All-optical intelligent spectrometer based on diffraction neural network

Abstract

The invention relates to an all-optical intelligent spectrometer based on a diffraction neural network. The all-optical intelligent spectrometer comprises an input module, a multi-stage diffraction neural network module and an output module, the input module is used for determining an input light field corresponding to the target multi-wavelength coherent light; the multi-stage diffraction neural network module is used for performing multi-stage phase modulation on the input light field and determining a modulation light field; the output module is used for determining a spectrum reconstruction result corresponding to the target multi-wavelength coherent light according to the modulation light field, and the spectrum reconstruction result is used for reflecting light intensity information of light rays with different wavelengths in the target multi-wavelength coherent light.

Description

Technical Field

[0001] The present disclosure relates to the field of optoelectronics and artificial intelligence technology, and in particular to an all-optical intelligent spectrometer based on a diffraction neural network. Background Art

[0002] Spectral reconstruction technology analyzes a substance's spectral lines to determine its composition and structure, and is widely used in fields such as remote sensing and medical imaging. Existing technologies primarily rely on traditional spectrometers or algorithms based on electronic neural networks. Traditional spectrometers rely on spectroscopic components for spectral separation and acquisition, making them bulky and lacking in integration. Neural network-based algorithms, on the other hand, require complex computational modules to optimize results, resulting in high spectrometer energy consumption. Furthermore, existing spectrometers often struggle to balance spectral resolution and bandwidth, limiting their application scenarios and scalability. Summary of the Invention

[0003] In view of this, the present disclosure proposes a technical solution for an all-optical intelligent spectrometer based on a diffraction neural network.

[0004] According to one aspect of the present disclosure, an all-optical intelligent spectrometer based on a diffraction neural network is provided, and the all-optical intelligent spectrometer includes: an input module, a multi-level diffraction neural network module, and an output module; the input module is used to determine the input light field corresponding to the target multi-wavelength coherent light; the multi-level diffraction neural network module is used to perform multi-level phase modulation on the input light field to determine the modulated light field; the output module is used to determine the spectral reconstruction result corresponding to the target multi-wavelength coherent light based on the modulated light field, wherein the spectral reconstruction result is used to reflect the light intensity information of light of different wavelengths in the target multi-wavelength coherent light.

[0005] In a possible implementation, the input light field includes: a plurality of two-dimensional plane light fields stacked in order of wavelength, wherein the amplitude distribution corresponding to any two-dimensional plane light field is uniform and the phase is fixed.

[0006] In one possible implementation, the modulation function corresponding to the multi-level diffraction neural network module is obtained by training based on a preset joint loss function and a sample spectrum reconstruction data set, wherein the sample spectrum reconstruction data set includes: multiple multi-wavelength coherent light samples, and a reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample.

[0007] In one possible implementation, the multi-level diffraction neural network module is further used to: obtain a predicted spectrum reconstruction result corresponding to any multi-wavelength coherent light sample after the spectrum is reconstructed by the all-optical intelligent spectrometer; iteratively train the modulation function based on the joint loss function, as well as the predicted spectrum reconstruction result and the reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample until the preset training conditions are met, thereby determining a trained modulation function.

[0008] In one possible implementation, the joint loss function includes spectral reconstruction accuracy loss and light intensity distribution contrast loss; wherein the spectral reconstruction accuracy loss is used to reflect the difference in spectral vector intensity between the predicted spectral reconstruction result obtained based on the reconstruction of the all-optical intelligent spectrometer and the reference spectral reconstruction result of the corresponding multi-wavelength coherent light sample; the light intensity distribution contrast loss is used to reflect the difference in light field distribution between the predicted spectral reconstruction result obtained based on the reconstruction of the all-optical intelligent spectrometer and the reference spectral reconstruction result of the corresponding multi-wavelength coherent light sample.

[0009] In a possible implementation, the multi-level diffraction neural network module includes multiple phase modulation units, and the modulation scale corresponding to each phase modulation unit is the same.

[0010] In a possible implementation, the modulation scale corresponding to any phase modulation unit is any one of 400×400, 600×600, or 800×800.

[0011] In a possible implementation, a pixel size corresponding to any phase modulation unit is smaller than or equal to half of a wavelength of a light field received by the phase modulation unit.

[0012] In a possible implementation, the spectral range corresponding to the target multi-wavelength coherent light includes the near-infrared C band; and the surface material of any phase modulation unit is any one of silicon, indium phosphide, or germanium-based materials.

[0013] In one possible implementation, the output module includes multiple detectors, where the number of detectors is equal to the number of light rays of different wavelengths in the target multi-wavelength coherent light; any one of the detectors is used to detect light rays of a wavelength corresponding to the detector in the modulated light field, and perform normalization processing to determine the light intensity information corresponding to the light rays.

[0014] The all-optical intelligent spectrometer of the disclosed embodiment can model the target multi-wavelength coherent light through an input module to determine the input light field corresponding to the target multi-wavelength coherent light; then, using a multi-level diffraction neural network module, perform multi-level phase modulation on the input light field to determine the modulated light field, directly realizing the mapping of the input light field to the modulated light field. This does not require reliance on traditional spectroscopic elements such as gratings and prisms, or complex electronic algorithm modules, and can significantly reduce the energy consumption of the all-optical intelligent spectrometer and improve the speed and accuracy of spectral reconstruction. The output module can determine the spectral reconstruction result corresponding to the target multi-wavelength coherent light based on the modulated light field to accurately reflect the light intensity information of the different wavelengths in the target multi-wavelength coherent light.

[0015] Further features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.

[0017] Figure 1 A block diagram of an all-optical intelligent spectrometer based on a diffraction neural network according to an embodiment of the present disclosure is shown;

[0018] Figure 2 A schematic structural diagram of an all-optical intelligent spectrometer based on a diffraction neural network according to an embodiment of the present disclosure is shown;

[0019] Figure 3 A schematic diagram showing a spectrum reconstruction result according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0020] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.

[0021] As used herein, the terms "comprises," "comprising," "having," or variations thereof are open ended and include one or more stated features, integers, elements, steps, parts, or functions, but do not preclude the presence or addition of one or more other features, integers, elements, steps, parts, functions, or groups thereof.

[0022] When an element is referred to as being "connected," "coupled," "responsive" or variations thereof to another element, it can be directly connected, coupled or responsive to the other element or intervening elements may be present.

[0023] Although the terms first, second, third, etc. may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another element / operation. Therefore, without departing from the teachings of the present invention, the first element / operation in some embodiments may be referred to as the second element / operation in other embodiments.

[0024] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0025] The term "and / or" herein simply describes an association relationship between associated objects, indicating that three relationships can exist. For example, "A and / or B" can represent the existence of three situations: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" herein refers to any combination of at least two of any one or more of a plurality of items. For example, "at least one of A, B, and C" can represent any one or more elements selected from the set consisting of A, B, and C.

[0026] In addition, numerous specific details are provided in the following detailed description to better illustrate the present disclosure. Those skilled in the art will appreciate that the present disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art are not described in detail in order to highlight the main points of the present disclosure.

[0027] Spectral reconstruction technology analyzes a substance's spectral lines to determine its composition and structure, and is widely used in fields such as remote sensing and medical imaging. Existing technologies primarily rely on traditional spectrometers or algorithms based on electronic neural networks. Traditional spectrometers rely on spectroscopic elements (such as gratings and prisms) and detector arrays for spectral separation and acquisition, making them bulky and lacking in integration. Neural network-based algorithms, on the other hand, require complex computational modules to optimize results, resulting in high spectrometer energy consumption.

[0028] On the other hand, spectrometers in the existing technology usually find it difficult to balance spectral resolution and spectral bandwidth. For example, although traditional spectroscopic elements have a high spectral bandwidth, their spectral resolution is limited by their physical size. The spectral resolution and accuracy of electronic neural networks are high, but computing speed is usually sacrificed to ensure accuracy. In addition, the electron-based diffraction neural network (DNN) in the existing technology lacks effective optimization in ultra-high resolution scenarios below 1nm, which limits its application in the field of precision analysis.

[0029] Therefore, due to problems such as structural redundancy, high energy consumption, low spectral resolution, and poor scalability, common spectrometers in the existing technology are difficult to meet the high-precision, low-power, and miniaturized spectral reconstruction requirements in remote sensing, geological exploration, medical imaging, and other fields.

[0030] Photonic Neural Networks (PNNs) are artificial neural networks implemented using photonics technology. Through the high-speed transmission and parallel processing capabilities of light, they can overcome the speed and energy efficiency limitations of traditional electronic computing, reducing energy consumption and latency in data processing, machine learning, and real-time analysis tasks. Furthermore, DNNs are highly compatible with PNNs due to their simplicity, efficiency, high generalization, and scalability. DNNs based on photonics technology typically consist of multiple modulation layers that can adjust the amplitude or phase of the light field. Deep learning methods can be used to refine the specific modulation elements, thereby achieving arbitrary input-output mapping functions.

[0031] However, although PNNs in existing technologies can use optical parallel computing to increase their speed, they usually still require an electronic calibration module to correct errors in the output results. The photoelectric conversion process involved will introduce additional time delays, affecting the real-time processing performance of the PNN. In addition, the training and inference processes of this PNN implementation method also rely on high-performance GPUs, which will also increase energy consumption costs and make it difficult to directly apply to spectrometers.

[0032] In light of this, the present disclosure provides an all-optical intelligent spectrometer based on a diffraction neural network. This spectrometer does not rely on traditional spectroscopic components such as gratings and prisms, or complex electronic algorithm modules. Instead, it utilizes a multi-level diffraction neural network module based on photonics technology to directly map the input light field to the modulated light field. This reduces the energy consumption of the all-optical intelligent spectrometer and improves the speed and accuracy of spectral reconstruction. The following is a detailed introduction to the all-optical intelligent spectrometer based on a diffraction neural network.

[0033] Figure 1 FIG. 1 is a block diagram of an all-optical intelligent spectrometer based on a diffraction neural network according to an embodiment of the present disclosure. Figure 1 As shown, the all-optical intelligent spectrometer 100 includes: an input module 101, a multi-level diffraction neural network module 102, and an output module 103.

[0034] The input module 101 is used to determine the input light field corresponding to the target multi-wavelength coherent light; the multi-level diffraction neural network module 102 is used to perform multi-level phase modulation on the input light field to determine the modulated light field; the output module 103 is used to determine the spectral reconstruction result corresponding to the target multi-wavelength coherent light based on the modulated light field, wherein the spectral reconstruction result is used to reflect the light intensity information of light of different wavelengths in the target multi-wavelength coherent light.

[0035] The target multi-wavelength coherent light may represent coherent light having multiple channels of different wavelengths. The specific form of the target multi-wavelength coherent light may be flexibly set according to actual usage requirements, and the present disclosure does not impose any specific limitation on this.

[0036] In the spatial domain, each channel of the target multi-wavelength coherent light has a uniformly distributed amplitude and a fixed phase, and can be viewed as a superposition of a series of single-wavelength light fields. Therefore, the target multi-wavelength coherent light can be modeled using input module 101 to determine the input light field corresponding to the target multi-wavelength coherent light, resulting in multiple light fields divided by wavelength. This allows spectral reconstruction to recover the light intensity information corresponding to each wavelength light field.

[0037] The implementation method of the input module 101 can refer to the implementation methods in the related art, and this disclosure does not make any specific restrictions on this. The specific form of the input light field can be flexibly set according to actual usage requirements, and this disclosure does not make any specific restrictions on this.

[0038] In a possible implementation, the input light field includes: a plurality of two-dimensional plane light fields stacked in order according to wavelength, wherein the amplitude distribution corresponding to any two-dimensional plane light field is uniform and the phase is fixed.

[0039] Since the target multi-wavelength coherent light has a fixed phase and the equal-phase plane is usually a two-dimensional plane, the input light field obtained by modeling the target multi-wavelength coherent light can be set as multiple two-dimensional plane light fields stacked in order according to the wavelength, and the amplitude distribution corresponding to any two-dimensional plane light field is uniform and the phase is fixed.

[0040] Figure 2 FIG. 1 shows a schematic structural diagram of an all-optical intelligent spectrometer based on a diffraction neural network according to an embodiment of the present disclosure. Figure 2 As shown, the input light field includes multiple two-dimensional plane light fields stacked in order according to wavelength, namely λ1, λ2, λ3 to λ n . Among them, the specific form of the wavelength sequence can be flexibly set according to actual usage requirements. For example, it can be set to the order of wavelengths from small to large, etc., and the present disclosure does not make specific limitations on this. The multi-level diffraction neural network module 102 can be based on photonics technology to realize a multi-level diffraction neural network layer including multiple modulation layers with adjustable light field amplitude or phase. It can be used to perform multi-level phase modulation on the input light field and determine the corresponding modulated light field, thereby directly realizing the mapping of the input light field to the modulated light field. It does not need to rely on traditional spectroscopic elements such as gratings and prisms, or complex electronic algorithm modules. It has low energy consumption and high processing speed.

[0041] The specific form of the multi-level diffraction neural network module 102 can be flexibly configured based on actual usage requirements, and this disclosure does not impose specific limitations on this. The specific form of the light of different wavelengths in the modulated light field can be flexibly configured based on actual usage requirements, and this disclosure does not impose specific limitations on this.

[0042] With the above Figure 2 For example, Figure 2 As shown, the multi-level diffraction neural network module 102 includes diffraction neural network layer 1, diffraction neural network layer 2, diffraction neural network layer 3 to diffraction neural network layer n.

[0043] The multi-level diffraction neural network module 102 will be described in detail later in conjunction with possible implementations of the present disclosure and will not be further described here. The output module 103 can receive the modulated light field from the multi-level diffraction neural network module 102 and detect the light intensity corresponding to each wavelength of light in the modulated light field, thereby determining the spectral reconstruction result corresponding to the target multi-wavelength coherent light, reflecting the light intensity information of the different wavelengths of light in the target multi-wavelength coherent light.

[0044] The specific form of output module 103 can be flexibly configured based on actual usage requirements and is not specifically limited in this disclosure. The specific form of the spectral reconstruction result can also be flexibly configured based on actual usage requirements, for example, it can be configured as a curve fitted by the light intensity of multiple light beams of different wavelengths, and is not specifically limited in this disclosure.

[0045] In one possible implementation, the output module 103 includes multiple detectors, where the number of detectors is equal to the number of light rays of different wavelengths in the target multi-wavelength coherent light; any one detector is used to detect light rays of a wavelength corresponding to the detector in the modulated light field, and perform normalization processing to determine the light intensity information corresponding to the light rays.

[0046] With the above Figure 2 For example, Figure 2 As shown, the output module 103 includes a detector array consisting of multiple detectors, and the number of detectors is equal to the number of different wavelengths of light in the target multi-wavelength coherent light. The specific form of any detector can be referred to the embodiments in the related art, and this disclosure does not specifically limit this.

[0047] Each detector corresponds to a wavelength of light and is used to detect light of that wavelength in the modulated light field. The detection results are normalized to determine the light intensity information corresponding to the light. The specific method of normalization processing can be referred to the implementation methods in the relevant technology and is not specifically limited in this disclosure.

[0048] Based on the multiple detectors included in the output module 103, after determining the light intensity information corresponding to each wavelength of light in the modulated light field, light intensity signal fitting can be performed to obtain the spectrum reconstruction result corresponding to the target multi-wavelength coherent light.

[0049] Figure 3 FIG. 1 is a schematic diagram showing a spectrum reconstruction result according to an embodiment of the present disclosure. Figure 3 As shown in Figure 2, (a) to (d) represent the spectrum reconstruction results corresponding to four different target multi-wavelength coherent lights, and each target multi-wavelength coherent light is a 10-band coherent light. Figure 3 Taking (a) in the figure as an example, the spectrum reconstruction result corresponding to the target multi-wavelength coherent light includes: a light intensity fitting curve corresponding to the target multi-wavelength coherent light, and a light intensity distribution diagram corresponding to the target multi-wavelength coherent light. Specifically, the horizontal axis of the light intensity fitting curve represents the wavelength component of the target multi-wavelength coherent light, the vertical axis represents the normalized light intensity, the orange broken line represents the light intensity fitting curve of each wavelength obtained by the all-optical intelligent spectrometer based on the embodiment of the present disclosure to perform spectral reconstruction on the target multi-wavelength coherent light, and the blue broken line represents the actual light intensity fitting curve of each wavelength in the target multi-wavelength coherent light; the light intensity distribution diagram includes the normalized light intensity corresponding to a wavelength of light determined by each detector in the output module.

[0050] The all-optical intelligent spectrometer of the disclosed embodiment can model the target multi-wavelength coherent light through an input module to determine the input light field corresponding to the target multi-wavelength coherent light; then, using a multi-level diffraction neural network module, perform multi-level phase modulation on the input light field to determine the modulated light field, directly realizing the mapping of the input light field to the modulated light field. This does not require reliance on traditional spectroscopic elements such as gratings and prisms, or complex electronic algorithm modules, and can significantly reduce the energy consumption of the all-optical intelligent spectrometer and improve the speed and accuracy of spectral reconstruction. The output module can determine the spectral reconstruction result corresponding to the target multi-wavelength coherent light based on the modulated light field to accurately reflect the light intensity information of the different wavelengths in the target multi-wavelength coherent light.

[0051] In one possible implementation, the modulation function corresponding to the multi-level diffraction neural network module 102 is obtained by training based on a preset joint loss function and a sample spectrum reconstruction data set, wherein the sample spectrum reconstruction data set includes: multiple multi-wavelength coherent light samples, and a reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample.

[0052] Typically, the multi-level diffraction neural network module 102 performs multiple phase modulations within the range of 0-2π on the input light field using a preset modulation function, so that light of different wavelengths in the modulated light field can be mapped to different positions in the input module 103. To ensure the accuracy and reliability of the all-optical intelligent spectrometer 100, the modulation function corresponding to the multi-level diffraction neural network module 102 can be trained based on a preset joint loss function and a sample spectrum reconstruction dataset.

[0053] Among them, the specific form of the modulation function can be flexibly set according to actual usage requirements. For example, it can include the phase modulation function corresponding to each diffraction neural network layer in the multi-level diffraction neural network module 102, etc. This disclosure does not make specific limitations on this.

[0054] The specific form of the joint loss function can be referred to the implementation methods in the relevant technology, and this disclosure does not make any specific limitations on this.

[0055] The spectral reconstruction data set may include multiple multi-wavelength coherent light samples and a reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample. The specific form of the multi-wavelength coherent light sample depends on the target multi-wavelength coherent light and can be flexibly set according to actual usage requirements. The present disclosure does not make specific restrictions on this. The number of multi-wavelength coherent light samples can be flexibly set according to actual usage requirements. The present disclosure does not make specific restrictions on this. The specific form of the reference spectrum reconstruction result corresponding to any multi-wavelength coherent light sample can be flexibly set according to actual usage requirements. The present disclosure does not make specific restrictions on this.

[0056] In one example, when the spectrum reconstruction result corresponding to the target multi-wavelength coherent light is a curve fitted by the light intensities of multiple light beams of different wavelengths, the reference spectrum reconstruction result corresponding to any multi-wavelength coherent light sample can be set as the spectrum amplitude label corresponding to the multi-wavelength coherent light sample to reflect the light intensity information of the multi-wavelength coherent light sample.

[0057] The specific method of training the modulation function corresponding to the multi-level diffraction neural network module 102 can refer to the implementation methods in the relevant technology, and this disclosure does not make any specific limitations on it.

[0058] In one possible implementation, the multi-level diffraction neural network module 102 is further used to: obtain a predicted spectrum reconstruction result corresponding to any multi-wavelength coherent light sample after the spectrum is reconstructed by the all-optical intelligent spectrometer 100; iteratively train the modulation function based on the joint loss function and the predicted spectrum reconstruction result and reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample until the preset training conditions are met, thereby determining a trained modulation function.

[0059] Specifically, any multi-wavelength coherent light sample is subjected to spectral reconstruction by the all-optical intelligent spectrometer 100, and the predicted spectral reconstruction result corresponding to the multi-wavelength coherent light sample can be determined. Based on the joint loss function, as well as the predicted spectral reconstruction result and the reference spectral reconstruction result corresponding to each multi-wavelength coherent light sample, the modulation function can be iteratively trained, and the parameter values of the modulation function are updated in each round of training until the preset training conditions are met, thereby obtaining a trained modulation function. The specific content of the preset training conditions can be flexibly set according to actual usage requirements. For example, it can include the number of iterative training times meeting the preset number, etc., and the present disclosure does not make specific limitations on this.

[0060] In one possible implementation, the joint loss function includes spectral reconstruction accuracy loss and light intensity distribution contrast loss; wherein, the spectral reconstruction accuracy loss is used to reflect the difference in spectral vector intensity between the predicted spectral reconstruction result obtained based on the all-optical intelligent spectrometer 100 and the reference spectral reconstruction result of the corresponding multi-wavelength coherent light sample; the light intensity distribution contrast loss is used to reflect the difference in light field distribution between the predicted spectral reconstruction result obtained based on the all-optical intelligent spectrometer 100 and the reference spectral reconstruction result of the corresponding multi-wavelength coherent light sample.

[0061] To improve the performance of the all-optical intelligent spectrometer 100, ensure the accuracy and reliability of the spectral reconstruction results, and increase the signal-to-noise ratio of the spectral reconstruction results, constrained optimization can be performed on the spectral reconstruction results in terms of spectral reconstruction accuracy and light intensity distribution contrast. Therefore, the joint loss function can include: spectral reconstruction accuracy loss and light intensity distribution contrast loss.

[0062] Among them, the spectral reconstruction accuracy loss can reflect the difference in spectral vector intensity between the predicted spectral reconstruction result obtained based on the all-optical intelligent spectrometer 100 and the reference spectral reconstruction result of the corresponding multi-wavelength coherent light sample. Its specific form can be flexibly set according to actual usage requirements, and the present disclosure does not make specific limitations on this.

[0063] In one example, the spectral reconstruction accuracy loss can be expressed as formula (1):

[0064] L profile =||A′-A||2 (1)

[0065] Wherein, A′ represents the normalized spectral intensity vector of the predicted spectral reconstruction result of any multi-wavelength coherent light sample reconstructed based on the all-optical intelligent spectrometer 100; A represents the spectral vector intensity corresponding to the reference spectral reconstruction result of the multi-wavelength coherent light sample.

[0066] The light intensity distribution contrast loss can reflect the difference in light field distribution between the predicted spectrum reconstruction result obtained based on the all-optical intelligent spectrometer 100 and the reference spectrum reconstruction result of the corresponding multi-wavelength coherent light sample. Its specific form can be flexibly set according to actual usage requirements, and the present disclosure does not make specific limitations on this.

[0067] In one example, the light intensity distribution contrast loss can be expressed as formula (2):

[0068] L distribution =||II g ||2 (2)

[0069] Wherein, I represents the light intensity distribution of the predicted spectrum reconstruction result of any multi-wavelength coherent light sample obtained by the all-optical intelligent spectrometer 100; I g The actual light intensity distribution represents the reference spectrum reconstruction result of the multi-wavelength coherent light sample.

[0070] In a possible implementation, the multi-level diffraction neural network module 102 includes multiple phase modulation units, and the modulation scale corresponding to each phase modulation unit is the same.

[0071] Specifically, any phase modulation unit can be considered a diffractive neural network layer, performing phase modulation within the range of 0-2π on the input light field during its propagation. The modulation scale corresponding to each phase modulation unit should remain the same to ensure consistency of different phase modulations. The specific form of the phase modulation unit can be referenced in the embodiments of the related art and is not specifically limited in this disclosure.

[0072] The modulation scale corresponding to any phase modulation unit can represent the number of surface pixels of the phase modulation unit. Its specific form can be flexibly set according to actual usage requirements, and this disclosure does not make any specific limitations on this.

[0073] The specific number of phase modulation units included in the multi-level diffraction neural network module 102 can be flexibly set according to actual usage requirements and depends on the number of bands covered by the target multi-wavelength coherent light. This disclosure does not make specific limitations on this.

[0074] In one example, when the target multi-wavelength coherent light is coherent light of 15 wavelength bands, the multi-level diffraction neural network module 102 can be set to include 3 phase modulation units.

[0075] The pixel size corresponding to any phase modulation unit and the spacing between any two adjacent phase modulation units (i.e., the spacing between adjacent diffraction layers of a multi-level diffraction neural network) can be determined based on the maximum half-angle cone theory of the full topological connection of interlayer neurons. The specific method can refer to the implementation methods in the relevant technology, and the present disclosure does not make specific limitations on this.

[0076] Compared with traditional spectrometers or spectrometers based on electronic neural networks, the all-optical intelligent spectrometer of the disclosed embodiment does not require spectroscopic elements and complex electronic computing modules. It can use a multi-level diffraction neural network module composed of a phase modulation unit to directly realize the mapping of the input light field to the modulated light field through a photonic neural network, which can significantly reduce the energy consumption of the all-optical intelligent spectrometer. Experiments show that the energy consumption of spectral reconstruction of common multi-wavelength coherent light is only 1 / 10 of that of a traditional spectrometer; and the speed of spectral reconstruction is improved. Experiments show that the delay of spectral reconstruction of common multi-wavelength coherent light is only 75% of that of the photonic neural network in the prior art, which can meet the real-time requirements while ensuring the accuracy of the spectral reconstruction results.

[0077] In a possible implementation, the modulation scale corresponding to any phase modulation unit is any one of 400×400, 600×600, or 800×800.

[0078] In order to improve the flexibility and scalability of the all-optical intelligent spectrometer and balance the spectral resolution and bandwidth, in the embodiment of the present disclosure, the modulation scale corresponding to any phase modulation unit can be set to any one of 400×400, 600×600 or 800×800, which can achieve high-precision light field control, as well as dynamic adjustment of the spectral resolution in the range of 0.005 nanometers (nm) to 10nm, and dynamic adjustment of the phase quantization in the range of 3 bits (bit) to 8 bits, thereby meeting the usage requirements in different application scenarios.

[0079] By dynamically adjusting the number of phase modulation units in the multi-level diffraction neural network module, the number of diffraction neural network layers can be adjusted. Combined with the adjustment of the modulation scale of each phase modulation unit, multi-dimensional dynamic adjustment of spectral resolution can be achieved, further improving the performance of the all-optical intelligent spectrometer.

[0080] In a possible implementation, a pixel size corresponding to any phase modulation unit is smaller than or equal to half of a wavelength of a light field received by the phase modulation unit.

[0081] To improve the phase modulation capability of each diffractive neural network layer and enhance the dispersion capability of the all-optical intelligent spectrometer, the pixel size corresponding to any phase modulation unit can be further reduced based on the pixel size determined by the maximum half-angle cone theory of the full topological connection of interlayer neurons. Specifically, the pixel size corresponding to any phase modulation unit can be set to be less than or equal to 1 / 2 the wavelength of the light field received by the phase modulation unit.

[0082] Among them, the implementation method of the pixel size corresponding to any phase modulation unit being less than or equal to 1 / 2 of the wavelength of the light field received by the phase modulation unit can refer to the implementation methods in the relevant technology. For example, a diffraction optical element processed by silicon photonics can be used as a phase modulation unit, etc., and the present disclosure does not make specific limitations on this.

[0083] In a possible implementation, the spectral range corresponding to the target multi-wavelength coherent light includes the near-infrared C band; and the surface material of any phase modulation unit is any one of silicon, indium phosphide, or germanium-based materials.

[0084] The applicable spectral range of common spectrometers in the existing technology is usually small and cannot meet the demand for the use of wide spectral bands in fields such as remote sensing and medical imaging. For example, when traditional visible light metasurface materials (such as titanium dioxide, silicon nitride, etc.) process light in the near-infrared band, the refractive index decreases and the absorption loss increases, resulting in reduced phase modulation efficiency.

[0085] Therefore, in the all-optical intelligent spectrometer of the embodiment of the present disclosure, the surface material of any phase modulation unit can use any one of silicon, indium phosphide, or germanium-based materials to increase the modulation efficiency of the phase modulation unit, so that the spectral range corresponding to the target multi-wavelength coherent light can include the near-infrared C band, and can achieve a high resolution of 0.05nm, which can fully meet the usage requirements in remote sensing, medical imaging and other fields.

[0086] The all-optical intelligent spectrometer of the disclosed embodiment can model the target multi-wavelength coherent light through an input module to determine the input light field corresponding to the target multi-wavelength coherent light; then use a multi-level diffraction neural network module to perform multi-level phase modulation on the input light field to determine the modulated light field, without relying on traditional spectroscopic elements such as gratings and prisms, or complex electronic algorithm modules. The multi-level diffraction neural network module composed of phase modulation units can be used to directly realize the mapping of input light field to modulated light field through a photonic neural network, thereby reducing the energy consumption of the all-optical intelligent spectrometer and improving the speed and accuracy of spectral reconstruction; by dynamically adjusting the number of phase modulation units in the multi-level diffraction neural network module, the number of diffraction neural network layers can be adjusted, and combined with the adjustment of the modulation scale of each phase modulation unit, multi-dimensional dynamic adjustment of spectral resolution can be achieved, further improving the performance of the all-optical intelligent spectrometer. The output module can determine the spectral reconstruction result corresponding to the target multi-wavelength coherent light based on the modulated light field to accurately reflect the light intensity information of different wavelengths in the target multi-wavelength coherent light.

[0087] It should be noted that although Figure 1 and Figure 2 The structure of the all-optical intelligent spectrometer disclosed herein, as well as partial structures of the multi-level diffraction neural network module and output module, are described above as examples. However, those skilled in the art will appreciate that the present disclosure is not limited thereto. In fact, users can flexibly set the specific structural form of the all-optical intelligent spectrometer based on their personal preferences and / or actual application scenarios. As long as the multi-level diffraction neural network module based on photonic technology can be used to directly implement the mapping of the input light field to the modulated light field, the energy consumption of the all-optical intelligent spectrometer can be reduced, and the speed and accuracy of spectral reconstruction can be improved.

[0088] While various embodiments of the present disclosure have been described above, the foregoing description is intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An all-optical intelligent spectrometer based on diffraction neural network, characterized in that: The all-optical intelligent spectrometer includes: an input module, a multi-level diffraction neural network module, and an output module; The input module is used to determine the input light field corresponding to the target multi-wavelength coherent light; The multi-level diffraction neural network module is used to perform multi-level phase modulation on the input light field to determine the modulated light field; The output module is used to determine a spectrum reconstruction result corresponding to the target multi-wavelength coherent light according to the modulated light field, wherein the spectrum reconstruction result is used to reflect the light intensity information of light of different wavelengths in the target multi-wavelength coherent light.

2. The all-optical intelligent spectrometer according to claim 1, characterized in that: The input light field includes: a plurality of two-dimensional plane light fields stacked in order according to wavelength, wherein the amplitude distribution corresponding to any two-dimensional plane light field is uniform and the phase is fixed.

3. The all-optical intelligent spectrometer according to claim 1 or 2, characterized in that: The modulation function corresponding to the multi-level diffraction neural network module is obtained by training based on a preset joint loss function and a sample spectrum reconstruction data set, wherein the sample spectrum reconstruction data set includes: multiple multi-wavelength coherent light samples, and a reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample.

4. The all-optical intelligent spectrometer according to claim 3, characterized in that: The multi-level diffraction neural network module is further used to: After any multi-wavelength coherent light sample is subjected to spectrum reconstruction by the all-optical intelligent spectrometer, a predicted spectrum reconstruction result corresponding to the multi-wavelength coherent light sample is obtained; Based on the joint loss function, and the predicted spectrum reconstruction result and the reference spectrum reconstruction result corresponding to each multi-wavelength coherent light sample, the modulation function is iteratively trained until a preset training condition is met, and a trained modulation function is determined.

5. The all-optical intelligent spectrometer according to claim 4, characterized in that: The joint loss function includes spectral reconstruction accuracy loss and light intensity distribution contrast loss; The spectrum reconstruction accuracy loss is used to reflect the difference in spectrum vector intensity between the predicted spectrum reconstruction result obtained based on the all-optical intelligent spectrometer and the reference spectrum reconstruction result of the corresponding multi-wavelength coherent light sample; The light intensity distribution contrast loss is used to reflect the light field distribution difference between the predicted spectrum reconstruction result obtained based on the all-optical intelligent spectrometer and the reference spectrum reconstruction result of the corresponding multi-wavelength coherent light sample.

6. The all-optical intelligent spectrometer according to claim 1 or 2, characterized in that: The multi-level diffraction neural network module includes multiple phase modulation units, and the modulation scale corresponding to each phase modulation unit is the same.

7. The all-optical intelligent spectrometer according to claim 5, characterized in that: The modulation scale corresponding to any phase modulation unit is any one of 400×400, 600×600 or 800×800.

8. The all-optical intelligent spectrometer according to claim 5, characterized in that: The pixel size corresponding to any phase modulation unit is smaller than or equal to 1 / 2 of the wavelength of the light field received by the phase modulation unit.

9. The all-optical intelligent spectrometer according to claim 5, characterized in that: The spectral range corresponding to the target multi-wavelength coherent light includes the near-infrared C band; The surface material of any phase modulation unit is any one of silicon, indium phosphide, or germanium-based materials.

10. The all-optical intelligent spectrometer according to claim 1 or 2, characterized in that: The output module includes a plurality of detectors, wherein the number of the detectors is equal to the number of light rays of different wavelengths in the target multi-wavelength coherent light; Any one of the detectors is used to detect light of a wavelength corresponding to the detector in the modulated light field, and perform normalization processing to determine light intensity information corresponding to the light.

Images