Optical tuning coding artificial visual neuron without information loss rate, preparation method and use method

Abstract

The invention discloses an information-loss-rate-free optical tuning coding artificial visual neuron, a preparation method and a use method. The preparation method comprises the following steps: firstly, growing an oxide layer on a silicon substrate, then forming a photosensitive channel by spin-coating photosensitive quantum dots on a channel layer, then forming an FET source-drain structure through photoetching and thermal evaporation, and then forming a TS parallel structure through atomic layer deposition, photoetching and thermal evaporation, thereby forming a 2TS-FET device and completing the preparation of the artificial visual neurons. The artificial visual neuron tests the peak current of the 2TS-FET device through the peripheral circuit as information coding. According to the invention, the TS parallel connection is utilized to reduce the current refractory period and the channel quantum dots have different sensing capabilities for light with different wavelengths, and the performance of the transistor is shown as the magnitude of channel current, so that the magnitude and frequency of coding current of a device are influenced, and the artificial visual neuron can be optically tuned without information loss. The artificial visual neuron is simple in structure, high in coding precision and low in power consumption and has CMOS compatibility.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an optically tuned coding artificial visual neuron with no information loss rate, and a preparation and use method thereof. Background Art

[0002] The sensory nervous system converts external information into biological signals, among which the visual perception system receives most of the information from the outside world and plays a vital role. The rise of neuromorphic technology has promoted the development of artificial intelligence, while biologically inspired devices have promoted neuromorphic visual engineering. At present, the devices used for neuromorphic visual computing mainly include CMOS and bionic visual neuromorphic hardware. The former is limited by the von Neumann bottleneck, and its development speed is greatly limited. In addition, traditional CMOS sensors cannot directly convert sensory input into pulse sequences. It needs to receive, process and transmit analog information in a continuous time. Such characteristics make it necessary to drive this device with great power and complex circuit structure. At the same time, due to the limitation of the von Neumann bottleneck, its development speed can no longer reach a new high. Unlike traditional CMOS sensors, artificial visual neurons can directly convert analog signals into peak signals, which can be directly used for the calculation of artificial neural networks. There are generally the following requirements for artificial visual neurons: 1. Ability to encode visually perceived analog signals into spike signals and provide them to artificial neural networks; 2. Efficiently and completely encode the captured information; 3. Wide recognition range and high recognition accuracy.

[0003] At present, artificial visual neurons generally adopt the information coding rules in biology, that is, judging what information is transmitted according to the discharge rate (the inverse of the average spike excitation interval). This coding method requires a long time interval for collection, usually hundreds of milliseconds to a few seconds, but in fact, the average information processing time of a single layer of biological neurons is only 10ms. This type of information coding method makes the current artificial visual neurons have the following disadvantages: 1. The discharge rate fluctuates too much in each short time interval, and it is impossible to accurately simulate biological information coding; 2. The randomness of a single volatile resistive memory is too strong. The damage or refractory period of a single memory will have a catastrophic effect on the representation of the code, causing the device to lose information during the information coding process.

[0004] At present, artificial visual neurons generally use a separate volatile resistive memory. The principle is that the threshold conversion switch TS itself has a memristive characteristic and can generate a spike current under a changing voltage signal. This type of artificial visual neuron has the following disadvantages: 1. A single volatile resistive memory cannot respond to changing light and is not light-tunable; 2. The active electrode layer usually uses active metals such as silver and copper, which are easily oxidized, causing the device to fail. Therefore, it is necessary to develop a new type of artificial visual neuron to overcome the problems faced by the current traditional artificial visual neurons that use a single volatile resistive memory. Summary of the invention

[0005] The purpose of the present invention is to address the shortcomings of existing artificial visual neurons and provide an artificial visual neuron based on a 2TS-FET structure, a preparation method and a use method. The artificial visual neuron has the advantages of no information loss rate, direct encoding of analog signals, tunable colors, and long stable life, and has broad application prospects in multiple fields such as visual coding and image recognition.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides an optically tuned coded artificial visual neuron with no information loss rate, the artificial visual neuron comprising a silicon substrate, an oxide layer and a silicon back gate formed on the silicon substrate, a channel layer formed on the oxide layer, the channel layer having a source electrode layer, a quantum dot layer and a drain electrode layer, the silicon substrate, the oxide layer, the channel layer, the quantum dot layer, the silicon back gate, the source electrode layer and the drain electrode layer constitute a field effect transistor FET structure; two parallel threshold conversion switch TS structures are formed on the drain electrode layer, each threshold conversion switch structure is formed by stacking a drain electrode layer, an intermediate layer and an active electrode layer; a common protective electrode layer is formed on the active electrode layer.

[0008] Furthermore, the oxide layer material is silicon oxide.

[0009] Furthermore, the channel layer material is molybdenum disulfide or boron nitride.

[0010] Furthermore, the quantum dot layer material is lead sulfide, perovskite or boron-doped silicon quantum dots.

[0011] Furthermore, the intermediate layer material is aluminum oxide or hafnium oxide.

[0012] Furthermore, the source electrode layer and the drain electrode layer are made of nickel, platinum or gold.

[0013] Furthermore, the active electrode layer material is silver or copper.

[0014] Furthermore, the common protective electrode layer is made of platinum or gold, which covers the active metal that can generate conductive filaments but is unstable, thereby improving the stability and life of the device while being able to work normally.

[0015] Furthermore, the neuron utilizes the different perception abilities of the same quantum dot on the channel to different wavelengths of light, as well as the different perception abilities of different quantum dots to different wavelengths of light, so that the channel resistance changes under different light illumination, which manifests itself as high and low channel currents in transistor performance, thereby affecting the high and low coding currents of the device, making the artificial visual neuron optically tunable; the use of a TS parallel structure can reduce the refractory period of the coding current, so that there is no information loss in the artificial visual neuron.

[0016] In a second aspect, the present invention provides a method for preparing an optically tuned coding artificial visual neuron with no information loss rate, the method comprising the following steps:

[0017] (1) forming an oxide layer on the surface of a silicon substrate, selecting a gate region, and removing the oxide layer in the gate region to form a silicon back gate;

[0018] (2) mechanically stripping the surface of the oxide layer to form a channel layer, and forming a source electrode layer and a drain electrode layer on the channel layer by photolithography and deposition;

[0019] (3) forming two parallel TS structure regions on the drain electrode layer by photolithography, then forming an intermediate layer by deposition, then forming an active electrode layer by deposition, and finally forming a common protection electrode layer by deposition;

[0020] (4) A quantum dot layer is formed on the channel layer by spin coating and annealing, and finally an optically tunable coding artificial visual neuron with no information loss rate based on a 2TS-FET quantum dot structure is formed.

[0021] Furthermore, in step (1), the method for forming an oxide layer on the surface of the silicon substrate is dry oxygen oxidation, the oxide layer material is silicon oxide, and the thickness is tens to hundreds of nanometers; the method for removing silicon oxide in the gate area is mechanical removal to form a silicon back gate.

[0022] Furthermore, in step (2), the method for mechanically stripping the channel layer is a tape peeling method, and the thickness of the channel layer material is several nanometers to tens of nanometers; the method for depositing the source and drain electrode layer is sputtering or evaporation, and the thickness of the source and drain electrode layer is tens of nanometers.

[0023] Furthermore, in step (3), the method for depositing the intermediate layer is atomic layer deposition, and the thickness of the atomic layer deposition is tens of nanometers. The atomic layer deposition method is used to grow the intermediate layer on the drain electrode layer; the method for depositing the active electrode layer is sputtering or evaporation, and the thickness of the active electrode layer is tens of nanometers; the method for depositing the common protection electrode layer is sputtering or evaporation, and the thickness of the common protection electrode layer is tens of nanometers.

[0024] Furthermore, in step (4), the method for spin coating the quantum dot layer is static spin coating; and the method for annealing the quantum dot layer is high temperature thermal annealing.

[0025] In the third aspect, the present invention provides a method for using an artificial visual neuron with optical tunable coding with no information loss rate, the method comprising: applying light above the quantum dot layer, applying a voltage to the common protection electrode layer, applying a constant voltage to the silicon back gate, grounding the source electrode layer, and measuring the curve of the current between the common protection electrode layer and the source electrode layer as a function of the wavelength of light and the voltage applied to the common protection electrode layer, to achieve optoelectronic bidirectional tunable coding with no information loss rate.

[0026] Furthermore, the quantum dot layer is irradiated with light within a wavelength range that enables the quantum dots to absorb light; the common protection electrode layer is applied with a voltage within a voltage range that enables the TS device to work normally; the silicon back gate is applied with a constant voltage within a voltage range that enables the FET device to work in a linear region; and the current between the common protection electrode layer and the source electrode layer of the 2TS-FET working under normal excitation is taken as the encoding information.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] The artificial visual neuron provided by the present invention is based on the light tuning of the channel quantum dot photosensitivity device, so it has a wide photosensitivity range and high photosensitivity accuracy, and can simultaneously realize dual perception of light wavelength and light intensity. Secondly, the artificial visual neuron adopts a TS parallel device structure, which can reduce the refractory period of the coding current, so that the discharge rate in each shorter time interval is stably output, so that the coding process without information loss can be realized, and the randomness is reduced. In addition, by utilizing the common protective electrode layer when TS is connected in parallel, by using stable metals such as platinum and gold, the active metal that can generate conductive filaments but is unstable is covered, so that the device can work normally while improving stability and life. Finally, the artificial visual neuron can directly encode analog signals into spike signals, which can be directly used for the calculation of artificial neural networks, and has a wide range of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0030] Figure 1 It is a front view schematic diagram of the 2TS-FET structure;

[0031] Figure 2 It is a three-dimensional schematic diagram of the 2TS-FET structure;

[0032] Figure 3 Schematic diagram of the states for spike coding in artificial visual neurons;

[0033] Figure 4 Schematic diagram of the information loss-free encoding results when performing spike encoding for artificial visual neurons, which is different from a single TS-FET structure;

[0034] Figure 5 Schematic diagram of the light-tuning function under different illumination when performing spike coding for artificial visual neurons;

[0035] In the figure, there are a silicon substrate 10 , an oxide layer 11 , a silicon back gate 12 , a channel layer 20 , a source electrode layer 21 , a drain electrode layer 22 , an intermediate layer 30 , an active electrode layer 31 , a common protective electrode layer 32 , and a quantum dot layer 33 . DETAILED DESCRIPTION

[0036] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0038] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0039] like Figure 1 , Figure 2 As shown, an embodiment of the present invention provides an optically tuned coding artificial visual neuron with no information loss rate. The artificial visual neuron is based on a 2TS-FET structure, including a silicon substrate 10, an oxide layer 11 and a silicon back gate 12 are formed on the silicon substrate 10, a channel layer 20 is formed on the oxide layer 11, and a source electrode layer 21, a quantum dot layer 33 and a drain electrode layer 22 are provided on the channel layer 20. The silicon substrate 10, the oxide layer 11, the channel layer 20, the quantum dot layer 33, the silicon back gate 12, the source electrode layer 21, and the drain electrode layer 22 constitute a field effect transistor FET structure, which is denoted as FET;

[0040] Two parallel threshold conversion switch TS structures are formed on the drain electrode layer 22, which are denoted as 2TS; each threshold conversion switch structure is formed by stacking the drain electrode layer 22, the intermediate layer 30, and the active electrode layer 31 in sequence;

[0041] A common protective electrode layer 32 is formed on the active electrode layer 31 .

[0042] Further, the oxide layer material includes but is not limited to silicon oxide, with a thickness of tens to hundreds of nanometers. The channel layer material includes but is not limited to molybdenum disulfide or boron nitride, with a thickness of several nanometers to tens of nanometers. The quantum dot layer material includes but is not limited to lead sulfide, perovskite or boron-doped silicon quantum dots, and the following settings can be adopted but are not limited to this: the lead sulfide concentration is 50mg / ml, and the spin coating amount is 10ml; the perovskite concentration is 20mg / ml, and the spin coating amount is 10ml; the boron-doped silicon quantum dots are 5mg / ml, and the spin coating amount is 20ml. The intermediate layer material includes but is not limited to aluminum oxide or hafnium oxide, with a thickness of tens of nanometers. The source electrode layer and drain electrode layer materials include but are not limited to platinum, nickel or gold, with a thickness of tens of nanometers. The active electrode layer material includes but is not limited to silver or copper, with a thickness of tens of nanometers. The common protection electrode layer material includes but is not limited to platinum or gold, with a thickness of tens of nanometers.

[0043] In the present invention, the 2TS-FET structure can adopt a MoS2 channel, and a quantum dot layer is formed on the channel by spin coating. Quantum dots, as a special photosensitive material, have a wide wavelength perception range, and different quantum dots have different wavelength absorption peaks. For example, the optical absorption peak of PbS quantum dots is at 780nm, i.e., in the near-infrared region; while the optical absorption peak of boron-doped silicon quantum dots is at 300nm, i.e., in the ultraviolet region. By utilizing the different light perception capabilities of the same quantum dot on the channel to different wavelengths, and the different light perception capabilities of different quantum dots to different wavelengths, the channel resistance of the device will also change under different illumination, which is manifested in the transistor performance as high and low channel currents, thereby affecting the high and low coding currents of the device, thereby affecting the firing rate of artificial visual neurons, so that artificial visual neurons can be optically tuned.

[0044] In the present invention, the common protective electrode layer when TS is connected in parallel is utilized, and the active metal that can generate conductive filaments but is unstable is covered by using stable metals such as platinum and gold, so that the device can work normally while improving the stability and life span, avoiding the device inactivation caused by the failure of single neuron excitation.

[0045] In the present invention, a TS parallel device structure is adopted, which can avoid the refractory period of the coding current, so that the discharge rate of the artificial visual neurons in each short time interval is stably output, avoiding the information loss caused by the random excitation of a single neuron, so that the coding process without information loss can be realized, the randomness is reduced, and the biological information coding can be accurately simulated, so that the artificial visual neurons have no information loss.

[0046] The preparation of the artificial visual neuron based on the 2TS-FET structure proposed in the present invention comprises the following steps:

[0047] (1) oxidizing the surface of the silicon substrate 10 to form an oxide layer 11, selecting a gate region, removing the oxide layer 11, and forming a silicon back gate 12;

[0048] (2) Mechanically stripping the surface of the oxide layer 11 to form a channel layer 20, and forming a source electrode layer 21 and a drain electrode layer 22 on the channel layer 20 by photolithography and deposition;

[0049] (3) forming two parallel TS structure regions on the drain electrode layer 22 by photolithography, then forming an intermediate layer 30 by deposition, then forming an active electrode layer 31 by deposition, and finally forming a common protection electrode layer 32 by deposition;

[0050] (4) A quantum dot layer 33 is formed on the channel layer 20 by spin coating and annealing, and finally an optically tuned coding artificial visual neuron with no information loss rate based on a 2TS-FET quantum dot structure is formed.

[0051] Furthermore, in step (1), the method for oxidizing the surface of the silicon substrate 10 to form the oxide layer 11 is dry oxygen oxidation, the oxide layer material is silicon oxide, and the thickness is tens to hundreds of nanometers; the method for removing silicon oxide in the gate area is mechanical removal to form a silicon back gate 12.

[0052] In the step (2), the method for mechanically stripping the channel layer 20 is a tape peeling method, and the material thickness of the channel layer 20 is several nanometers to tens of nanometers; the method for depositing the source and drain electrode layers is sputtering or evaporation, and the thickness of the source and drain electrode layers is tens of nanometers, and the source electrode layer 21 and the drain electrode layer 22 are deposited on the channel layer 20.

[0053] In the step (3), the method for depositing the intermediate layer 30 is atomic layer deposition, and the thickness of the atomic layer deposition is tens of nanometers. The intermediate layer 30 is grown on the drain electrode layer 22 by the atomic layer deposition method; the method for depositing the active electrode layer 31 is sputtering or evaporation, and the thickness of the active electrode layer 31 is tens of nanometers; the method for depositing the common protection electrode layer 32 is sputtering or evaporation, and the thickness of the common protection electrode layer 32 is tens of nanometers.

[0054] The method for spin coating the quantum dot layer 33 in step (4) is static spin coating, that is, dripping the quantum dot solution in a stationary state, and then starting the coating machine to rotate at high speed to make the quantum dot solution evenly distributed; the method for annealing the quantum dot layer 33 is high temperature thermal annealing.

[0055] The method for using the artificial visual neuron based on the 2TS-FET structure proposed in the present invention is specifically as follows:

[0056] Apply light above the quantum dot layer 33, apply voltage to the common protection electrode layer 32, apply a constant voltage to the silicon back gate 12, and ground the source electrode layer 21. Measure the curve of the current between the common protection electrode layer 32 and the source electrode layer 21 as the wavelength of light changes and the voltage applied to the common protection electrode layer 32, and realize photoelectric bidirectional tunable coding. The light applied to the quantum dot layer 33 is the wavelength range of light that enables the quantum dots to achieve light absorption, the voltage applied to the common protection electrode layer 32 is the voltage range that enables the TS device to work normally, and the constant voltage applied to the silicon back gate 12 is the voltage range that enables the FET device to work in the linear region. The current between the common protection electrode layer 32 and the source electrode layer 21 of the 2TS-FET working under normal excitation is taken as the coding information. Figure 3 As shown, when the artificial visual neuron is working, the source electrode layer 21 is grounded, and the common protection electrode layer 32 is connected to an external voltage signal V in , the silicon back gate 12 is connected to a constant voltage signal V g , the quantum dot layer 33 receives light. A voltage signal V is applied to the common protective electrode layer 32 in During the periodic change, the current I between the common protection electrode layer 32 and the source electrode layer 21 is measured as the voltage V applied to the common protection electrode layer 32 changes. in The change curve of the current I between the common protection electrode layer 32 and the source electrode layer 21 is taken as the neuron discharge rate over time to realize the information encoding of the artificial visual neuron.

[0057] In a specific example of the present invention, the emission rate of the current between the common protection electrode layer 32 and the source electrode layer 21 is measured over time when a single TS and a TS are connected in parallel under different illumination conditions. The more the number of currents emitted over time and the more uniform the distribution, the lower the information loss rate of the artificial visual neuron. The larger the photosensitivity range of the quantum dot layer and the more prominent the absorption peak, the longer and more precise the light tuning range of the artificial neuron. Figure 4 As shown, in the above specific example, the emission rate of the current between the common protection electrode layer 32 and the source electrode layer 21 over time is shown when a single TS and a TS are connected in parallel. The voltage measurement range of the common protection electrode layer 32 is 0 to 1V, which is the voltage range for normal operation of the device. Figure 5 Only the firing rate curve of artificial visual neurons within the range of 300-800nm ​​illumination wavelength is shown. The neurons have different light absorption wavelengths in different quantum dot layers, and thus different firing rates. Taking three quantum dots: lead sulfide quantum dots, perovskite quantum dots, and silicon quantum dots as examples, with light wavelength as the horizontal axis and neuron firing rate as the vertical axis, the optoelectronic bidirectional tuning curve of the artificial visual neuron is shown.

[0058] The above description is merely a preferred embodiment of one or more embodiments of the present specification and is not intended to limit one or more embodiments of the present specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present specification shall be included in the scope of protection of one or more embodiments of the present specification.

Claims

1. An artificial visual neuron with light-tuned coding and no information loss rate, characterized in that: It includes a silicon substrate, an oxide layer and a silicon back gate are formed on the silicon substrate, a channel layer is formed on the oxide layer, a source electrode layer, a quantum dot layer and a drain electrode layer are provided on the channel layer, and the silicon substrate, the oxide layer, the channel layer, the quantum dot layer, the silicon back gate, the source electrode layer and the drain electrode layer constitute a field effect transistor FET structure; two parallel threshold conversion switch TS structures are formed on the drain electrode layer, and each threshold conversion switch structure is formed by stacking a drain electrode layer, an intermediate layer and an active electrode layer; a common protective electrode layer is formed on the active electrode layer.

2. The optically tuned coding artificial visual neuron according to claim 1, characterized in that: The channel layer material is molybdenum disulfide or boron nitride.

3. The optically tuned coding artificial visual neuron according to claim 1, characterized in that: The quantum dot layer material is lead sulfide, perovskite or boron-doped silicon quantum dots.

4. The optically tuned coding artificial visual neuron according to claim 1, characterized in that: The source electrode layer and the drain electrode layer are made of nickel, platinum or gold; the active electrode layer is made of silver or copper.

5. The optically tuned coding artificial visual neuron according to claim 1, characterized in that: The intermediate layer material is aluminum oxide or hafnium oxide.

6. The optically tuned coding artificial visual neuron according to claim 1, characterized in that: The material of the common protective electrode layer is platinum or gold, which covers the active metal that can generate conductive filaments but is unstable, so that the device can work normally while improving stability and life.

7. The optically tuned coding artificial visual neuron according to claim 1, characterized in that: The neuron utilizes the different perception capabilities of the same quantum dot on the channel to light of different wavelengths, and the different perception capabilities of different quantum dots to light of different wavelengths, so that the channel resistance changes under different light illumination, which is manifested as high and low channel currents in transistor performance, thereby affecting the high and low coding currents of the device, making the artificial visual neuron optically tunable; the use of a TS parallel structure can reduce the refractory period of the coding current, so that there is no information loss in the artificial visual neuron.

8. A method for preparing an optically tuned coding artificial visual neuron with no information loss rate as claimed in any one of claims 1 to 7, characterized in that: The steps include: (1) forming an oxide layer on the surface of a silicon substrate, selecting a gate region, and removing the oxide layer in the gate region to form a silicon back gate; (2) mechanically stripping the surface of the oxide layer to form a channel layer, and forming a source electrode layer and a drain electrode layer on the channel layer by photolithography and deposition; (3) forming two parallel TS structure regions on the drain electrode layer by photolithography, then forming an intermediate layer by deposition, then forming an active electrode layer by deposition, and finally forming a common protection electrode layer by deposition; (4) A quantum dot layer is formed on the channel layer by spin coating and annealing, and finally an optically tunable coding artificial visual neuron with no information loss rate based on a 2TS-FET quantum dot structure is formed.

9. A method for using the light-tuned coding artificial visual neuron with no information loss rate as claimed in any one of claims 1 to 7, characterized in that: Light is applied above the quantum dot layer, voltage is applied to the common protection electrode layer, a constant voltage is applied to the silicon back gate, the source electrode layer is grounded, and the curve of the current between the common protection electrode layer and the source electrode layer changing with the wavelength of light and the voltage applied to the common protection electrode layer is measured to achieve optoelectronic bidirectional tunable coding.

10. The method of use according to claim 9, characterized in that: The quantum dot layer is illuminated to achieve a wavelength range of light that enables the quantum dots to absorb light; the common protection electrode layer is applied with a voltage within a voltage range that enables the TS device to work normally; the silicon back gate is applied with a constant voltage within a voltage range that enables the FET device to work in a linear region; the current between the common protection electrode layer and the source electrode layer of the 2TS-FET working under normal excitation is taken as the encoding information.

Images