Conventional machine vision systems based on photodiode technology have been widely applied in intelligent robotics^[1], autonomous vehicles^[2], and speech interactions^[3]. However, the physical separation of optical sensor arrays, processing units, and memory units restricts sensing and computing efficiency, resulting in high power consumption, delays, and complex chip designs^[4−7]. Emerging in-sensing vision computing systems that utilize advanced optical devices—such as optomemristors, optotransistors, and optocapacitors—offer promising solutions to these limitations by emulating human vision functions in sensing, memory, and computing^[8−10]. For example, Zhou et al. developed a 2D retina-morphic hardware device based on heterostructures that can detect and recognize moving carts, combining perceptual memory with computational capabilities^[11]. Sun et al. introduced a 1024-pixel flexible photosensor array integrating carbon nanotubes and perovskite quantum dots, designed for biological processing and neural reinforcement learning, highlighting the feasibility of high-performance neuromorphic visual sensors^[12]. Additionally, a one-photodiode and one-memristor (1P-1M) design based on MoS₂ mimics retinal cells for real-time image encoding^[13]. This system leverages van der Waals effects in 2D material heterostructures for enhanced light sensitivity, enabling simultaneous image detection and processing^[14]. The development of a 2D active pixel sensor based on a single-layer MoS₂ photoconductive transistor array aims to reduce footprint and energy consumption while improving efficiency in optical in-sensing computing^[15]. Considering the advantages of the bio-inspired optomemristor such as easy integration, high biomimicry, and fast response speed, we have built a full optomemristor implementation in-sensing vision computing system for faithfully mimicing the function of the human retina and visual cortex for image processing^[16], demonstrating that this optomemristor technology has great potential in a future computing system that is desirable for complex primary information processing.

Recent advancements in optical in-sensing reservoir computing have demonstrated effectiveness in applications like pattern classification and temporal signal processing^{[17, 18]}. Notably, Chai et al. realized optoelectronic random access memory (ORRAM) devices for efficient neuromorphic systems, demonstrating non-volatile optoelectronic switch and optically tunable synaptic behavior^[19]. While Yang et al. utilized storage layers of diffusive memristors and readout layers of drift memristors for temporal pattern classification^[20]. Lu et al. employed dynamic tungsten oxide (WO_x) memristors to nonlinearly map time inputs to reservoir states, successfully achieving spoken digit recognition and chaos prediction^[21]. Other studies have explored the application of bio-inspired design in intelligent systems through reliable memory switching in biological synapses^[22] and high-density nanowire arrays mimicking human retina photoreceptors^[23]. Meanwhile, Shen et al. utilized artificial optoelectronic synapses to achieve long-term and short-term memory along with adaptive learning, further advancing the development of neuromorphic computing^[24]. Sun et al. introduced a sensor-integrated reservoir computing system that optimized real-time signal processing for language learning^[25], complementing Wang's development of adaptive light modulation delay reservoir computing technology based on biological retinas, enhancing the system's generalization ability and stability across various environments^[26].

Despite these advancements^[27−29], the organic optic active layer-mediated flexible optomemristor-based in-sensing vision computing system has not been thoroughly demonstrated on complex analogue signal processing^{[30, 31]}.

Herein, We propose a speech signal processing system based on flexible memristors Au/C₂₇H₃₀O₁₅/FeO_x/ITO, where C₂₇H₃₀O₁₅ represents a lotus flower solution, integrating into a complete optical memristive implementation within sensory visual computing systems. The C₂₇H₃₀O₁₅/FeO_x memristors exhibit excellent photoelectric properties, demonstrating a significant positive photoeffect under light stimulation and successfully mimicking synaptic behavior akin to human retina and visual cortex functionalities. To elucidate the observed positive photoeffect and simulate the switching behavior, we propose a defect-assisted energy band physical model. The combination of the speech signal processing system with flexible memristors demonstrates outstanding accuracy, showcasing immense potential for future machine vision systems.

Fig. 1 provides an overview of the flexible optoelectronic memristor device based on the Au/C₂₇H₃₀O₁₅/FeO_x/ITO heterostructure, including its composition, structure, characteristics, and mechanisms. Fig. 1(a) shows the molecular structure of the C₂₇H₃₀O₁₅ lotus component, extracted from lotus petals to prepare a purple solution for spin-coating the C₂₇H₃₀O₁₅ switch functional layer. Fig. 1(b) shows a synaptic connection within the cochlear nerve and the structure of the flexible memristor device. The cell bodies of cochlear nerve cells are located inside the cochlear spiral, each with cilia containing numerous micro-hair cells that sense vibrations of sound waves and convert them into neural signals. These signals are transmitted via synapses to the brainstem auditory nuclei, and then further to the cerebral cortex for processing and interpretation. In the auditory system, these synapses of cochlear nerve cells are crucial for accurate processing and understanding of sound information.

The excellent optosynaptic characteristics of Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristors enable them to effectively mimic the biological synaptic behavior of cochlear nerve cells. The inset on the right in Fig. 1(b) shows a field emission scanning electron microscopy (FE-SEM) image of the cross-section of the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor, indicating approximate thicknesses of C₂₇H₃₀O₁₅ and FeO_x switch functional layers and top electrode Au of ~6, ~274, and ~15 nm respectively. The flexible Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor was fabricated using magnetron sputtering, sol-gel, and physical vapor deposition techniques. For the 274 nm thick FeO_x layer, magnetron sputtering was conducted on the flexible ITO substrate at 100 W for 120 min, divided into two growth sessions of 30 min each, with a 1 h pause in between to prevent overheating. Subsequently, the ITO substrate with the deposited FeO_x functional layer was placed in a spin coater, where C₂₇H₃₀O₁₅ solution was spun at 4500 rpm for 30 s. After uniform coating, the processed device was heated on a hot engine at 150 ^°C for 10 min and then dried in an electrically heated constant-temperature oven at 40 ^°C for 450 min. Finally, the top electrode Au was deposited under high vacuum (~1 × 10⁻⁴ Pa) at approximately 13 V, resulting in the formation of the Au/C₂₇H₃₀O₁₅/FeO_x/ITO flexible memristor.

Fig. 1(c) presents the I−V characteristic curves of the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor under the same scan rate and small bias voltage (0.2 V), with and without light exposure. It clearly shows a pronounced positive light effect under initial conditions without light exposure, laying the foundation for subsequent optical tests. Fig. 1(d) displays the UV−Vis absorption results of the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor switch functional layer (FeO_x, C₂₇H₃₀O₁₅ + FeO_x), fitted with (αhv)²−Energy (hv), where α, h, and v represent absorbance intensity, Planck's constant, and frequency, respectively. The fittings indicate semiconductor bandgaps of 2.64 and 2.20 eV for FeO_x and FeO_x with spin-coated C₂₇H₃₀O₁₅, respectively, demonstrating that spin-coating with a lotus solution on FeO_x films enhances the tunability of the optical bandgap opening and closing, resulting in improved responsiveness.

Figs. 1(e) and 2(h) show the high resistance (HRS) and low resistance states (LRS) extracted for the SCLC fitting mechanism in the simulated switch memory. Figs. 1(f) and 2(i) display the variations in the concentration of free electron defect states (N_t) in the high/low resistance states. The high and low resistance states correspond to the first to the fifteenth scanning cycles (five cycles per group). For HRS, the slope ranges from 0.85 to 1.12 at low voltages, increasing to 0.87 to 0.94 in relatively higher voltage regions, and from 2.16 to 1.52 at high voltages, with threshold voltage (V_th) ranges of 0.080 to 0.058 V and 0.626 to 0.509 V increasing with scan cycles. The Poisson equation associated with defects is as follows:

1ϵ0ϵredE(x)d(x)=n(x).

Here ϵ0, ϵr, and e represent the vacuum capacitance, relative capacitance, and electron, respectively, n(x) and E(x) represent the charge density and electric field. Charge injection is directly related to the external electric field. At low bias voltages, the defect site captures most injected electrons. Therefore, the concentration of free electrons (nf) is negligible. In the SCLC mechanism, N_t and nt are defined as total defects and defects after electron filling, respectively. nt is given by the following formula:

2nt=Nt1+exp(Et−EFkT).

It is easy to obtain, charge density: n = nf + nt. Through the Poisson equation:

3ϵ0ϵredE(x)d(x)=(nf+nt)(x).

Therefore, considering the Mott-Gurrey law to describe current density:

4J=98ϵ0ϵrθμV2L2.

Here μ and L represent the mobility and thickness of the functional layer, respectively, and θ=nf/(nf+nt) is the proportion of free electrons. Therefore, under low electric fields, the charge density is in the nf≪nt state. As the voltage increases, the current density is dominated by the Mott-Gorrey law^[32−34]. In the high voltage region, the current−voltage relationship satisfies I − V^{m > 2}. Therefore, the defect concentration in the C₂₇H₃₀O₁₅/FeO_x switch layer can be calculated as:

5Nt=2ϵ0ϵreL2Vth.

Here V_th is the threshold voltage. Fig. 1(f) shows the defect concentration calculated by Eq. (5), combined with Fig. 2(h) showing that the defect concentrations in HRS and LRS decrease gradually with increasing scan cycles. Studies have shown that in oxide materials or other composite materials, oxygen vacancies may react with water molecules as charge traps or active metal atoms (such as iron), contributing to the functionality of resistive switching memories^[35]. As the current−voltage scan cycle extends, the defect concentration decreases, indicating that electrons in deep traps do not escape under negative bias. Based on experimental phenomena and fitting results, a physical model of band structure is proposed to explain the observed switching and memory behaviors. To analyze the device mechanisms more comprehensively, we also fitted the flexible memristor using mechanisms such as Fowler-Nordheim tunneling, Schottky tunneling, Ohmic-contact conduction, and trap-assisted tunneling. The analysis results indicate that none of these mechanisms effectively match the actual situation. Fig. 1(g) demonstrates the excellent flexibility of Au/C₂₇H₃₀O₁₅/FeO_x/ITO flexible memristors at the different curvature radii (2.967, 1.25, 1.0607, 1.0104, and 1.0001 cm).

Fig. 2 focuses on the electrical testing of Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor devices and explains their electrical characteristics. Typical current−voltage (I−V) tests demonstrate significant resistive switching behavior in the device. Initially, in the first stage (0→1 V), as the bias voltage increases, the device switches from a high resistance state (HRS) to a low resistance state (LRS), with the current density gradually rising to its maximum value. In the second stage (1→0 V), the device remains in LRS, and the current density gradually decreases as the bias voltage lowers. Similarly, during the reverse bias scanning, in the third stage (0→−1 V), the device remains in LRS, and in the fourth stage (−1→0 V), the device resets from LRS back to HRS, with the current density gradually decreasing to a minimum. From these processes, it is clear that the device exhibits significant resistance switching (RS) behavior. Studies confirm that variations in bias voltage magnitude and scanning rate affect the resistive switching phenomenon through their influence on interface charges. Therefore, the current density response of the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor under various bias voltages and scan rates was investigated in this study (Figs. 2(a) and 2(b)). Results indicate that changes in current density and conductivity correspond to variations in bias voltage and scan rate within certain ranges. The instability between cycles of the memristor is influenced by random switching processes, while device-to-device instability is affected by uneven material deposition.

To assess the impact of mechanical bending on the current response of flexible memristors, devices were bent at 60^° angles for 1, 10¹, 10², 10³, and 10⁴ cycles, revealing a decrease in device conductivity with increasing bend cycles (Fig. 2(c)). Deployment and application of memristors require ensuring stability across cycles and between devices, necessitating tests for cyclic stability (Fig. 2(d)) and inter-device stability (Fig. 2(g)). Testing at a 0.2 V reading voltage confirms the device maintains stability for at least 1000 consecutive cycles. Additionally, the devices exhibit excellent stability between units, supporting deployment and application of Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristors in algorithms and hardware. Furthermore, in computational operations, aside from the aforementioned requirements, memristors are also expected to possess non-volatility and usable computing precision. At a 0.3 V reading voltage, the device demonstrates a significant on/off resistance ratio and sustains high and low resistance states for at least 10⁴ s (Fig. 2(e)), highlighting its non-volatile characteristics. Testing for computing precision involves applying different reading voltage stimuli to identify at least 32 states for the memristor (Fig. 2(f)), capable of achieving at least 5-bit computing precision in applications, facilitating computational tasks in conjunction with algorithms and expanding the device's application scope. The aforementioned electrical characterization test results affirm that the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor exhibits outstanding electrical performance, providing a robust foundation for its applications.

Fig. 2(h) shows the current−voltage fitting results of the device in the LRS state. At lower voltages, the slope ranges from 0.85 to 0.89, while in the relatively higher voltage region, it ranges from 1.06 to 1.03, and at higher voltages, it ranges from 1.87 to 1.50. Additionally, with increasing scan cycles, the V_th ranges are 0.060 to 0.040 V and 0.583 to 0.464 V respectively. Fig. 2(i) illustrates the calculated defect density, indicating that the defect concentrations in both HRS and LRS decrease gradually with increasing scan cycles, demonstrating that the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor conforms to the SCLC mechanism.

Fig. 3 details the optical characteristics of the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor, laying the groundwork for advancing optical computing. To assess the device's cyclic stability under light conditions, we conducted I−V scans at 0.2 V bias voltage and 19 mW laser power, yielding results shown in Fig. 3(a). The increasing current response with accumulated light exposure suggests excellent analog properties of the flexible memristor. Figs. 3(b) and 3(c) show the device tests based on light intensity and light exposure time (19.0−90.4 mW and 2−6 s, respectively). Testing shows that light pulses within 19.9 to 90.4 mW and 2 to 6 s can progressively program the device in the nA range, demonstrating dependency on both light pulse intensity and width for modulating photoconductive states. Under light stimulation of 19.0 −102.0 mW at frequencies of 0.3−2.0 Hz, the device exhibits nearly linear ΔI (deviation between consecutive pulses) accumulation (Fig. 3(d)), meeting requirements for broad-response tasks like speech recognition. Applying a light pulse with a light intensity of 19 mW and a pulse width of 0.5 s, gradually activates separable photoconductive states, sustaining conductive states for over 16 s (Figs. 3(e) and 3(f)). This indicates that the device's 150 stable separable multi photoconductive states may achieve 7-bit computational accuracy, ensuring artificial intelligence tasks such as speech recognition.

Fig. 3(g) demonstrates that consecutive light pulses (102 mW, 0.5 s) can toggle high and low photoconductive states, indicating short-term plasticity (STP) transitions to long-term plasticity (LTP) following 102 mW light exposure. Fundamental to learning and memory is synaptic plasticity, where dendritic "mitochondrial signals" promote synaptic transitions from short to long term, mirroring biological synapses. In biological neural systems, PPF (paired-pulse facilitation) serves as a typical function of synapses, demonstrating the ability to process continuous temporal information. When two presynaptic stimuli are applied successively with a certain time interval, the postsynaptic response triggered by the second stimulus is greater than that of the first, and this effect depends on the time interval between the two stimuli. On the developed memristor, varying light pulses (64.6 mW) at intervals from 0.5 to 13.0 s study PPF characteristics (Fig. 3(h)). Illustratively, as intervals increase, facilitation diminishes following an exponential trend, demonstrating the device's ability to mimic rapid and transient synaptic efficacy changes, akin to biological synapses. In Fig. 3(i), under a reading voltage of 0.05 V, we first applied a light pulse with a power of 19 mW and a pulse width of 0.5 s to the device. The flexible memristor exhibited a significant positive optical effect, responding positively to the light pulse, which led to an increase in current density and conductance, resulting in a transition from low to high conductance. Subsequently, we applied a negative voltage pulse of −0.1 V with a pulse width of 0.5 s, which caused a decrease in conductance, completing the transition from high to low. Through this series of operations, we achieved conductance updates driven by the electro−optical synergy effect. This synergy not only enhances the device's performance but also opens up broad prospects for its application in hardware, further elevating its developmental potential (Fig. 3(i)). In summary, the flexible memristor of Au/C₂₇H₃₀O₁₅/FeO_x/ITO demonstrates excellent performance and application potential, providing an important foundation for the realization of future optical computing and artificial intelligence tasks.

This study utilized the NIST TI-46 public dataset to train and test a speech recognition system. The TI-46 dataset consists of audio waveforms of spoken digits (0−9 in English pronunciation) from five different female speakers. The goal of the speech signal recognition is to distinguish between different audio signals of digits while disregarding speaker variations. The overall architecture of the analog signal processing system based on Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristors is depicted in Fig. 4(a). The raw digital speech signals are used as input, processed using Lyon's passive ear model to generate preprocessed input signals^[36]. After dimensional conversion, each row of the feature matrix is treated as an input vector with a duration of τ. This decodes into a series of spectral signals, reducing unnecessary information and improving system processing speed. Fig. 4(a) on the left shows a schematic diagram of the preprocessed input signal after masking.

Fig. 4(b) illustrates how the Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristor responds to 16 types of encoding modulation (0000−1111) with light pulses. Through these 16 encoding modulation states, it is observed that each sequence of encoding outputs different photocurrents. This demonstrates that the device can achieve differentiated responses to inputs in different sequences, laying the theoretical groundwork for the application of Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristors in speech recognition systems.

The processed signals are mapped into light pulse intensity, and modeled according to C₂₇H₃₀O₁₅/FeO_x memristors to obtain corresponding dynamic photovoltaic currents with temporal characteristics. By applying a sequence of random optical power temporal pulses to the Au/C₂₇H₃₀O₁₅/FeO_x/ITO photonic memristor (input), the memristor generates a dynamic response with specific features to the random input signals. Based on the photovoltaic response, the photonic memristor is modeled to deeply explore its dynamic response characteristics, as detailed in Fig. 4(c). We utilize a ten-dimensional vector as a target vector to represent the classification results of these ten digits. For instance, when the target digit is 5, the sixth element in the target vector will be 1, while the others should be 0. The system employs a parallel reservoir system (integrating masks, matrix operations, and other complex functionalities) for deep training. Within the parallel reservoir system, multiple different mask matrices are used to repetitively mask and record the same input signal, totaling N times. This process aims to capture the dynamic responses of each memory resistor within each time interval τ, and combine these responses into reservoir states for use in subsequent MLP classification tasks. The core objective of the training is to find an appropriate set of weights W_out to make the output vector as close as possible to the corresponding target vector. To achieve this goal, linear regression methods are employed in this study to precisely calculate W_out. During the prediction phase of new input speech signals, to verify the reliability and accuracy of the model, we introduce an eight-fold cross-validation strategy. Specifically, we repeat the training and testing process eight times, randomly selecting datasets each time for model training and performance testing. Finally, we average the recognition rates across all test datasets from these eight cross-validation runs to obtain the model's ultimate recognition rate.

Fig. 4(d) intuitively demonstrates the performance of the parallel RC system based on flexible Au/C₂₇H₃₀O₁₅/FeO_x/ITO memristors in predicting digit speech recognition tasks. The system achieves an accuracy of 94.88%, with the depth of individual digit recognition correlating directly with the number of digits correctly classified by the model. This highlights the substantial potential of Au/C₂₇H₃₀O₁₅/FeO_x/ITO flexible photonic memristors in practical applications.

The flexible memristors fabricated from Au/C₂₇H₃₀O₁₅/FeO_x/ITO using physical vapor deposition, magnetron sputtering, and sol-gel methods exhibit significant positive photoconductive response under optical pulse stimulation. These devices not only demonstrate excellent electrical performance but also accurately emulate various plasticity features of biological synapses, such as the STP-LTP transition from short-term learning to long-term memory and the PPF phenomenon. This supports the application of memristors in temporal data processing.

The research on high-performance parallel RC systems based on Au/C₂₇H₃₀O₁₅/FeO_x/ITO photo memristors is implemented on individual memristors operating in serial mode. This configuration optimizes system structure, reducing power consumption while enhancing efficiency, thus demonstrating the feasibility and effectiveness of memristor-based RC systems in future artificial intelligence memory networks. Overall, the study's findings expand the application scope of electronic devices, provide a new technological path for bioelectronics, showcase the integration of materials science and electronic engineering, and establish crucial theoretical and experimental foundations for future smarter, more energy-efficient electronic devices.