. Nat. Commun. 14, 468 (2023).

[0026] Sun, Y. et al. In-Sensor Reservoir Computing Based on Optoelectronic Synapse. Adv. Intell. Syst. 5, 2200196 (2023).

[0027] Tan, H. & Van Dijken, S. Dynamic machine vision with retinomorphic photomemristor-reservoir computing. Nat. Commun. 14, 2169 (2023).

[0028] Zha, J. et al. Electronic / Optoelectronic Memory Device Enabled by Tellurium-based 2D van der Waals Eleterostructure for in-Sensor Reservoir Computing at the Optical Communication Band. Adv. Mater. 35, 2211598 (2023).

[0029] Xing, G. et al. Long-Range Balanced Electron- and Hole-Transport Lengths in Organic-Inorganic CH 3 NH 3 Pbl 3. Science 342, 344-347 (2013).

[0030] Yantara, N. etal. Ion-mediated Recombination Dynamics in Perovskite-Based Memory Light-Emitting Diodes for Neuromorphic Control Systems. Adv. Mater. 2305857 (2023) doi: 10.1002 / adma.202305857.

[0031] John, R. A. et al. Synergistic Gating of Electro-Iono-Photoactive 2D Chalcogenide Neuristors: Coexistence of Hebbian and Homeostatic Synaptic Metaplasticity. Adv. Mater. 30, 1800220 (2018).

[0032] Zhang, W. et al. Edge learning using a fully integrated neuro-inspired memristor chip. Science 381 , 1205-121 1 (2023).

[0033] Ambrogio, S. el al. An analog-AI chip for energy-efficient speech recognition and transcription. Nature 620, 768-775 (2023).

[0034] Shi, J. et al. From Ultrafast to Ultraslow: Charge-Carrier Dynamics of Perovskite Solar Cells. Joule 2, 879-901 (2018).

[0035] Qiu, C. et al. Probing charge carrier dynamics in metal halide perovskite solar cells. EcoMat 5, el2268 (2023).

[0036] Sakhatskyi, K. et al. Assessing the Drawbacks and Benefits of Ion Migration in Lead Halide Perovskites. ACS Energy Lett. 7, 3401-3414 (2022).

[0037] Chen, J. et al. Optoelectronic graded neurons for bioinspired in-sensor motion perception. Nat. Nanotechnol. (2023) doi: 10.1038 / s41565-023-01379-2.

[0038] Appeltant, L. et al. Information processing using a single dynamical node as complex system. Nat. Commun. 2, 468 (201 1).

[0039] Cao, J. et al. Emerging dynamic memristors for neuromorphic reservoir computing. Nanoscale 14, 289-298 (2022).

[0040] Wang, Z. et al. Memristors with diffusive dynamics as synaptic emulators for neuromorphic computing. Nat. Mater. 16, 101-108 (2017).

[0041] Moon, J. et al. Temporal data classification and forecasting using a memristor-based reservoir computing system. Nat. Electron. 2, 480-487 (2019).

[0042] Zhong, Y. et al. Dynamic memristor-based reservoir computing for high-efficiency temporal signal processing. Nat. Commun. 12, 408 (2021).

[0043] Koh, S.-G., Shima, H , Naitoh, Y., Akinaga, H. & Kinoshita, K. Reservoir computing with dielectric relaxation at an electrode-ionic liquid interface. Set. Rep. 12, 6958 (2022).

[0044] Zhong, Y. et al. A memristor-based analogue reservoir computing system for real-time and power-efficient signal processing. Nat. Electron. 5, 672-681 (2022).

[0045] Nishioka, D. et al. Edge-of-chaos learning achieved by ion-electron-coupled dynamics in an ion-gating reservoir. Sci. Adv. 8, eadel 156 (2022).

[0046] Liang, X. et al. A Physical Reservoir Computing Model Based on Volatile Memristor for Temporal Signal Processing, in 2022 29th IEEE International Conference on Electronics, Circuits and Systems (ICECS) 1-4 (IEEE, 2022). doi : 10.1 109 / ICECS202256217.2022.9970880.

[0047] Du, C. et al. Reservoir computing using dynamic memristors for temporal information processing. Nat. Commun. 8, 2204 (2017).

[0048] Chen, Z. et al. All-ferroelectric implementation of reservoir computing. Nat. Commun. 14, 3585 (2023).

[0049] Atiya, A. F. & Parlos, A. G. New results on recurrent network training: unifying the algorithms and accelerating convergence. IEEE Trans. Neural Netw. 11, 697-709 (2000).

[0050] Khalil, H. K. Nonlinear systems. Hauptbd. (Prentice Hall, 2002).

[0051] Kamisalic, A , Fister, I , Turkanovic, M. & Karakatic, S. Sensors and Functionalities of Non-Invasive Wrist-Wearable Devices: A Review. Sensors 18, 1714 (2018).

[0052] Kearney, M , Fay, M. & Fitzmaurice, D. A Stroke prevention in atrial fibrillation: we can do better. Br. J. Gen. Pract. 66, 62-63 (2016).

[0053] Pereira, T. etal. Photoplethysmography based atrial fibrillation detection: a review. Npj Digit. Med. 3, 3 (2020).

[0054] Moody, B., Moody, G , Villarroel, M , Clifford, G. & Silva, I. MIMIC-III Waveform Database Matched Subset. (2017) doi: 10.13026 / C2294B.

[0055] Bashar, S. K. Atrial Fibrillation annotations of electrocardiogram from MIMIC III matched subset. 14891 Bytes (2020) doi: 10.6084 / M9.FIGSHARE. 12149091 VI.

[0056] Liu, J. el al. Correlations between Electrochemical Ion Migration and Anomalous Device Behaviors in Perovskite Solar Cells. ACS Energy Lett. 6, 1003-1014 (2021).

[0057] Xiao, Z. el al. Giant switchable photovoltaic effect in organometal trihalide perovskite devices. Nat. Mater. 14, 193-198 (2015).

[0058] Caprioglio, P. et al. On the Relation between the Open-Circuit Voltage and QuasiFermi Level Splitting in Efficient Perovskite Solar Cells. Adv. Energy Mater. 9, 1901631 (2019).

[0059] Calado, P. et al. Driftfusion: an open source code for simulating ordered semiconductor devices with mixed ionic-electronic conducting materials in one dimension. J. Comput. Electron. 21, 960-991 (2022).

Claims

CLAIMS1. A photovoltaic cell comprising: a first layer comprising a substrate arranged to include a transparent conducting material, a second layer configured as a hole transport layer, a third layer configured as a semiconductor layer comprising a halide perovskite, a fourth layer configured as an electron transport layer, and a fifth layer comprising a conductive contact.

2. The photovoltaic cell of claim 1, wherein the halide perovskite is of the formula ABX3, where:A is selected from a monovalent metal cation or an organic cation;B is a metal cation; andX is a halide ion.

3. The photovoltaic cell of claim 2, wherein:A is selected from Cs, CH3NH3+or [HCTNH:):] ";B is selected from Pb or Sn; andX is selected from Cl, Br or I.

4. The photovoltaic cell of any of claims 1-3, wherein the transparent conducting material comprises one or more of indium tin oxide (ITO), fluorinated tin oxide (FTO), silver nanowires, carbon nanotubes, metal mesh, aluminium zinc oxide, gallium zinc oxide or indium zinc oxide.

5. The photovoltaic cell of any of claims 1-4, wherein the second layer comprises Spiro- OMeTAD, poly[bis(4-phenyl)(2,4,6-trimethylphenyl)amine]) (PTAA), Poly(3- hexylthiophene) (P3HT), Tris(4-carbazoyl-9-ylphenyl)amine (TCTA), nickel oxide, copper iodide, copper thiocyanate, molybdenum oxide, vanadium oxide or poly(3,4- ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS).

6. The photovoltaic cell of any of claims 1-5, wherein the fourth layer comprises one or more of titanium oxide, zinc oxide, stannic oxide, tungsten trioxide, 2,2',2"-(l,3,5- benzinetriyl)-tris(l -phenyl- 1-H-benzimidazole) (TBPi), C60 and phenyl-C61 -butyric acid methyl ester (PCBM).

7. The photovoltaic cell of any of claims 1-6, wherein the fifth layer comprises one or more of lithium fluoride, calcium fluoride, titanium oxide, aluminium, silver, copper, titanium or gold.

8. The photovoltaic cell of any of claims 1-7, wherein: the first layer comprises an indium tin oxide (TTO)-coated glass substrate, the second layer comprises poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS), the third layer comprises CHjNHjPbB or CEfiNHjPbE, the fourth layer comprises phenyl -C 61 -butyric acid methyl ester (PCBM), and the fifth layer comprises lithium fluoride and aluminium.

9. The photovoltaic cell of any of claims 1-8, wherein the third layer has a thickness of between about 180 nm to about 600 nm.

10. A method of preparing a photovoltaic cell, the method comprising the steps of:(a) applying a second layer to a first layer, wherein the first layer comprises a substrate arranged to include a transparent conducting material and the second layer is configured as a hole transport layer,(b) applying a third layer to the second layer, wherein the third layer comprises a halide perovskite,(c) applying a fourth layer to the third layer, wherein the fourth layer is configured as an electron transport layer, and(d) applying a fifth layer to the fourth layer, wherein the fifth layer comprising a conductive contact.

11. The method of claim 10, wherein step (a) further comprises, before step (a), the step of (al) cleaning the first layer.

12. The method of claim 11, wherein step (al) further comprises the step of (a2) treating the cleaned first layer using UV-ozone.

13. The method of any of claims 10-12, wherein step (b) further comprises solvent annealing after applying the third layer.

14. The method of any of claims 10-13, further comprising the step of (e) encapsulating the photovoltaic cell with epoxy resin and glass.

15. A photovoltaic cell obtained or obtainable from the method of any of claims 10-14.

16. A photovoltaic panel array comprising a plurality of photovoltaic cell of any of claims 1-9 and 15.

17. A photovoltaic reservoir computing (PVRC) system, comprising: a receiver configured to receive optical pulse signals associated with spatiotemporal characteristics of an entity, said optical pulse signals characterised by variations at least in intensity due to changes in the spatiotemporal characteristics; and a photodetector configured to provide, by way of in-sensor reservoir computing based upon said received optical pulse signals, corresponding unique photovoltages as output to enable determination of the spatiotemporal characteristics, wherein the unique voltages are respectively stateful representations in time for the spatiotemporal characteristics, wherein the photodetector includes at least one photovoltaic cell of any of claims 1 -9 and 15, or the photovoltaic panel array of claim 1618. The PVRC system of claim 17, wherein the PVRC system is adapted for photoplethysmography (PPG) and the entity is a human subject, further comprising: light illuminating means configured to illuminate green light pulses onto the skin tissue of the human subject, wherein the receiver is further configured to receive, based at least in part on the illuminated green light, light pulses reflected off the skin tissue, whereby the reflected lightpulses vary temporally in intensity being responsive to changes in blood volume in a microvascular bed under the skin tissue; and wherein the reflected light pulses are the optical pulse signals and the spatiotemporal characteristics is associated with atrial fibrillation (AF).

19. The PVRC system of claim 18, wherein the PVRC system is configured as a wearable device.

20. The PVRC system of claim 18 or 19, wherein the light illuminating means is a light emitting diode (LED) configured as an emitter of green light pulses.

21. The PVRC system of any of claims 17-20, further comprising output means configured to provide readout of the photovoltages as the output, wherein the photovoltages are opencircuit voltages (Foe).

22. The PVRC system of any of claims 17-21, wherein the PVRC system is configured to operate with power less than 1 fW.

23. The PVRC system of any of claims 17-22, wherein the optical pulse signals are in the form of light pulse trains.

24. A method of operating the PVRC system of any of claims 17-23, comprising: receiving, by the receiver, optical pulse signals associated with spatiotemporal characteristics of an entity, said optical pulse signals characterised by variations at least in intensity due to changes in the spatiotemporal characteristics; and providing, by the photodetector by way of in-sensor reservoir computing based upon said received optical pulse signals, corresponding unique photovoltages as output to enable determination of the spatiotemporal characteristics, wherein the unique voltages are respectively stateful representations in time for the spatiotemporal characteristics.

Images