AI regulation and control high-frequency artificial nerve device based on homogeneous integrated OECT

Abstract

The invention discloses an AI regulation and control high-frequency artificial nerve device based on homogeneous integrated OECT. The AI regulation and control high-frequency artificial nerve device comprises a receptor module, a neuron module and a synaptic module which are integrated on the same flexible polyimide substrate. The receptor module is used for collecting and converting optical, electric and chemical signals; the neuron module is used for signal amplification and logic processing; the synaptic module is used for signal storage and synaptic plasticity simulation. Core devices adopted by the three modules are OECTs manufactured through the same manufacturing method, a vertical stacking framework is adopted in the manufacturing process, a polymer semiconductor electron transport layer-gradient ion conductor transport layer bicontinuous network is constructed, and transmission efficiency matching is achieved through path separation and collaborative optimization. According to the artificial nerve device, high-sensitivity response of multi-mode signals is achieved, the sensing-processing-memorizing integrated function is completed, and the requirements of medical scenes such as neural restoration for biocompatibility and long-term working stability are met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biosensors, and particularly relates to an AI-regulated high-frequency artificial neural device based on homogeneous integrated OECTs. BACKGROUND

[0002] In the field of neural electronics, the development of artificial neural devices has always been a research hotspot. Organic electrochemical transistors (OECTs) are widely used in the construction of brain-like electronic devices due to their low-voltage driving, strong biocompatibility, and large-area flexible preparation. However, the current artificial neural devices based on OECTs still have many limitations in performance. For example, the traditional planar OECT has the problem of cross-interference of electron-ion transmission path, which makes the transmission efficiency of electrons and ions unbalanced, resulting in slow device response. Moreover, the memory retention time of such devices is limited, and when facing high-frequency signals, it is difficult to stably complete the perception-processing-memory integrated function. In addition, the stability of the existing device structure under high-frequency driving is poor, which cannot meet the requirements of medical applications with high stability requirements.

[0003] Traditional silicon-based circuits also have many limitations in responding to neurophysiological signals. Especially when dealing with weak and high-frequency signals (up to 1 kilohertz, less than 50 millivolts), silicon-based circuits perform poorly and cannot efficiently process these signals, greatly limiting their application in neural repair and brain-computer interface fields.

[0004] In the field of artificial neural function integration, the current technology mainly adopts heterogeneous integration process, which has interface compatibility problems, and lacks effective regulation mechanism for chemical signals in medical scenarios. It does not have a specific process of real-time feedback-performance prediction-parameter closed-loop regulation of biological signals, and it is difficult to meet the requirements of low power consumption and real-time performance in medical scenarios.

[0005] In summary, the existing artificial neural devices have deficiencies in ion and electron transmission efficiency, storage performance, response speed, and stability, which cannot meet the needs of neural repair and brain-computer interface applications in medical high-frequency scenarios. Therefore, it is of great practical significance and research necessity to develop a high-frequency artificial neural device that can solve the above problems and realize multi-performance compatibility, functional integration stability, and intelligent regulation. SUMMARY

[0006] In view of the deficiencies of artificial neural devices in ion and electron transmission efficiency, storage performance, response speed, and stability, the present application provides an AI-regulated high-frequency artificial neural device based on homogeneous integrated OECTs.

[0007] The application provides an AI regulation high-frequency artificial nerve device based on homogeneous integrated OECT, mainly including a receptor module, a neuron module and a synapse module integrated on the same substrate. The spacing of each module is greater than or equal to 50 μm, so as to avoid signal crosstalk. The receptor module is used for light, electricity and chemical signal collection and conversion, and photoelectric response is greater than or equal to 0.5 A / W (450 nm light). Electrochemical response time is less than or equal to 50 ms (calcium ion concentration is 10 -6 mol / L). The neuron module is used for signal amplification and logic processing, and a two-stage GIBS-OECT amplification circuit is adopted, voltage gain is greater than or equal to 40 dB, and cutoff frequency is greater than or equal to 1.5 kHz. The synapse module is used for signal storage and synapse plasticity simulation, and the synapse weight adjustment range is 0.1-1.0 mS; conductance memory error is less than or equal to 5% at 100 kHz.

[0008] The core device adopted by the receptor module, the neuron module and the synapse module is an organic electrochemical transistor (OECT) made by the same preparation method. The preparation method of the OECT is as follows: S1, cleaning and plasma treatment are performed on the substrate.

[0009] S2, a PEDOT:PSS-PEG solution is coated on the substrate, and annealing treatment is performed in a vacuum environment to form a semiconductor layer; an ultraviolet lithography and oxygen plasma etching are adopted to form a vertical channel pattern. The PEDOT:PSS-PEG solution used for preparing the OECT can be an aqueous solution containing poly (3,4-ethylenedioxythiophene)-poly (styrene sulfonate) (PEDOT:PSS) and polyethylene glycol (PEG), or an aqueous solution of polyethylene glycol modified poly (3,4-ethylenedioxythiophene)-poly (styrene sulfonate).

[0010] S3, a poly (vinyl alcohol)-sodium alginate solution (PVA-SA solution) is used to deposit an ion conductor by coating and annealing process to form an ion conductor layer.

[0011] S4, the device obtained in step S3 is immersed in an ethanol solution of camphor sulfonic acid, soaked at 30℃ for 30 min, and then vacuum dried, so as to complete n-type doping.

[0012] S5, an Au electrode is deposited by an electron beam evaporation process, and the electrode pattern is formed by a photoresist stripping method, the spacing between the source electrode and the drain electrode is 2 μm, the gate electrode is vertically aligned with the channel to ensure the gate control efficiency; finally, packaging treatment is performed, polydimethylsiloxane is used for packaging, and oxygen plasma bonding is used to realize the close combination with the surface of the device and improve the biological stability.

[0013] The preparation methods of the OECTs of different modules are slightly different: For the OECT in the sensor module, step S3 specifically is: gradient ion conductor deposition is performed using a layer-by-layer coating-annealing process, and PVA-SA solutions with increasing concentrations are coated in sequence to form a gradient ion conductor layer. The PVA-SA solution also contains ethylene glycol bis-tetraacetic acid. Preferably, PVA-SA solutions with concentrations of 1 wt%, 2 wt%, and 3 wt% are coated in sequence to form an ion conductor gradient layer with a number of layers of 3.

[0014] For the OECT in the neuron module, step S3 specifically is: ion conductor deposition is performed using a PVA-SA solution with a mass percentage concentration of 3% by a coating-annealing process to form an ion conductor layer; and the electrode size W / L = 20 is set in step S5. For the OECT in the synapse module, step S3 specifically is: ion conductor deposition is performed using a PVA-SA solution with a mass percentage concentration of 5% by a coating-annealing process to form an ion conductor layer; and the electrode size W / L = 5.

[0015] The above-described organic electrochemical transistor is prepared using a vertical stack architecture to construct a “polymer semiconductor electronic transport layer-gradient ion conductor transport layer” dual-continuous network. Electrons are transported along the polymer semiconductor channel in the vertical direction, and ions are diffused along the ion conductor channel with a gradient distribution, so that the transmission efficiency is matched through path separation and collaborative optimization. The structural parameters are designed as follows: the vertical channel height is 500-800 nm, which is calculated based on the electron mobility formula to balance the electron transmission rate and the device integration; the number of ion conductor gradient layers is 3-5 to achieve smooth transition of the ion concentration gradient and reduce the diffusion resistance; the semiconductor / ion conductor interface roughness is ≤5 nm to reduce the interface charge trapping and improve the carrier mobility; the n-type doping concentration is 10 17 - 10 18 cm -3 , which optimizes the electron concentration and improves the transconductance value.

[0016] In the OECT preparation method, a PEDOT:PSS-PEG solution is selected as the polymer semiconductor material, and the biocompatibility and electron mobility are improved by polyethylene glycol (PEG) modification. The electron mobility is 0.15 cm 2 / (V s), which is 40% higher than that of traditional PEDOT:PSS. A poly (vinyl alcohol)-sodium alginate (PVA-SA) composite system is used as the gradient ion conductor material, and a gradient (1 wt%-5 wt%) is formed by adjusting the sodium alginate concentration. The ion conductivity gradient range is 10 -3 - 10 - 2S / cm, adaptable to ion transport requirements in different regions. Camphor sulfonic acid (CSA), with good biocompatibility, is selected as the n-type dopant. The n-type conductivity modification of PEDOT is achieved through proton transfer, avoiding the biotoxicity of traditional heavy metal dopants.

[0017] The connecting wires between the modules of the artificial neural device of the present invention are prepared by inkjet printing, and the surface of the wires is coated with a PEDOT:PSS-PEG solution to form a protective layer. After the modules are integrated, they are globally packaged using a composite encapsulation of polydimethylsiloxane and polylactic acid, with polylactic acid as the outer support structure and polydimethylsiloxane as the inner sealing layer.

[0018] Preferably, the artificial nerve device uses a flexible polyimide substrate with a thickness of 25μm and an elongation at break of ≥30%, which is suitable for conforming to the curved surfaces of human tissue.

[0019] Furthermore, the artificial neural device of the present invention employs an AI adaptive control algorithm with a closed-loop architecture of "signal acquisition layer - feature extraction layer - model prediction layer - parameter control layer".

[0020] The model's prediction layer constructs a lightweight Long Short-Term Memory (LSTM) network model. The input is a 4-dimensional feature vector, and the output is the predicted device performance parameters. The LSTM network structure is input layer 4 - hidden layer 16 - output layer 2, with 1000 training iterations and a loss function MSE ≤ 0.01. The input features of the LSTM model include the current transconductance g. m The current synaptic weight w, calcium ion concentration, and temperature are used as inputs; the output is the optimal g for the next time step. m The model training data came from 1000 sets of performance-environment parameter samples from in vitro testing. The training process used the Adam optimizer with a learning rate of 0.001.

[0021] The core update formula of the LSTM model is as follows:

[0022] In the formula, , , These are the input gate, forget gate, and output gate, respectively. Candidate memory units; For memory units; σ is the hidden layer output; W is the weight matrix; b is the bias vector; σ is the sigmoid activation function; ⊙ is the Hadamard product. Based on the model prediction results, a PID control algorithm is used to achieve precise parameter control. Taking the control of synaptic weight w as an example, the control formula is as follows: ,in e (t) = w*(t) - w (t) This is the weighting error; = 0.5、 = 0.1、 = 0.05 is the PID parameter (obtained using the Ziegler-Nichols tuning method); V GS (t) represents the output gate-source control voltage. This algorithm achieves a weighted control error of ≤3% and a control response time of ≤10ms, ensuring real-time optimization in high-frequency scenarios.

[0023] Compared with the prior art, the advantages of the present invention are: (1) This invention effectively solves the problem of cross-interference in electron-ion transport paths in traditional planar OECT through the gradient double continuous structure (GIBS) design, achieving transmission efficiency matching. This results in a volatile response time ≤30μs, a non-volatile memory frequency of 100kHz, a memory retention time ≥1000h, a neuronal circuit cutoff frequency ≥1.5kHz, and a synaptic high-frequency conductance read / write frequency ≥100kHz, significantly improving high-frequency response performance and memory stability compared to existing technologies. In in vitro electrical performance testing, an impedance analyzer was used to test the frequency response of the neuronal circuit, scanning in the 10Hz-10kHz frequency band to verify a cutoff frequency ≥1.5kHz. A 100kHz pulse signal was input using a pulse generator, and the conductance memory error of the synaptic module was tested to be ≤5%, fully demonstrating the advantages of the artificial nerve of this invention in high-frequency performance.

[0024] (2) The device of the present invention can be adapted to medical scenarios such as nerve repair. Its good biocompatibility and long-term working stability meet the standards for clinical application, which helps to reduce the risks and costs in the medical process and reduce the cost of treatment failure or repeated treatment due to poor device performance. Through animal implantation experiments, a nerve function injury model was constructed using SD rats. The device was implanted near the injury site for 12 weeks for functional verification. The motor function score and nerve conduction velocity of the experimental group were significantly improved, which proved that the artificial nerve device of the present invention can promote the recovery of nerve function, reduce the economic burden of long-term rehabilitation treatment for patients, and has significant economic benefits.

[0025] (3) This invention provides a new and effective means for the treatment of neurological diseases, mainly for nerve repair and brain-computer interfaces. In nerve repair, it can help damaged nerves recover function, such as helping patients with spinal cord injury and peripheral nerve injury to recover nerve function. In terms of brain-computer interfaces, it can realize direct communication between the brain and external devices, assist paralyzed patients in recovering motor ability, or provide new technical support for the diagnosis and treatment of neurological diseases. In animal experiments, the artificial nerve device of this invention successfully restored the conditioned reflex ability of mice with impaired nerve function, which provides a strong experimental basis for future applications in humans.

[0026] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the fabrication process of the AI-controlled high-frequency artificial nerve device based on homogeneous integrated OECT according to the present invention.

[0028] Figure 2 This is a schematic diagram of the AI ​​adaptive control closed-loop process of the artificial neural device of the present invention. Detailed Implementation

[0029] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0030] like Figure 1 As shown, the fabrication method of the AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT of the present invention adopts a homogeneous integration strategy of "same substrate-same process" to avoid interface compatibility issues of heterogeneous integration. On the same PI substrate, the device regions of the receptor module, neuron module, and synapse module are defined by photolithography partitioning, with the spacing between each module ≥50μm. Then, the core device OECT of each module is fabricated one by one.

[0031] The preparation method of the PEDOT:PSS-PEG solution used in the preparation of OECT in this embodiment is as follows: First, prepare a 50mL centrifuge tube or beaker, add 10mL of PEDOT:PSS aqueous dispersion (the common concentration of PEDOT:PSS aqueous dispersion is 1.0~1.3wt%, commercially available for direct use, no dilution required), then add 35mL of deionized water, and stir with a magnetic stirrer at low speed (300~500rpm) for 5 minutes until well mixed.

[0032] Then add 5g of PEG (molecular weight 2000~10000, liquid or solid) to the mixture and stir for 30 minutes to allow the PEG to dissolve completely (without visible particles). Then add 0.5mL of dimethyl sulfoxide (DMSO) and stir for 10 minutes (DMSO can promote PEDOT aggregation and improve conductivity; it can be omitted if biocompatibility is not required).

[0033] The stirred mixture was placed in an ultrasonic cleaner and ultrasonically defoamed for 20 minutes (to remove air bubbles and prevent pores in the subsequent film formation). After ultrasonication, the mixture was stirred at low speed for 1 hour to allow the components to fully integrate, resulting in a uniform, transparent, dark blue dispersion, namely the PEDOT:PSS / PEG solution.

[0034] The preparation method of the PVA-SA solution used in OECT preparation is as follows: First, calculate the raw material mass according to the target concentration (1%, 2%, 3%, 5%): taking 100mL of solution as an example, the total amount of PVA and SA used are 1g, 2g, 3g, and 5g, respectively, where the mass ratio of PVA to SA is 8:2 (taking into account both ion conductivity and film-forming properties). For example, to prepare a 3% concentration PVA-SA solution, 2.4g of PVA and 0.6g of SA are required.

[0035] Then, add PVA to 80 mL of deionized water and stir in an 80°C water bath for 1-2 hours until completely dissolved; add SA to the remaining deionized water and stir at room temperature for 30 minutes until no flocculent matter remains (avoid high temperature from damaging the SA structure).

[0036] Finally, slowly pour the SA solution into the PVA solution, stir at 50°C for 30 minutes to mix, ultrasonically defoam for 20 minutes (or let stand overnight), and then filter before coating.

[0037] The specific fabrication steps for the OECT in the receptor module are as follows: (1) First, the substrate was pretreated: the PI substrate was ultrasonically cleaned with acetone, ethanol and deionized water for 15 min in sequence, dried with nitrogen, and baked in an oven at 120℃ for 30 min to remove residual moisture; plasma treatment (power 100W, time 3 min) was used to improve the hydrophilicity of the substrate surface and the contact angle was reduced from 85° to 30°.

[0038] (2) Preparation of polymer semiconductor channel: PEDOT:PSS-PEG solution (concentration 5%) was coated onto PI substrate by spin coating process (rotation speed 3000r / min, time 30s), and annealed in vacuum oven at 80℃ for 60min to form a semiconductor layer with a thickness of 500nm; vertical channel pattern was formed by ultraviolet lithography (exposure dose 100mJ / cm²) and oxygen plasma etching (etching rate 50nm / min).

[0039] (3) Gradient ion conductor deposition was performed using a layer-by-layer spin-coating-annealing process: PVA-SA solutions with concentrations of 1%, 2%, and 3% were sequentially coated. Each layer was spin-coated at a speed of 2000 r / min, annealed at 60℃ for 40 min, and the thickness of each layer was controlled at 150 nm to form a gradient ion conductor layer. Ethylene glycol ditetraacetic acid (EGTA) was also added to the three concentrations of PVA-SA solutions. The gradient ion conductor layers of the sensor module have a clear functional division: the 1% concentration PVA-SA layer is the front-end stimulation capture sub-region of the sensor module, the 2% concentration PVA-SA layer is the signal transition and conduction sub-region of the sensor module, and the 3% concentration PVA-SA layer is the core signal integration sub-region of the sensor module. The amount of EGTA added needs to be precisely matched with the requirements of each region. The specific requirements for the amount of EGTA added to the three concentrations of PVA-SA solutions are as follows: 1% concentration PVA-SA layer (front-end stimulus capture sub-region): Prioritizes capturing up to 10% PVA in the physiological environment. -7 A weak calcium ion signal at mol / L. 1% PVA-SA has a low concentration and large network pores, resulting in fast ion diffusion but weak signal response. Therefore, a maximum of 1% EGTA is needed to increase the density of sensitive sites, ensuring a calcium ion binding rate of over 80%, triggering a significant change in dielectric constant Δε ≥ 20, thus meeting the "signal capture sensitivity" requirement. The percentage of EGTA added is the percentage of EGTA mass to the total mass of the solution.

[0040] 2% PVA-SA layer (signal transition and conduction sub-region): Its core function is to efficiently transmit the signal captured at the front end to the core region, requiring a balance between "responsiveness" and "conduction efficiency". The 2% PVA-SA network structure already has a certain foundation for ion conduction, and the addition of a moderate amount of 0.7% EGTA can avoid ion transport obstruction caused by excessive sensitizer, keeping the signal attenuation rate within 8% while maintaining dielectric response stability.

[0041] 3% PVA-SA layer (core signal integration sub-region): The goal is to stably aggregate signals from multiple pathways, rather than pursuing high sensitivity. The 3% PVA-SA concentration is high, the network is dense, and its ion conduction efficiency is inherently excellent (ionic conductivity can reach 10). -2 (S / cm), but excessive EGTA can easily cause aggregation, leading to signal fluctuations. A low addition of 0.5% can reduce the impact of uneven distribution of sensitive sites, making the conductance memory error ≤3%, which meets the requirement of "signal integration consistency".

[0042] Method of adding EGTA: First, use medical-grade EGTA (purity ≥99.5%) and prepare a 5wt% EGTA stock solution with deionized water (stir at room temperature for 20 minutes until completely dissolved). Then, slowly add the EGTA stock solution dropwise to the PVA-SA solution according to the target dosage (e.g., when preparing a 1% PVA-SA + 1.0% EGTA solution, take 20mL of 5% EGTA stock solution and add it to 80mL of 1wt% PVA-SA solution, stir at 50℃ and 300rpm for 40 minutes, and then sonicate to remove foam for 20 minutes before use).

[0043] (4) n-type doping treatment: The device was immersed in a CSA ethanol solution (2wt%) at 30°C for 30 min, and the residual ethanol was removed by vacuum drying (50°C, 2 h) to complete the n-type doping. The n-type doping concentration was 10. 17.5 cm -3 X-ray photoelectron spectroscopy (XPS) was used to characterize the doping effect and ensure the uniformity of doping concentration.

[0044] (5) An Au electrode (100 nm thick) is deposited using an electron beam evaporation process. The electrode pattern is formed by photoresist stripping. The source and drain electrode spacing is 2 μm. The gate electrode is vertically aligned with the channel to ensure gate control efficiency.

[0045] (6) Encapsulation: Polydimethylsiloxane (PDMS) is used for encapsulation. The encapsulation layer is 50 μm thick and is tightly bonded to the device surface through oxygen plasma bonding (80 W power, 2 min time). The water permeability is ≤10%. -6 g / (m To enhance biostability, the specific method is as follows: Sylgard 184 PDMS is used, with the prepolymer and curing agent mixed at a mass ratio of 10:1. After vacuum degassing for 30 minutes, the mixture is coated (50 μm thick) and cured in an oven at 80℃ for 2 hours. The oxygen plasma bonding parameters are: power 80 W, Ar / O2 mixed gas (volume ratio 9:1) flow rate 20 sccm, treatment time 2 minutes, and bonding pressure 0.1 MPa.

[0046] The method for preparing OECT in the neuronal module differs from the method described above in that step (3) uses only a 3wt% PVA-SA solution.

[0047] The method for preparing OECT in the synaptic module differs from the method described above in that step (3) uses only a 5wt% PVA-SA solution.

[0048] The electrode dimensions are also differentiated: the width-to-length ratio (W / L) of the receptor module is 10 (W = 20μm, L = 2μm), the neuron module has a W / L of 20, and the synapse module has a W / L of 5.

[0049] Each fabricated OECT device needs to undergo electrical performance testing. If the test fails, return to step (2) and re-fabricate according to the methods in steps (2)-(6). If the test passes, proceed to the next module of OECT fabrication.

[0050] The core performance parameter of OECT devices is transconductance (g). m ( ) is a key indicator for measuring high-frequency response capability, and the calculation formula is as follows:

[0051] Where, μ n Electron mobility (cm) 2 / (V s), obtained through Hall effect test; C ox Gate oxide capacitance (F / cm) 2 The value is calculated from the dielectric constant and thickness of the ionic conductor; W / L is the aspect ratio of the device; V GS This refers to the gate-source voltage (V), set to 0.5 - 1.5V for medical applications (low-voltage drive requirements); V th The threshold voltage (V) was obtained by fitting the transfer characteristic curve, with a target value ≤0.3V. This was achieved by optimizing μ. n (increased to 0.15 cm) 2 / (V s)) and C ox (Upgraded to 10) -8 F / cm²), making g m ≥1ms, a 5-fold improvement over traditional OECT, laying the foundation for high-frequency response.

[0052] Connecting wires between the receptor module, neuron module, and synapse module (silver nanowire ink, 10 μm linewidth, resistivity ≤10) -6 Ω m) The device was fabricated using inkjet printing. A 10nm thick PEDOT:PSS-PEG solution was coated on the surface of the wires as a protective layer to improve biocompatibility and conductivity stability. After integration, global encapsulation was performed using a composite encapsulation of PDMS and polylactic acid (PLA). PLA served as the outer support structure, and PDMS served as the inner sealing layer, ensuring the flexibility and sealing of the integrated device. This resulted in an AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT. Specific composite encapsulation method: PLA (model 4032D) was 3D printed with a thickness of 200μm. Medical epoxy adhesive (model EPOTEK 301) was applied between PLA and PDMS and cured at 60℃ for 1 hour, achieving a bond strength ≥5MPa.

[0053] The ion conductor layer of the sensor module is doped with a calcium ion sensor (ethylene glycol ditetraacetic acid, EGTA), and the binding constant of EGTA to calcium ions is Kd = 10. -6 mol / L, suitable for physiological concentration range (10 -7 - 10 -3 (mol / L). When calcium ions combine with EGTA, the dielectric constant of the ionic conductor changes (Δε≥20), as shown by the formula... (ε0 is the vacuum permittivity, ε The relative permittivity (where S is the electrode area) causes a change in the gate capacitance, which in turn modulates the drain current (Io) of the OECT. D This system converts chemical signals into electrical signals. AI algorithms analyze the converted electrical signals to identify physiological states (such as nerve cell activity) corresponding to changes in calcium ion concentration, dynamically adjusting the conductance memory parameters of the synaptic module to achieve memory updates under chemical regulation. In this embodiment, the photoelectric responsivity of the sensor module reaches 0.6 A / W (450 nm light); the electrochemical response time is 40 ms (calcium ion concentration 10... -6 (mol / L). The voltage gain of the neuron module is 45dB; the cutoff frequency is 1.6kHz.

[0054] The synaptic weight adjustment range of the synaptic module is 0.1 - 1.0 mS; the conductivity memory error at 100kHz is ≤3%.

[0055] The AI ​​adaptive control algorithm used in the artificial neural device of this invention is as follows: Figure 2 As shown. The algorithm runs on a low-power microcontroller (MCU, model STM32L476), with power consumption ≤5mW (4mW in this embodiment), suitable for implantable scenarios. The signal acquisition layer acquires the operating parameters of the acquisition device (I... D V G V DThe system collects biological feedback signals (calcium ion concentration, local temperature, pH value) at a sampling frequency of 1kHz and a sampling precision of 12 bits. Specifically, the MCU acquires the drain current I of the OECT via a 12-bit ADC module (integrated within the MCU). D Gate voltage V G (Sampling rate 1kHz, sampling accuracy ±0.5LSB) Data from an external temperature sensor (model DS18B20) and a calcium ion concentration sensor (model SEN0199) is read via I2C interface. The feature extraction layer uses wavelet transform (db4 wavelet, decomposed into 3 layers) to extract the time-domain features (peak value, mean, variance) and frequency-domain features (center frequency, bandwidth) of the signal. Principal component analysis (PCA) reduces the feature dimension from 12 to 4, lowering computational complexity. The model prediction layer constructs a lightweight long short-term memory (LSTM) network, with a 4-dimensional feature vector as input and predicted device performance parameters (transconductance g) as output. m The LSTM network structure is "Input Layer 4 - Hidden Layer 16 - Output Layer 2", with 1000 training iterations and a loss function MSE ≤ 0.01. The parameter control layer, based on the prediction results, uses the MCU to control the digital-to-analog converter (DAC) (model AD5624, 12-bit resolution) via the SPI interface to output a control voltage (0-2V). A 0.1Ω precision resistor is connected in series at the DAC output to limit the current. The control voltage is transmitted to the gate electrode of the OECT via a flexible wire (silver nanowire ink), with a response time ≤ 10ms, adjusting the gate-source voltage (V) of the device. GS ) and drain-source voltage (V DS This enables performance optimization.

[0056] I. External Electrical Performance Testing 1.1 High-frequency response test method An impedance analyzer (Agilent E4990A) was used to construct a test system. The neuron module of the artificial neural device of this invention was connected to the test circuit. The test frequency band was set to 10Hz-10kHz, the scan step size was 10Hz, and the input signal amplitude was 0.5Vpp. The voltage gain and phase changes at different frequencies were recorded. A pulse generator (Tektronix AFG3021C) was used to input a 100kHz square wave pulse signal (amplitude 0.8V, duty cycle 50%) to the synapse module. The synaptic conductance value was acquired in real time using a source meter (Keithley 2450), and the conductance memory error over 1000 consecutive pulse cycles was calculated.

[0057] Test results show that the neuron circuit has a voltage gain of 40dB and a phase shift of ≤10° at a frequency of 1.5kHz, which meets the design requirement of a cutoff frequency of ≥1.5kHz; the synaptic module has a conductance memory error of 3.2% (≤5%) under 100kHz pulse signal excitation, and has excellent high-frequency memory stability.

[0058] 1.2 Stability Testing Methods A simulated bodily fluid environment test chamber was constructed, with the temperature set at 37℃±0.5℃ and relative humidity at 95%±2%. The encapsulated device was immersed in the simulated bodily fluid (composition: NaCl 0.9wt%, KCl 0.04wt%, CaCl2 0.02wt%) and continuously powered on for 1000 hours. The transconductance (g) was tested every 24 hours using a transconductance testing system. m The synaptic weights (w) are used to calculate the rate of change of performance parameters relative to their initial values.

[0059] Test results show that after 1000 hours of continuous operation, the transconductance (g) m The change rate was 8.5%, and the change rate of synaptic weight (w) was 9.2%, both ≤10%, indicating that the long-term working stability met the standard.

[0060] 1.3 Multimodal Response Testing Methods Optical signal response test: using a 450nm laser light source (power density 1mW / cm²). 2 The multimodal sensor was vertically irradiated, and the output current was recorded using a photoelectric testing system to calculate the photoelectric responsivity. For the electrical signal response test: a 1kHz sinusoidal electrical signal (amplitude 0.5V) was input, and the amplitude change before and after signal amplification was measured to calculate the conversion efficiency. For the chemical signal response test: a solution with a concentration of 10... -6 A mol / L calcium ion solution was added dropwise to the surface of the sensor, and the time from the addition to the stable output current (response time) was recorded. The signal-to-noise ratio of the output signal was then tested using a spectrum analyzer.

[0061] Test results show that the photoelectric responsivity of the 450nm optical signal is 0.55A / W, which is greater than 0.5A / W; the conversion efficiency of the 1kHz electrical signal is 82%, which is greater than 80%; 10 -6 The response time of the mol / L calcium ion signal is 45ms, which is less than 50ms, and the signal-to-noise ratio of the output signal is 32dB, which is greater than 30dB, indicating that the multimodal response performance meets the design specifications.

[0062] II. Biocompatibility Assessment 2.1 Cytotoxicity test methods Prepare device extraction solution according to GB / T 16886.5-2017 standard: Cut the device into 1cm × 1cm samples, according to 1cm... 2Immerse L929 fibroblasts in RPMI 1640 medium at a ratio of / mL, incubate at 37℃ for 72h, filter to sterilize, and set aside; 5×10⁹ L929 fibroblasts are then added to the medium. 3 Cells were seeded per well in a 96-well plate and cultured for 24 hours. The culture medium was then replaced with the extraction medium and cultured for another 48 hours. The absorbance was measured using the MTT assay with a microplate reader (wavelength 490 nm) to calculate cell viability and assess cytotoxicity levels.

[0063] Test results showed that the survival rate of L929 fibroblasts was 93%, which is greater than 90%. According to the GB / T 16886.5-2017 standard, the cytotoxicity level was 1, indicating no significant cytotoxicity.

[0064] 2.2 Hemolysis Test Method Take 10 mL of fresh venous blood from healthy New Zealand rabbits, add 3.8 wt% sodium citrate solution for anticoagulation (blood:anticoagulant = 9:1), centrifuge to separate red blood cells, wash 3 times with physiological saline to prepare a 2 wt% red blood cell suspension; mix 0.5 g of device sample with 10 mL of physiological saline, incubate at 37℃ for 30 min, add 2 mL of red blood cell suspension, continue incubation for 60 min, centrifuge, take the supernatant, and measure the absorbance using a UV spectrophotometer (wavelength 545 nm); use physiological saline as a negative control and distilled water as a positive control to calculate the hemolysis rate.

[0065] Test results showed that the hemolysis rate was 1.2%, less than 2%, which meets the industry standard of ≤5% for hemolysis rate of medical implant materials, and there is no risk of hemolysis.

[0066] 2.3 Inflammatory Response Testing Methods Ten SD rats (weighing 200-250g) were selected. After anesthesia, a 1cm×0.5cm device was implanted subcutaneously in their backs. Three rats were randomly selected and sacrificed on postoperative days 1, 7, and 30. Tissue samples were taken from a 0.5cm radius around the implantation site. After formaldehyde fixation, paraffin embedding, sectioning, and HE staining, the number of infiltrating inflammatory cells (lymphocytes and macrophages) was observed under an optical microscope, and the ratio of the number of inflammatory cells on day 30 to that on day 1 was calculated.

[0067] Test results showed that a small number of inflammatory cells were infiltrated 1 day after surgery, the number of infiltrates decreased by 50% 7 days after surgery, and the number of inflammatory cells decreased by 85% 30 days after surgery compared with 1 day. There was no obvious chronic inflammatory response, and the biocompatibility was excellent.

[0068] III. Animal Implantation Experiment Testing Twenty SD rats (weighing 200-250g) were randomly divided into an experimental group (n=10) and a control group (n=10). A neurological function injury model was established using a spinal cord hemisection. The experimental group received the artificial nerve device of this invention, while the control group received a traditional OECT artificial nerve device without AI regulation. The implantation sites were near the spinal cord injury area in both groups. The control group received a traditional planar OECT device, which used a PI substrate, a pure PEDOT:PSS semiconductor layer, a gradient-free PVA-SA ion conductor layer (3wt% single concentration), and no AI regulation module. Other preparation processes were consistent with those of this invention. Twelve weeks post-operation, hind limb motor function was assessed using the BBB motor function scoring method, and the nerve conduction velocity of the hind limb gastrocnemius muscle was tested using an electromyography (Nihon Kohden MEB-9404) instrument. Twelve weeks post-operation, the rats were sacrificed, the implanted device was removed, and the rate of change of electrical performance parameters was measured using an impedance analyzer to observe device degradation and tissue adhesion.

[0069] Motor function assessment results: The average BBB score of the experimental group was 18.5 points, while that of the control group was 11.2 points. The recovery of motor function in the experimental group was significantly better than that in the control group. This indicates that the artificial neural device of the present invention can effectively promote the recovery of neural function and help experimental animals regain some motor ability.

[0070] Electrophysiological test results: The average nerve conduction velocity in the hind limb muscles of the experimental group was 31.5 m / s, while that of the control group was 21.0 m / s. The experimental group showed a 50% improvement over the control group, indicating a significant improvement in nerve conduction function. This demonstrates that the artificial nerve device of this invention can effectively improve nerve conduction performance.

[0071] Device safety assessment results: The device removed from the experimental group showed no significant degradation, no adhesion to surrounding tissues, and its transconductance (g) remained normal. m The average rate of change in core electrical performance parameters such as synaptic weight (w) was 12%, less than 15%; the control group device showed slight degradation, and three rats showed slight tissue adhesion, with an average rate of change in performance parameters of 28%. This indicates that the artificial nerve device of the present invention has significant advantages in long-term implantation safety and stability.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A high-frequency artificial neural device for AI modulation based on homogeneous integrated OECT, characterized in that, It includes a receptor module, a neuron module, and a synapse module integrated on the same substrate; the receptor module is used for the acquisition and conversion of optical, electrical, and chemical signals; the neuron module is used for signal amplification and logic processing; and the synapse module is used for signal storage and synaptic plasticity simulation. The core device used in all three modules is an organic electrochemical transistor (OECT) fabricated using the same method. The fabrication method of the OECT is as follows: S1. Clean and plasma-treat the substrate; S2. A PEDOT:PSS-PEG solution is coated onto a substrate and annealed under vacuum to form a semiconductor layer; a vertical channel pattern is formed by ultraviolet lithography and oxygen plasma etching; the PEDOT:PSS-PEG solution is an aqueous solution containing both PEDOT:PSS and polyethylene glycol PEG, or an aqueous solution containing polyethylene glycol-modified poly(3,4-ethylenedioxythiophene)-poly(styrene sulfonate); S3. An ion conductor layer is formed by depositing an ion conductor using a poly(vinyl alcohol)-sodium alginate solution through a coating-annealing process. S4. Immerse the device obtained in step S3 in an ethanol solution of camphor sulfonic acid, soak it at 30°C for 30 min, and then vacuum dry it to complete the n-type doping. S5. Set the source electrode, drain electrode and gate electrode, and finally perform the encapsulation process; For the OECT in the sensor module, step S3 specifically involves: using a layer-by-layer coating-annealing process to deposit a gradient ion conductor, sequentially coating with poly(vinyl alcohol)-sodium alginate solution of increasing concentration to form a gradient ion conductor layer; the poly(vinyl alcohol)-sodium alginate solution contains ethylene glycol ditetraacetic acid. For OECT in the neuronal module, step S3 specifically involves: depositing an ion conductor layer using a 3% (w / w) poly(vinyl alcohol)-sodium alginate solution via a coating-annealing process; the electrode size set in step S5 is W / L = 20. For the OECT in the synaptic module, step S3 specifically involves: depositing an ion conductor layer using a 5% (w / w) poly(vinyl alcohol)-sodium alginate solution through a coating-annealing process; the electrode size W / L = 5.

2. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 1, characterized in that, In step S5, an Au electrode is deposited using an electron beam evaporation process, and the electrode pattern is formed by photoresist stripping. The gate electrode is vertically aligned with the channel. Finally, polydimethylsiloxane is used for encapsulation, and oxygen plasma bonding is used to achieve a tight bond with the device surface.

3. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 1, characterized in that, The connecting wires between the modules are prepared by inkjet printing, and the surface of the wires is coated with a PEDOT:PSS-PEG solution to form a protective layer.

4. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 3, characterized in that, After integration, each module is globally encapsulated using a composite encapsulation of polydimethylsiloxane and polylactic acid, with polylactic acid serving as the outer support structure and polydimethylsiloxane as the inner sealing layer.

5. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 1, characterized in that, The substrate is a flexible polyimide with a thickness of 25 μm and an elongation at break of ≥30%.

6. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 1, characterized in that, This artificial neural device employs an AI adaptive control algorithm with a closed-loop architecture consisting of a "signal acquisition layer - feature extraction layer - model prediction layer - parameter control layer".

7. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 6, characterized in that, The model prediction layer constructs a lightweight long short-term memory (LSTM) network model. The input is a 4-dimensional feature vector, and the output is the predicted device performance parameters. The LSTM network structure is input layer 4 - hidden layer 16 - output layer 2. The training iterations are 1000 times, and the loss function MSE≤0.

01.

8. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 7, characterized in that, The input features of the LSTM model include the current transconductance g. m Current synaptic weight w, calcium ion concentration, and temperature; The output is the optimal g for the next time step. m With w.

9. The AI-controlled high-frequency artificial neural device based on homogeneous integrated OECT as described in claim 8, characterized in that, The core update formula of the LSTM model is as follows: In the formula, , , These are the input gate, forget gate, and output gate, respectively. Candidate memory units; For memory units; σ is the hidden layer output; W is the weight matrix; b is the bias vector; σ is the sigmoid activation function; ⊙ is the Hadamard product.

Images