A reconfigurable electro-optical nonlinear activation function device

Abstract

The present invention provides a reconfigurable electro-optical nonlinear activation function device, comprising: an input splitter module, a control module and a reconfigurable nonlinear output module; the input splitter module adopts a directional coupler to split the input optical signal to obtain a low-power splitter signal and a remaining splitter signal; the control module comprises: a photodetector, a linear transimpedance amplifier TIA and an external voltage; the photodetector is used to convert the low-power splitter signal into a current signal; the linear transimpedance amplifier TIA is used to amplify the current signal into a first voltage signal; the first voltage signal and the external voltage are combined to obtain a second voltage signal; the reconfigurable nonlinear output module adopts an All-Pass type micro-ring resonator doped with a PN junction and embedded with a non-volatile phase change material Sb2Se3; the All-Pass type micro-ring resonator is used to modulate the remaining splitter signal according to the second voltage signal to obtain the output result of the activation function device. The present invention is expected to cope with the application of low-threshold reconfigurable optical nonlinear activation functions in ONN, thereby improving the adaptability and performance of optical neural networks.

Description

Technical Field

[0001] This invention belongs to the field of optical signal processing, and in particular relates to a reconfigurable electro-optic nonlinear activation function device. Background Technology

[0002] With the rapid development of information technology, Artificial Neural Networks (ANNs), as computational models that mimic biological nervous systems, have been widely applied in various fields. However, traditional manufacturing processes are gradually approaching their limits, Moore's Law for traditional electronic integrated circuits is becoming increasingly ineffective, and electronic chips face some intractable bottlenecks in further improving computing speed and reducing power consumption. Simultaneously, with the continuous increase in information capacity and the throughput limitations of general-purpose electronic processors in parallel processing, traditional electrical neural networks cannot effectively execute highly complex artificial intelligence tasks. Optical Neural Networks (ONNs), due to their high parallelism and high modulation speed, offer higher efficiency and speed while reducing energy consumption. Therefore, optical neural networks are expected to become an important development direction for future data processing hardware, bringing faster and more efficient data processing capabilities to more fields.

[0003] Neurons in a neural network consist of linear computational units and activation functions (AFs). When an artificial neural network architecture lacks nonlinear activation, regardless of the number of hidden layers, it simply performs linear combinations of inputs, resulting in a weak ability to represent data. Introducing nonlinear activation functions maps the inputs to the output nonlinearly, allowing the neural network to approximate any nonlinear function and enhancing its data representation capabilities. This capability stems from the finiteness and invariance of network complexity, enabling neural networks to learn complex mappings between inputs and outputs, thus solving complex real-world problems. However, most research on optical neural networks focuses on linear matrix-vector multiplication. Many mature optical linear matrix multiplication schemes have emerged in free-space diffraction elements or integrated photonic circuits, such as Mach-Zehnder interferometer (MZI) grids, micro-ring resonator (MRR) matrices, and phase-change material-assisted matrix multiplication units. Research on another crucial component of ANNs—nonlinear activation—is relatively limited. However, nonlinear activation plays a vital role in the network. Applying nonlinear activation functions to the output after the MultiplyAccumulate (MAC) operation remains a challenge in optical neural network research.

[0004] Achieving nonlinearity in optical hardware platforms is more challenging, partly because optical nonlinearities are relatively weak. This means that larger interaction lengths and higher signal power are required to achieve the desired effect. This not only impacts the physical size of the device but also increases energy consumption and raises the nonlinearity threshold, both of which limit the design and performance of ONNs. A more fundamental limitation of optical nonlinearities is that their responses tend to become fixed during device fabrication. This limited tunability of nonlinear optical responses prevents ONNs from being reprogrammed to implement different forms of nonlinear activation functions, severely restricting their flexibility and tunability. Summary of the Invention

[0005] To address the aforementioned technical challenges and enhance the adaptability and generalization ability of neural network models, this invention utilizes a directional coupler to split the input optical signal. The low-power light is amplified through photoelectric conversion and then transformed into a control signal for the remaining majority of the signal power. An input splitting module and a control module are designed to achieve a low activation threshold without increasing the optical signal power. By leveraging the characteristics of the output spectrum altered by applying voltage to a micro-ring resonator and changing the crystallinity (effective refractive index) of the phase change material, along with the non-volatile nature of the phase change material, a reconfigurable nonlinear output module structure is designed. This enables a reconfigurable nonlinear relationship between the input and output optical signals, achieving a low-threshold, reconfigurable optical nonlinear activation function, thereby improving the adaptability and generalization ability of the neural network model.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:

[0007] A reconfigurable electro-optic nonlinear activation function device includes: an input beam splitting module, a control module, and a reconfigurable nonlinear output module;

[0008] The input beam splitter module employs a directional coupler to split the input optical signal P in The low-power spectrophotometer αP was obtained by performing spectral splitting. in and the remaining spectroscopic signal (1-α)P in ;

[0009] The control module includes: a photodetector, a linear transimpedance amplifier (TIA), and an external voltage V. b ;

[0010] The photodetector is used to split the low-power spectral signal αP in The current signal is converted into a current signal; the linear transimpedance amplifier TIA is used to amplify the current signal into a first voltage signal; the first voltage signal and the external voltage V b The second voltage signal is obtained by merging the signals.

[0011] The reconfigurable nonlinear output module employs an All-Pass type microring resonator doped with a PN junction and embedded with the non-volatile phase change material Sb2Se3.

[0012] The All-Pass type microring resonator is used to split the remaining beam signal (1-α)P according to the second voltage signal. in The output of the activation function device is obtained by modulation.

[0013] Preferably, the photodetector is a photodiode.

[0014] Preferably, the second voltage signal is applied to the PN junction of the All-Pass type micro-ring resonator as a control signal.

[0015] Preferably, the All-Pass type microring resonator includes: a substrate, and a straight waveguide and a ring waveguide disposed on the substrate, wherein the straight waveguide is arranged laterally; the ring waveguide is disposed in front of or behind the straight waveguide, and the straight waveguide and the ring waveguide have a coupling region; on the left side of the ring waveguide, an N-type node is doped along the circumference of the ring waveguide, and a P-type node is disposed in the inner circumferential region of the ring waveguide, forming the PN junction of the All-Pass type microring resonator.

[0016] Preferably, a non-volatile phase change material Sb2Se3 layer is disposed on the upper right end of the ring waveguide along the circumference of the ring waveguide.

[0017] Preferably, the width of the non-volatile phase change material Sb2Se3 layer is the same as the width of the ring waveguide; the length of the non-volatile phase change material Sb2Se3 layer is 0.01 to 0.2 times the circumference of the ring waveguide.

[0018] The present invention has at least the following beneficial effects

[0019] This invention proposes a reconfigurable electro-optic nonlinear activation function device. By utilizing signal amplification in the electronic domain, the low-power split signal obtained after splitting the input optical signal is converted to the electronic domain and amplified as a control signal. Signal amplification in the electronic domain overcomes the need for high optical signal power, achieving a lower activation threshold without increasing optical signal power, thereby reducing the demand for high optical signal power and improving the performance and efficiency of the ONN. Furthermore, since most of the signal power (the remaining split signal) remains in the optical domain, there is no need for a new light source at each nonlinear layer of the network. In addition, a non-volatile phase change material, Sb₂Se₃, is introduced and embedded in the MRR. By utilizing the non-volatile properties of Sb₂Se₃, different crystallinities of Sb₂Se₃ are switched to achieve reconfigurable nonlinear modulation. Using a low-power nanosecond laser, rapid and reversible switching of Sb₂Se₃ between amorphous and crystalline states can be achieved. Considering the non-volatile properties of Sb₂Se₃, it can maintain its state even without a power source. The use of non-volatile phase-change materials to achieve reconfigurable nonlinear modulation gives the system greater flexibility and tunability. This means the structure can adapt to different types of data processing tasks, thus achieving better adaptability and generalization ability. Therefore, this invention is expected to address the application of low-threshold reconfigurable optical nonlinear activation functions in ONNs, thereby improving the adaptability and performance of optical neural networks. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the reconfigurable electro-optic nonlinear activation function device of the present invention;

[0021] Figure 2 The cross-section of the MRR structure of the non-volatile phase change material Sb2Se3 embedded in the present invention and the transmittance of microring resonators with different crystallinities of Sb2Se3 after doping PN junction are shown.

[0022] Figure 3 This is a schematic diagram illustrating the implementation process of the reconfigurable nonlinear activation function in an embodiment of the present invention; Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0025] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0026] Please see Figure 1 The present invention provides a reconfigurable electro-optic nonlinear activation function device, comprising: an input beam splitting module, a control module, and a reconfigurable nonlinear output module;

[0027] The input beam splitter module employs a directional coupler to split the input optical signal P in The low-power spectrophotometer αP was obtained by performing spectral splitting. in and the remaining spectroscopic signal (1-α)P in ;

[0028] The control module includes: a photodetector, a linear transimpedance amplifier (TIA), and an external voltage V. b ;

[0029] The photodetector is used to split the low-power spectral signal αP in The current signal is converted into a current signal; the linear transimpedance amplifier TIA is used to amplify the current signal into a first voltage signal; the first voltage signal and the external voltage V b The second voltage signal is obtained by merging the signals.

[0030] The reconfigurable nonlinear output module employs an All-Pass type microring resonator doped with a PN junction and embedded with the non-volatile phase change material Sb2Se3.

[0031] The All-Pass type microring resonator is used to split the remaining beam signal (1-α)P according to the second voltage signal. in The output of the activation function device is obtained by modulation.

[0032] In this embodiment, Figure 1 (a) is a schematic diagram of a reconfigurable electro-optic nonlinear activation function device proposed in this invention. In this embodiment, it is assumed that no additional loss is generated during the transmission and coupling process of the directional coupler, and its scattering matrix S is:

[0033]

[0034] Where k is the coupling coefficient, z is the effective coupling length of the coupler, and t1 and k1 represent the self-coupling coefficient and mutual coupling coefficient of the coupler, respectively. Without considering losses, the two coefficients satisfy: t1 2 +k1 2 =1; ρ is the distribution ratio of the optical signal between the two output ports O1 and O2.

[0035] The relationship between the optical signal energy at the output port and the optical signal energy at the input port of the directional coupler can be represented by a transmission matrix:

[0036]

[0037] Here, I1 and I2 represent the input ports. The splitting ratio of the directional coupler, i.e. the ratio of the intensity of the output signal, is adjusted by designing the structural dimensions.

[0038] The lower input port I2 of the directional coupler is the input signal light P in The output port O1 outputs a low-power split signal αP. in The output O2 from the lower port is the remaining split signal (1-α)P. in .

[0039] It is understood that the size of α here can be set arbitrarily, and those skilled in the art can determine the corresponding size according to the actual situation. This invention does not impose specific limitations on this.

[0040] In this embodiment of the invention, the control module includes: a photodetector, a linear transimpedance amplifier (TIA), and an external voltage V. b ; αP output from the upper port of the directional coupler in The current signal is obtained after conversion by a balanced photodiode, and then converted into a first voltage signal V by a linear transimpedance amplifier (TIA). TIA First voltage signal V TIA and external DC voltage V b Together they serve as the input second voltage signal V of the PN junction of the micro-ring modulator. sum When the input voltage V of the PN junction of the micro-ring modulator sum Less than the threshold voltage V thWhen the (micro-ring PN junction's) on-state voltage is reached, the PN junction remains closed, and no charge carriers are injected into the PN junction. When the input V... sum Exceeding the threshold voltage V th When the PN junction is turned on, the refractive index of the microring resonator changes as charge carriers are injected into the PN junction, causing the resonant wavelength of the microring modulator to shift, thus achieving nonlinear control of the input optical signal of the microring resonator.

[0041] Preferably, the photodetector is a photodiode.

[0042] Preferably, the second voltage signal is applied to the PN junction of the All-Pass type micro-ring resonator as a control signal.

[0043] The optical signal (1-α)P output from the lower port O2 of the directional coupler in As the input optical signal of the All-Pass microring resonator, the voltage signal V TIA and external DC voltage V b Together they serve as the control input (second voltage signal V) of the PN junction of the micro-ring modulator. sum The first voltage signal V TIA and external DC voltage V b (Added together); with the input optical signal P in The increase of αP at the output port of the directional coupler in Increase, the first voltage signal V after photoelectric conversion TIA As it increases, the input second voltage signal V of the PN junction eventually becomes... sum With the input optical signal P in The PN junction input voltage V increases with the increase of [something]. sum An increase in the value of (1-α)P will cause a shift in the output spectrum of the microring resonator, i.e., the input optical signal (1-α)P of the microring resonator will shift. in With output optical signal P out It generates nonlinear changes, realizing the output of the nonlinear function of the input optical signal and the output optical signal.

[0044] Preferably, the All-Pass type microring resonator includes: a substrate, and a straight waveguide and a ring waveguide disposed on the substrate, wherein the straight waveguide is arranged laterally; the ring waveguide is disposed in front of or behind the straight waveguide, and the straight waveguide and the ring waveguide have a coupling region; on the left side of the ring waveguide, an N-type node is doped along the circumference of the ring waveguide, and a P-type node is disposed in the inner circumferential region of the ring waveguide, forming the PN junction of the All-Pass type microring resonator.

[0045] This invention provides an embodiment in which an N-type node is disposed on the left side of the ring waveguide along the circumference of the ring waveguide; an Sb2Se3 layer is disposed on the upper surface of the right end of the ring waveguide along the circumference of the ring waveguide; similarly, an N-type node can also be disposed on the right side of the waveguide along the circumference of the ring waveguide; and an Sb2Se3 layer is disposed on the upper surface of the left end of the ring waveguide along the circumference of the ring waveguide.

[0046] Preferably, the thickness of the Sb₂Se₃ layer is 5–40 nm. In this embodiment, the substrate includes a Si (silicon) buried substrate and a SiO₂ buried oxide layer stacked sequentially. A ring waveguide and a straight waveguide are disposed on the SiO₂ buried oxide layer. Both the straight waveguide and the ring waveguide include a Si waveguide layer and a ridge disposed on the Si waveguide layer. The width of the ridge (material: Si (silicon)) is equal to the width of the Sb₂Se₃ layer. The thickness of the SiO₂ buried oxide layer is 3.3 μm; the thickness of the Si waveguide layer is 70 nm; the thickness of the ridge is 150 nm; the width of the Sb₂Se₃ layer is 450 nm; the thickness of the Sb₂Se₃ layer is 5–40 nm, and in this embodiment it is set to 30 nm. This is merely one embodiment of the present invention. For those skilled in the art, it is possible to adjust the above parameters through limited experiments based on an understanding of this solution, all of which fall within the protection scope of the present invention.

[0047] Figure 1 Figure (b) shows a schematic diagram of the PN junction doping of an All-Pass microring resonator doped with a PN junction and embedded with the non-volatile phase-change material Sb₂Se₃. N-doping is performed in the central region of the MRR (equivalent to the inner circumference of the ring waveguide), and p-doping is performed at the other end (equivalent to an N-type junction doped along the circumference of the ring waveguide on the left side). When a forward bias voltage is applied to the center (N region) of the All-Pass microring resonator, a redshift occurs at the resonant frequency. This is because the forward bias causes current to flow from the P region to the N region, increasing the carrier concentration in the P region and decreasing the carrier concentration in the N region. This changes the refractive index near the PN junction, increasing the speed of light propagation in the microring, thus causing the resonant frequency to shift towards a higher frequency (i.e., a redshift). Figure 1 As shown in (c), Figure 1 (c) represents a cross-sectional view of an MRR structure with a doped PN junction.

[0048] Preferably, a non-volatile phase change material Sb2Se3 layer is disposed on the upper right end of the ring waveguide along the circumference of the ring waveguide.

[0049] Preferably, the width of the non-volatile phase change material Sb2Se3 layer is the same as the width of the ring waveguide; the length of the non-volatile phase change material Sb2Se3 layer is 0.01 to 0.2 times the circumference of the ring waveguide.

[0050] Figure 2 The cross-section of the MRR structure of the embedded non-volatile phase change material Sb2Se3 proposed in this invention and the transmittance of microring resonators with different crystallinities of Sb2Se3 after doping PN junction are shown.

[0051] Figure 2 (a) shows a schematic diagram of the cross-section of an all-pass microring resonator with embedded Sb2Se3. It can be seen that a 30nm thick non-volatile phase change material Sb2Se3 is embedded on the through-type ring resonator and is the same width as the ring waveguide, so it does not increase the additional area.

[0052] Figure 2 Figure (b) shows the effect of different crystallization levels of Sb2Se3 on the output spectrum of the All-Pass microring resonator after doping the PN junction. Different crystallization levels correspond to the resonant wavelength of the initial All-Pass microring resonator (when a voltage is applied to the PN junction). Therefore, the initial resonant wavelength can be shifted by changing the crystallization rate of Sb2Se3.

[0053] In summary, a non-volatile phase change material, Sb₂Se₃, was introduced and embedded in an All-Pass type microring resonator. By utilizing its non-volatility and the high contrast between the amorphous and crystalline phases, the effective refractive index of the microring resonator was changed by altering the state of Sb₂Se₃, thereby shifting the resonant wavelength of the initial microring modulator and achieving reconfigurability of the nonlinear activation function.

[0054] Figure 3 This is a schematic diagram illustrating the implementation process of the reconfigurable electro-optic nonlinear activation function device proposed in this invention. Assume the wavelength of the input signal light is λ. laser The resonant wavelength of the micro-ring modulator is λ res .

[0055] Figure 3 Image (a) is a schematic diagram of the implementation process of the ReLU function proposed in this invention. When the wavelength of the input signal light is λ laser With the resonant wavelength λ of the All-Pass type micro-ring resonator res When consistent, when V sum <V th At this time, the PN junction remains closed, and no charge carriers are injected into the PN junction. Because the input light is filtered by the notch response of the All-Pass type microring resonator, the output power P of the microring modulator is... out Almost zero; when V sum >V th When the PN junction is turned on, the refractive index of the All-Pass microring resonator changes as charge carriers are injected into the PN junction, causing a shift in the resonant wavelength of the All-Pass microring resonator and increasing the optical output power P.out In conclusion, when V sum <V th At that time, the output power P of the All-Pass type micro-ring resonator out Almost zero; when V sum >V th At that time, the output power P of the All-Pass type micro-ring resonator out With V sum P increases with the increase of out Follow as V sum The ReLU function characteristics. And because V sum =V TIA +V b It can be achieved by changing the external voltage V b To adjust the ReLU threshold (V) th ).

[0056] Figure 3 Image (b) is a schematic diagram of the implementation process of the ELU function proposed in this invention. When the wavelength of the input signal light is λ laser With the resonant wavelength λ of the All-Pass type micro-ring resonator res When they match, set the external voltage V. b The threshold voltage V of the PN junction th V b =V th With the input optical power P in With the increase of αP, the output αP of the directional coupler at the upper port in Increase, the voltage V after photoelectric conversion and amplification TIA It is also increasing, so the modulation voltage V applied to the PN junction in the All-Pass microring resonator is... sum With the input optical power P in The value increases with the increase of [the value]. The output of the lower port of the directional coupler is (1-α)P. in Also varies with the input optical power P in The power increases with the increase of [something]. Therefore, the output power P of the All-Pass type micro-ring resonator [is] [increases]. out With V sum P increases with the increase of out Follow as V sum The characteristics of the ELU function.

[0057] Figure 3 Image (c) is a schematic diagram of the RBF function implementation process proposed in this invention. By switching the state of the non-volatile phase change material Sb₂Se₃, i.e., changing the effective refractive index of the All-Pass microring resonator, the resonant wavelength λ of the initial All-Pass microring resonator is changed. res The offset occurs, causing the resonant wavelength λ of the All-Pass type microring resonator to change.res With the input signal light wavelength λ laser If inconsistent, set the external voltage V. b For [0, V th As the optical input power P in The increase of P out Follow as V sum The radial basis function properties.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A reconfigurable electro-optical nonlinear activation function device, characterized in that: include: Input optical splitting module, control module and reconfigurable nonlinear output module; The input optical splitter module uses a directional coupler to convert the input optical signal Split the light to get a low-power split signal and the remaining split signal ; The control module includes: a photodetector, a linear transimpedance amplifier TIA and an external voltage ; The photodetector is used to split the low-power light signal The linear transimpedance amplifier TIA is used to amplify the current signal into a first voltage signal; the first voltage signal and the external voltage Combining to obtain a second voltage signal; The reconfigurable nonlinear output module adopts an All-Pass type micro-ring resonator doped with a PN junction and embedded with a non-volatile phase change material Sb2Se3; The All-Pass type micro-ring resonator comprises: a substrate, and a straight waveguide and a ring waveguide arranged on the substrate, wherein the straight waveguide is arranged in a transverse direction; the ring waveguide is arranged in front of or behind the straight waveguide, and the straight waveguide and the ring waveguide have a coupling region; on the left side of the ring waveguide, an N-type node is mixed along the circumference of the ring waveguide, and a P-type node is arranged in the inner circumference area of ​​the ring waveguide, forming a PN junction of the All-Pass type micro-ring resonator; A non-volatile phase change material Sb2Se3 layer is arranged on the right end of the annular waveguide along the circumference of the annular waveguide; The All-Pass type microring resonator is used to modulate the remaining split light signal according to the second voltage signal to obtain the output result of the activation function device.

2. A reconfigurable electro-optical nonlinear activation function device according to claim 1, characterized in that: The photodetector is a photodiode.

3. A reconfigurable electro-optical nonlinear activation function device according to claim 1, characterized in that: The second voltage signal is applied to the PN junction of the All-Pass type microring resonator as a control signal of the All-Pass type microring resonator.

4. A reconfigurable electro-optical nonlinear activation function device according to claim 1, characterized in that: The width of the nonvolatile phase-change material Sb2Se3 layer is the same as the width of the ring waveguide; the length of the nonvolatile phase-change material Sb2Se3 layer is 0.01 to 0.2 times the circumference of the ring waveguide.

Images