Monolithic integrated computing reflectance spectrometer and spectral acquisition and reconstruction method

Abstract

The application discloses a monolithic integrated computing reflectance spectrometer and a spectrum acquisition and reconstruction method, which comprises a monolithic integrated optoelectronic device, and double-wavelength LEDs and PDs are integrated on the same substrate based on gallium nitride material; the LEDs realize dynamic wavelength tuning through current regulation, forming active spectral coding; the PDs fixedly receive sample reflected light and output photoelectric current, without the need of receiving end filtering, so that light intensity and signal-to-noise ratio are improved. The system comprises a current driving module for regulating LED wavelength, and a processing module for combining a PD photoelectric current sequence and LED emission spectrum to reconstruct sample reflectance spectrum. Physical constraints and data driving can be fused, data dependence is reduced, and high-precision reconstruction is realized. The design is small in size, high in integration and low in power consumption, and is suitable for wearable, biomedical detection and other scenes, and has strong industrialization potential.

Description

Technical Field

[0001] This invention relates to spectrometers, and more particularly to a monolithic integrated computational reflectance spectrometer and a method for spectral acquisition and reconstruction. Background Technology

[0002] With the widespread application of spectral analysis in fields such as biomedicine, environmental monitoring, and industrial testing, the demand for miniaturized, highly stable, and low-power spectral detection equipment is increasing. Traditional spectroscopic instruments are usually built based on dispersive elements (such as gratings and prisms) and spatial scanning or detection arrays. Although they have high resolution and high sensitivity, their complex optical path structure and mechanical components result in large system size, high cost, and high power consumption, making them unsuitable for integrated or portable applications.

[0003] In recent years, miniaturized computational spectrometers have attracted widespread attention from researchers as a novel spectral detection architecture. These systems achieve compressed sampling and algorithmic reconstruction of incident spectra by working in conjunction with optical coding structures (such as filter arrays, speckle structures, and nano-optical devices), breaking through the traditional limitations of balancing size and performance in spectrometers.

[0004] Existing computational spectrometers can be mainly divided into three categories: those based on spatial coding, time coding, and light source coding.

[0005] Spatial coding refers to the spatial separation and measurement of spectral responses at different wavelengths. This method encompasses various structures. Among them, speckle coding structures generate characteristic speckle patterns using random optical structures (such as photonic chips, helical waveguides, and tapered optical fibers), with different spatial intensity distributions corresponding to different wavelengths. Typical devices have achieved sub-nanometer resolution. Filter array structures utilize filter units fabricated from materials such as quantum dots, metasurfaces, and resonant cavities to encode spectral information in a two-dimensional array, which is then read by an image sensor. Devices based on CdSe quantum dots, freeform meta-atoms, and Fabry-Pérot cavities have been widely reported. Detector array coding, on the other hand, achieves direct coding without filters by designing photodetector units with different wavelength response characteristics (such as GaN nanowires and gradient bandgap nanowires), further promoting the monolithic integration and miniaturization of systems.

[0006] Time-modulated spectral coding significantly reduces the size of computational spectrometers by measuring the response of different wavelengths of the spectrum in a time-division manner. This technique primarily uses laser pulses, voltage pulses, or thermal energy to modulate the response of filters or photodetectors, achieving spectral coding in the time dimension. Systems based on tunable filters utilize structures such as Fabry-Pérot cavities, phase change materials, waveguide resonators, and metasurfaces to dynamically sample the spectrum by adjusting the filter's wavelength response to external stimuli, achieving high resolution and rapid measurement. These systems have high integration but require high precision and speed in modulation. Schemes based on tunable photodetectors complete the coding by adjusting the spectral response of a single detector; typical materials include black phosphorus, perovskite, and two-dimensional van der Waals heterojunctions. These methods are smaller, have faster response, and avoid the sensitivity of filters to the incident angle, but require high stability in detector modulation.

[0007] Light source modulation achieves temporal or spatial encoding by controlling the emission of a series of known spectra from a light source. Common methods include multi-band LEDs, quantum dot (QD) arrays, organic light-emitting materials, and bandgap gradient semiconductor films. The encoded light is transmitted through the sample and collected by a fixed detector to reconstruct the spectrum. Existing research has realized light source modulation systems in the 400–1400 nm range, which have the advantages of compact structure and no need for spectroscopic devices. For example, QD arrays can directly realize multi-peak emission and also have spectroscopic functions; VO2 phase change materials can be used for thermally tunable emission in the infrared band; ZnCdSeS bandgap gradient films generate linearly tunable spectra through ultraviolet excitation for chip-level integrated spectral detection. However, these methods usually rely on transmission structures, requiring samples with good light transmittance, and are limited by the efficiency and stability of the light-emitting materials. Light source modulation still has certain limitations in achieving high precision, wide spectral band, and real-time measurement.

[0008] Despite significant progress in recent years in terms of compact design, low cost, and integration of computational spectrometers, existing technologies still have the following shortcomings:

[0009] 1. Computational spectrometers are passive designs that rely on receiver adjustments (such as filter arrays, tunable filters, PD arrays, etc.) to achieve spectral separation and encoding. They require active broadband light sources, and existing integrated chips mostly only integrate the receiving part (filters and PDs), without achieving collaborative integration of the light source and the receiving module. This type of structure usually requires multiple discrete components and complex optical paths, which not only makes it difficult to achieve monolithic integration, but also leads to light energy loss and reduced signal-to-noise ratio due to receiver filtering, limiting its application in wearable and portable terminal scenarios.

[0010] 2. For tunable light source spectrometers, array-type light source structures are currently commonly used for spectral encoding. However, due to inconsistencies in the emission angle, intensity, and coupling efficiency of different light-emitting units, the spectral energy distribution incident on the analyte becomes unstable. Furthermore, this also presents problems such as structural dispersion and difficulty in system integration.

[0011] 3. Existing computational spectrometers mainly rely on traditional data-driven methods in terms of algorithms, failing to effectively integrate the physical processes in spectral measurements. This results in poor model interpretability, weak robustness to changes in light source and detector response, and reconstruction accuracy that depends on a large amount of calibration data. Furthermore, they are difficult to generalize and apply to different devices or scenarios.

[0012] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0013] The main objective of this invention is to overcome the deficiencies in the aforementioned background technology and provide a monolithic integrated computational reflectance spectrometer and a method for spectral acquisition and reconstruction.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] A monolithic integrated computational reflectance spectrometer, comprising:

[0016] A monolithic integrated optoelectronic device includes a dual-wavelength light-emitting diode (LED) and a photodiode (PD). The LED emits a wavelength-tunable light signal through current modulation, and the PD receives the reflected signal from the sample and outputs a photocurrent. The LED and PD are integrated on the same substrate to form a monolithic structure that coordinates light source emission and reflection reception. Active spectral encoding is achieved through dynamic wavelength tuning of the LED.

[0017] A current-driven module is connected to the LED and regulates its injected current, enabling the LED to emit a light signal with an adjustable wavelength, thereby achieving dynamic wavelength tuning of the emission spectrum.

[0018] The processing module reconstructs the sample reflectance spectrum based on the photocurrent sequence output by the PD and the LED emission spectrum data.

[0019] Furthermore, the monolithic integrated optoelectronic device is based on gallium nitride material, and its hierarchical structure includes:

[0020] The gallium nitride layer, n-type gallium nitride layer, InGaN / GaN multi-quantum well layer, and p-type gallium nitride layer are stacked sequentially from bottom to top.

[0021] Indium tin oxide (ITO) current diffusion layer covering the surface of a p-type gallium nitride layer;

[0022] The PD region surrounding the LED has a multi-quantum well layer structure consistent with the LED.

[0023] A distributed Bragg reflector (DBR) layer is applied to the device surface to enhance light extraction efficiency.

[0024] Furthermore, the wavelength tuning mechanism of the LED is specifically as follows:

[0025] By increasing the injection current to drive the recombination of carriers in multiple quantum wells, the emission spectrum is continuously shifted from 575nm green light dominance to 400nm blue light dominance.

[0026] Furthermore, the processing module includes a Physical Information Convolutional Neural Network (PI-CNN) processing module, which, based on the photocurrent sequence output by the PD and the LED emission spectrum data, retrieves the sample reflectance spectrum through a reconstruction algorithm that integrates physical model constraints and data-driven methods. The PI-CNN processing module includes:

[0027] The feature extraction encoder, consisting of two layers of one-dimensional convolutional neural networks, is used to extract deep features from photocurrent sequences.

[0028] The decoder, consisting of fully connected layers, maps feature vectors to reflectance spectra;

[0029] The physical consistency constraint unit integrates the predicted spectrum with the system response function to generate the predicted photocurrent, and compares it with the measured photocurrent to calculate the loss.

[0030] Furthermore, the training of the PI-CNN processing module employs a two-stage strategy:

[0031] In the first stage, the network parameters are jointly optimized based on the spectral mean square error loss, smoothing regularization loss, and total variational loss.

[0032] In the second stage, a physical consistency loss is added on top of the loss in the first stage to constrain the physical matching relationship between the predicted spectrum and the measured photocurrent.

[0033] Furthermore, the PI-CNN processing module also includes a multi-task classification branch:

[0034] After completing the spectral reconstruction training, the encoder-decoder parameters are frozen, and the reconstructed spectrum is trained for a classification task using convolutional layers and fully connected layers.

[0035] Furthermore, the physical consistency loss function is defined as:

[0036] The integral of the product of the predicted spectrum and the system response kernel function is then added to the weighted mean square error of the measured photocurrent.

[0037] The system response kernel function includes the LED emission spectrum and the PD response spectrum.

[0038] A method for acquiring and reconstructing reflectance spectra using the aforementioned spectrometer, comprising:

[0039] S1. The injection current of the LED is controlled by the current driving module to emit a light signal with continuously adjustable wavelength to the sample under test;

[0040] S2. Utilize the PD integrated on the same chip to receive the sample reflected light signal and output a photocurrent sequence corresponding to the reflection intensity;

[0041] S3. Based on the photocurrent sequence and LED emission spectrum data, the sample reflectance spectrum is reconstructed by using a data-driven inversion algorithm that integrates the physical constraints of the system's optical response with the data of known reflectance spectrum training samples.

[0042] Further, in step S3, a PI-CNN model is established, and the sample reflectance spectrum is reconstructed through an inversion algorithm that fuses physical constraints and data-driven approaches. The PI-CNN inversion algorithm includes:

[0043] (a) Constructing a training set: Collect LED emission spectra and PD photocurrent responses corresponding to multiple sets of known standard reflection spectra;

[0044] (b) First stage training: Optimize network parameters by weighting and summing the spectral reconstruction loss, smoothing regularization loss, and total variational loss;

[0045] (c) Second stage training: Physical consistency loss is added on the basis of the loss in stage (b), and the photocurrent matching degree is calibrated by the integral result of the predicted spectrum and the system response function;

[0046] (d) Learning rate scheduling and gradient pruning techniques are used to improve training stability, and the reconstruction performance is evaluated by mean absolute error and spectral angle mapping.

[0047] Furthermore, the calculation of the physical consistency loss specifically includes:

[0048] The predicted photocurrent is generated by multiplying the predicted reflectance spectrum with the system response kernel function by wavelength.

[0049] After normalizing the measured photocurrent, the weighted mean square error between the measured photocurrent and the predicted photocurrent is calculated.

[0050] The system response kernel function is composed of the product of the LED emission spectrum, the PD response spectrum, and the wavelength step.

[0051] The present invention has the following beneficial effects:

[0052] This invention proposes a monolithically integrated computational reflectance spectrometer and a method for spectral acquisition and reconstruction. The core idea lies in constructing an active spectrometer, significantly different from traditional passive spectrometers. Traditional spectrometers typically emit broadband light, relying on the receiver to adjust the spectral response for spectral separation and reception. Their structure requires an optical bandpass filter before the receiver for beam splitting, and often only integrates the filter and receiver (e.g., a photodiode, PD), requiring an active broadband light source to function. This invention, however, focuses on the emission end, achieving spectral encoding by adjusting the wavelength of the light source, while the PD at the receiver remains fixed. This design eliminates the filtering stage at the receiver, resulting in stronger light signals at various wavelengths and a significantly improved signal-to-noise ratio. It also overcomes the limitation of existing integrated chips that often only integrate the receiver, integrating the light source (dual-wavelength LED) and receiver (PD) on the same substrate, with both using identical device structures for both emission and reception.

[0053] This innovative monolithic integrated design offers multiple advantages: it achieves collaborative integration of the light source and detection module at the chip level, effectively avoiding common problems of discrete devices such as alignment difficulties, high coupling losses, and system complexity. It significantly reduces the size of the spectral measurement module, improves integration, and lowers power consumption. Based on this, the entire system eliminates the need for mechanical scanning or filter arrays, resulting in a more compact and efficient structure. It overcomes the limitations of traditional computational spectrometers, such as large size, long optical paths, and poor integration, making it particularly suitable for scenarios requiring miniaturization and high stability, such as material identification, biomedical detection, and wearable analysis. It possesses greater potential for practical application and commercialization.

[0054] In the preferred embodiment, this invention also employs a Physical Information Convolutional Neural Network (PI-CNN) algorithm for spectral inversion calculations, further enhancing system performance. This algorithm embeds a spectral integral model into a deep learning framework, effectively integrating physical constraints and data-driven approaches. This not only improves the model's accuracy and generalization ability but also addresses the dependence of traditional data-driven algorithms on the number and balance of training samples—compared to traditional computational spectrometers requiring a large number of calibration samples, this invention achieves high-quality spectral reconstruction with only a small amount of data. This combination of an integrated optoelectronic structure and the PI-CNN algorithm constitutes a simple, computationally efficient, and highly accurate intelligent miniature spectral measurement system.

[0055] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the hierarchical structure of the spectrometer based on GaN monolithically integrated LED and PD, according to an embodiment of the present invention.

[0057] Figure 2This is a schematic diagram of the InGaN / GaN multi-quantum-well structure parameters according to an embodiment of the present invention.

[0058] Figure 3 This is a curve showing the tuning characteristics of the LED emission wavelength as a function of the injected current in an embodiment of the present invention.

[0059] Figure 4 This is a schematic diagram illustrating the working principle of a GaN-based monolithic integrated optoelectronic device according to an embodiment of the present invention.

[0060] Figure 5 This is a schematic diagram illustrating the matching between the LED emission spectrum and the PD response spectrum in an embodiment of the present invention.

[0061] Figure 6 This is a flowchart illustrating the principle of the PI-CNN-based spectral reconstruction method according to an embodiment of the present invention.

[0062] Figure 7 This is a graph showing the relationship between LED driving current and PD photocurrent in an embodiment of the present invention.

[0063] Figure 8 This is a schematic diagram of the physical information convolutional neural network (PI-CNN) according to an embodiment of the present invention.

[0064] Figure 9 This is a system block diagram of the monolithic integrated computational reflectance spectrometer of the present invention.

[0065] Figure 10 This is a flowchart illustrating the overall process of the reflectance spectrum acquisition and reconstruction method of the present invention. Detailed Implementation

[0066] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0067] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as "connected to" another component, it can be directly connected to or indirectly connected to that other component. Furthermore, a connection can be used for fixing, coupling, or communication.

[0068] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of 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, they should not be construed as limitations on the present invention.

[0069] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0070] This invention aims to address the problems of large size, poor integration, and reliance on large amounts of calibration data in traditional spectrometers, proposing a monolithic integrated computational reflectance spectrometer based on active coding. Its core is the integration of a dual-wavelength LED and a photodiode (PD) on the same substrate. The wavelength of the emitting end is adjustable by controlling the LED (active spectral coding), while the receiving end PD remains fixed, thus eliminating the need for receiver filtering, improving the signal-to-noise ratio, and achieving system miniaturization and high integration. This design avoids the problems of difficult alignment and high coupling loss associated with discrete components, making it suitable for wearable devices, biological detection, and other scenarios. In a preferred embodiment, this invention also introduces a Physical Information Convolutional Neural Network (PI-CNN) algorithm, fusing physical constraints and data-driven approaches. This allows for high-precision spectral reconstruction with only a small amount of data, solving the dependence on training samples inherent in traditional algorithms.

[0071] See Figure 9 This invention provides a monolithically integrated computational reflectance spectrometer, comprising a monolithically integrated optoelectronic device, a current-driven module, and a processing module. The monolithically integrated optoelectronic device includes a dual-wavelength light-emitting diode (LED) and a photodiode (PD). The LED emits a wavelength-tunable light signal through current modulation, while the PD receives the sample's reflected signal and outputs a photocurrent. The LED and PD are integrated on the same substrate, forming a monolithic structure with coordinated light source emission and reflection reception. Active spectral encoding is achieved through dynamic wavelength tuning of the LED. The current-driven module is connected to the LED and modulates its injected current, enabling the LED to emit a wavelength-tunable light signal, thus achieving dynamic wavelength tuning of the emission spectrum. The processing module reconstructs the sample reflectance spectrum based on the photocurrent sequence output by the PD and the LED emission spectral data.

[0072] like Figure 1As shown, in some embodiments, the monolithic integrated optoelectronic device is based on gallium nitride (GaN) material, and its layered structure includes: a GaN layer, an n-type GaN layer, an InGaN / GaN multi-quantum-well layer, and a p-type GaN layer stacked sequentially from bottom to top; an indium tin oxide (ITO) current diffusion layer covering the surface of the p-type GaN layer; a PD region surrounding the LED, whose multi-quantum-well layer structure is consistent with that of the LED; and a distributed Bragg reflector (DBR) layer covering the device surface to enhance light extraction efficiency.

[0073] In some embodiments, the wavelength tuning mechanism of the LED is specifically as follows: by increasing the injection current to drive the recombination of multiple quantum well carriers, the emission spectrum is continuously shifted from 575nm green light dominance to 400nm blue light dominance.

[0074] In some embodiments, the processing module includes a Physical Information Convolutional Neural Network (PI-CNN) processing module, which, based on the photocurrent sequence output by the PD and the LED emission spectrum data, retrieves the sample reflectance spectrum by fusing physical model constraints and a data-driven reconstruction algorithm. The PI-CNN processing module includes: a feature extraction encoder, consisting of two one-dimensional convolutional neural networks, used to extract deep features of the photocurrent sequence; a decoder, consisting of fully connected layers, which maps the feature vectors to reflectance spectra; and a physical consistency constraint unit, which integrates the predicted spectrum with the system response function to generate a predicted photocurrent and compares it with the measured photocurrent to calculate the loss.

[0075] In some embodiments, the training of the PI-CNN processing module adopts a two-stage strategy: in the first stage, the network parameters are jointly optimized based on the spectral mean square error loss, the smoothing regularization loss and the total variational loss; in the second stage, a physical consistency loss is added on the basis of the loss in the first stage to constrain the physical matching relationship between the predicted spectrum and the measured photocurrent.

[0076] In some embodiments, the PI-CNN processing module further includes a multi-task classification branch: after completing the spectral reconstruction training, the encoder-decoder parameters are frozen, and the reconstructed spectrum is trained for a classification task through convolutional layers and fully connected layers.

[0077] In some embodiments, the physical consistency loss function is defined as: the integral of the product of the predicted spectrum and the system response kernel function, plus the weighted mean square error of the measured photocurrent; wherein the system response kernel function includes the LED emission spectrum and the PD response spectrum.

[0078] See Figure 10 The present invention also provides a method for acquiring and reconstructing the reflectance spectrum using the spectrometer, comprising:

[0079] S1. The injection current of the LED is controlled by the current driving module to emit a light signal with continuously adjustable wavelength to the sample under test;

[0080] S2. Utilize the PD integrated on the same chip to receive the sample reflected light signal and output a photocurrent sequence corresponding to the reflection intensity;

[0081] S3. Based on the photocurrent sequence and LED emission spectrum data, the sample reflectance spectrum is reconstructed by using a data-driven inversion algorithm that integrates the physical constraints of the system's optical response with the data of known reflectance spectrum training samples.

[0082] The physical constraints of the system's optical response can specifically include the integral relationship between the reflection spectrum, the LED emission spectrum, and the PD response spectrum in the wavelength dimension; the system response kernel function composed of the LED emission spectrum, the PD response spectrum, and the wavelength step; and the matching relationship between the predicted photocurrent calculated from the predicted spectrum using the kernel function and the measured photocurrent.

[0083] In some embodiments, in step S3, a PI-CNN model is established, and the sample reflectance spectrum is reconstructed through an inversion algorithm that fuses physical constraints and data-driven approaches. The PI-CNN inversion algorithm includes:

[0084] (a) Constructing a training set: Collect LED emission spectra and PD photocurrent responses corresponding to multiple sets of known standard reflection spectra;

[0085] (b) First stage training: Optimize network parameters by weighting and summing the spectral reconstruction loss, smoothing regularization loss, and total variational loss;

[0086] (c) Second stage training: Physical consistency loss is added on the basis of the loss in stage (b), and the photocurrent matching degree is calibrated by the integral result of the predicted spectrum and the system response function;

[0087] (d) Learning rate scheduling and gradient pruning techniques are used to improve training stability, and the reconstruction performance is evaluated by mean absolute error and spectral angle mapping.

[0088] In some embodiments, the calculation of the physical consistency loss specifically includes:

[0089] The predicted photocurrent is generated by multiplying the predicted reflectance spectrum with the system response kernel function by wavelength.

[0090] After normalizing the measured photocurrent, the weighted mean square error between the measured photocurrent and the predicted photocurrent is calculated.

[0091] The system response kernel function is composed of the product of the LED emission spectrum, the PD response spectrum, and the wavelength step.

[0092] The main technical advantages of this invention are as follows: Through innovative active spectral coding design, it abandons the traditional passive spectrometer's reliance on receiver filters or tunable filters for spectral splitting. Instead, it utilizes the dynamic wavelength tuning of the LED at the light-emitting end to achieve spectral coding, enabling the receiver PD to fix the received reflected signal without any adjustment or filtering structure, significantly reducing optical path loss and improving the signal-to-noise ratio. At the same time, it breaks through the limitation of existing integrated chips that only contain a receiver module, co-integrating the light source and receiver on the same substrate and using the same device hierarchy structure, realizing the monolithicization of light emission and reception at the chip level, completely solving the problems of alignment deviation, coupling loss, and system complexity of discrete devices, forming a compact spectral measurement module that does not require mechanical scanning or filter arrays, and has the advantages of miniaturization, low power consumption, and high stability, providing an efficient solution for wearable terminals and other scenarios. In the preferred scheme, the physical constraints of the system's optical response are further integrated with data-driven inversion algorithms (such as PI-CNN). By embedding a spectral integral model, the physical interpretability of the reconstruction process is enhanced, significantly reducing the dependence on a large amount of calibration data and improving the cross-scene generalization ability. Together, they form an intelligent spectral measurement system with simple structure, high computational efficiency, and high reconstruction accuracy.

[0093] The features, working principle and advantages of specific embodiments of the present invention are further described below.

[0094] An integrated micro-computational spectrometer system, specifically a monolithic integrated optoelectronic device based on gallium nitride (GaN) material, achieves rapid acquisition and high-precision reconstruction of sample reflectance spectra by integrating dual-wavelength light-emitting diodes (LEDs) and photodiodes (PDs) onto a single chip. The system uses current regulation to drive the LEDs to emit wavelength-tunable light sources. After reflection from the sample, the PD receives the reflected signal and converts it into a current response. This response is then combined with a Physical Information Neural Network (PI-CNN) algorithm to perform spectral inversion calculations, thus achieving compact structure and efficient spectral measurement without the need for mechanical scanning or filter arrays. This invention overcomes the limitations of traditional computational spectrometers, such as large size, long paths, and poor integration, and is suitable for scenarios such as material identification, biomedical detection, and wearable analysis, possessing promising prospects for miniaturization and industrialization.

[0095] The device consists of a GaN-based LED and a PD with identical structures. The LED is located in the central region, surrounded by a PD region that operates without bias. The hierarchical structure of the LED and PD is as follows: Figure 1As shown, the exhibits include unintentionally doped gallium nitride layers, silicon-doped n-type gallium nitride, InGaN / GaN multiple quantum wells (MQWs) with varying In content, and magnesium-doped p-type gallium nitride, all of which are fabricated using metal-organic chemical vapor deposition (MOCVD). Among them, InGaN / GaN multiple quantum wells (MQWs) serve as the core functional layer of the device. Their manufacturing parameters include the number of cycle layers, the thickness of each sublayer, and the In composition. Specific examples of parameters are as follows: Two quantum well cycle structures are used in combination. One is a 12-cycle (loop=12) B-QW / B-QB structure, where the B-QB layer (GaN composition) is 10.5 nm thick with a Si doping concentration of 3.68E+17 in each cycle, and the B-QW layer (InGaN composition) is 3.5 nm thick, for a total thickness of 126 nm over 12 cycles. The other is a 6-cycle (loop=6) G-QW / G-QB structure, where the G-QB layer (GaN composition) is 12.5 nm thick with a Si doping concentration of 3.68E+17 in each cycle, and the G-QW layer (InGaN composition) is 2.5 nm thick, for a total thickness of 81 nm over 6 cycles. For details of these parameters, please refer to [link to relevant documentation]. Figure 2 These epitaxial layers are grown on a 4-inch c-plane sapphire substrate. To facilitate current diffusion, a 120 nm thick indium tin oxide (ITO) layer is deposited on the surface of the gallium nitride layer. The diode grid is formed using photolithography and inductively coupled plasma etching. It should be noted that one of the core advantages of this device is that it eliminates the need for complex process design: although LEDs and PDs have functional differences, no additional process compensation measures are required to address their functional requirements. High-performance monolithic integration of LEDs and PDs on the same substrate can be achieved using conventional LED manufacturing processes (such as metal-organic chemical vapor deposition, photolithography, and etching), simplifying the fabrication process and reducing costs.

[0096] Subsequently, a silicon dioxide passivation layer was deposited using plasma-enhanced chemical vapor deposition, followed by the deposition of a silicon dioxide / titanium dioxide distributed Bragg reflector (DBR) layer using an optical thin-film deposition machine. The device employs a flip-chip design, with both light emission and reception occurring on the sapphire substrate. The silicon dioxide / titanium dioxide DBR layer is deposited on the non-sapphire substrate side of the device. This configuration effectively enhances the light extraction efficiency of the LED without affecting the PD's reception of reflected light from the sample on the sapphire substrate, ensuring the integrity of the PD's received signal.

[0097] The final step involves fabricating p-type and n-type electrodes using processes such as photolithography, electron beam evaporation, and boom-bust technology. During current injection, carrier recombination within the MQW induces the LED to emit light in the wavelength range of 400nm to 575nm. Furthermore, as the driving current increases, the emission wavelength shifts from being dominated by green light to being dominated by blue light. Figure 3As shown. The device of this invention employs a dual-wavelength quantum well structure. The core function of increasing the driving current is to shift the dominant peak of the emission spectrum from 575nm green light to 400nm blue light. The mechanism of increasing the current to drive carrier recombination in multiple quantum wells to achieve wavelength tuning has been verified by Apsys simulation: In this structure, during the increase of the driving current, the wavelength shift of the electroluminescence spectrum is clearly correlated with the carrier concentration distribution within the multiple quantum wells (MQWs). As the injection current increases, the carrier concentration in the multiple quantum well region increases significantly, and the recombination probability of carriers changes in quantum wells with different In composition. The carrier recombination ratio in low In composition quantum wells (corresponding to short-wavelength blue light emission) gradually increases to higher than that in high In composition quantum wells (corresponding to long-wavelength green light emission), ultimately manifesting as a continuous shift of the electroluminescence spectrum from 575nm green light dominance to 400nm blue light dominance. During light injection, the MQWs of the PD generate carriers, thereby generating photocurrent. The photodiode employs a multiple quantum well (MQW) structure, which endows it with low noise characteristics. This allows the photodiode to effectively suppress noise interference even when the absorption rate of reflected light is low, thereby providing a high signal-to-noise ratio and ensuring the stability and accuracy of the photocurrent signal. Figure 4 The working principle of this GaN optoelectronic device is demonstrated. When a sample is placed above the device, the light signal from the LED reaches the sample surface, is reflected, and is then received by the PD on the same side, allowing the acquisition of specific current information. This device incorporates a silicon dioxide dielectric layer between the LED and the PD. This dielectric layer provides some reflection, reducing parasitic photocurrent. Simultaneously, during the measurement of the spectral reflectance signal, the remaining parasitic photocurrent remains constant, preventing interference with the effective signal that varies with the sample's reflectance characteristics, thus ensuring the accuracy of the PD's received signal. Figure 5 The emission spectra of the LED and the response spectra of the PD under different injection currents are shown. The large overlap between the two over a wide wavelength range demonstrates the good responsivity of the device, which is the basis for encoding the spectral information of the analyte as photocurrent.

[0098] This device can operate with a drive current of up to 25mA, and this current is applied in a short-duration pulse manner, ensuring long-term stable operation. Furthermore, this device uses an FPC package, within which a steel reinforcement structure can be integrated for mechanical strengthening and enhanced thermal diffusion.

[0099] like Figure 6As shown, the spectral reconstruction method is based on a Physical Information Convolutional Neural Network (PI-CNN). It acquires multiple sets of LED emission data with different known spectra and corresponding photodetector response signals to construct training and validation sets. The data includes the emission spectral distribution of the LED under different driving states and the normalized response values ​​of the photodetector at each wavelength sampling point. An integral model composed of the emission spectrum, response spectrum, and a series of photocurrents is used as a priori equation, combined with the standard reflectance spectrum of known samples as a data-driven supervised term for optimization. It should be noted that the trend of photocurrent variation with LED driving current (e.g., ...) is observed in the experiment. Figure 7 As shown in the figure, the PD is far from reaching the light saturation state; at the same time, even if there are nonlinear factors caused by the device itself, they can be automatically corrected during the learning process of the neural network, thereby ensuring the accuracy of spectral analysis based on the PD response.

[0100] like Figure 8 As shown, the PI-CNN model includes a feature extraction encoder consisting of two one-dimensional convolutional neural networks and a decoder consisting of multiple fully connected networks. The training process employs a two-stage strategy:

[0101] In the first stage, the spectral mean square error loss, spectral smoothness loss, and total variational loss are calculated based on the training set, and then weighted and summed. The AdamW optimization algorithm is used to adaptively update the network parameters until the preset number of iterations is reached. The total loss function can be expressed as:

[0102]

[0103] in, , and The weighting coefficients are represented by N, which represents the number of wavelength sampling points in the spectrum. and These represent the predicted reflectance value and the true reflectance value, respectively. These three loss components play complementary roles in guiding spectral reconstruction: reconstruction loss... To ensure an accurate fit of the spectral shape, Let be the mean square error function. To predict reflectance values, The true reflectance value; smoothed regularization loss. Sudden oscillations were suppressed; i represents the wavelength sampling point index; total variation loss. It promotes edge sparsity, effectively reduces noise, and enhances the structural stability of the spectrum.

[0104] In the second stage, a physical consistency loss is added to the loss from the first stage. This loss is calculated by convolving the predicted spectrum with the system response to obtain the predicted photocurrent, and then calculating a weighted mean square error with the measured photocurrent to ensure consistency between the spectral reconstruction result and the physical measurement. The AdamW algorithm is then used to optimize the parameters up to the second preset number of iterations. The loss function is updated as follows:

[0105]

[0106]

[0107]

[0108]

[0109] in, , , and Represents the weighting coefficient. For physical consistency loss, and Let be the measured photocurrent and the predicted photocurrent corresponding to the i-th wavelength sampling point, respectively, and K be the global calibration parameter. This is the maximum value in the measured photocurrent sequence. It is a tiny constant. Let E be the system response kernel function, and E be the emission spectrum of the LED under different bias states. This is the corresponding spectrum of PD. This represents the wavelength step size.

[0110] During the two training phases of spectral reconstruction, the encoder and decoder parameters remain trainable and are not frozen.

[0111] During model training, learning rate scheduling and gradient pruning techniques are employed to improve training stability. Spectral reconstruction performance is evaluated using mean absolute error (MAE) and spectral angle mapping (SAM) metrics, comprehensively reflecting the matching degree between spectral amplitude and shape. This method effectively suppresses noise influence, ensures the physical consistency of spectral reconstruction, and is suitable for multi-channel miniature spectral detection systems and real-time spectral analysis.

[0112] Furthermore, this embodiment of the invention also includes a multi-task classification training phase. After completing the two-stage training for spectral reconstruction, the parameters of the encoder and decoder are frozen, and only the classification branch is trained to achieve the multi-task classification objective of the spectral data. The classification branch consists of convolutional layers and fully connected layers, used to improve the model's generalization ability and task adaptability. Classification performance is evaluated using accuracy metrics, effectively enhancing the model's discriminative ability in practical applications.

[0113] It should be understood that the algorithm for reconstructing the sample reflectance spectrum of the present invention is not limited to using PINN. For example, traditional neural networks (such as MLP, 1D-CNN) can also be used instead of PINN for spectral regression. By learning the photocurrent sequence and LED emission spectrum data, the sample reflectance spectrum reconstruction function of the present invention can also be realized.

[0114] In summary, the embodiments of the present invention provide a monolithic integrated computational reflectance spectrometer and a method for spectral acquisition and reconstruction. Compared with the prior art, the significant advantages of the embodiments of the present invention are reflected in the following aspects:

[0115] This invention presents a miniature reflectance computational spectrometer based on a GaN monolithically integrated LED and PD, combined with a Physical Information Neural Network (PI-CNN) algorithm, to achieve high-precision reconstruction of the reflectance spectra of substances. Compared to existing computational spectrometers that are complex in structure, bulky in size, and rely on multi-device arrays or external optical systems, the single-device solution proposed in this invention is easier to miniaturize and integrate, making it particularly suitable for wearable devices, biosensor applications, and other scenarios, and possessing greater potential for practical application and commercialization. Furthermore, traditional computational spectrometers rely on a large number of calibration samples for reconstruction, while the PINN algorithm introduced in this invention requires only a small amount of data to achieve high-quality spectral reconstruction, effectively alleviating the problem of training data dependence.

[0116] The technical advantages of this invention fully utilize the excellent optoelectronic properties and high monolithic integration capabilities of gallium nitride (GaN) materials. This material not only possesses the dual capabilities of high-efficiency light emission and high-response detection, but also exhibits excellent thermal stability and a scalable spectral response range. By achieving monolithic integration of the emitter (LED) and detector (PD) on the same substrate using GaN material, problems common in discrete devices, such as alignment difficulties, high coupling losses, and system complexity, are avoided. This enables the collaborative integration of the light source and detection module at the chip level, resulting in a smaller, more integrated, and lower-power spectral measurement module. Furthermore, the introduction of a physical information neural network embeds the spectral integration model into a deep learning framework, effectively merging physical constraints and data-driven approaches. This improves the model's accuracy and generalization ability, solving the dependence of traditional data-driven algorithms on the number and balance of training samples. The combination of these two aspects constitutes a simple, computationally efficient, and highly accurate intelligent micro-spectral measurement system.

[0117] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A monolithic integrated computational reflectance spectrometer, characterized in that, include: Monolithic integrated optoelectronic devices based on gallium nitride materials include dual-wavelength light-emitting diodes (LEDs) and photodiodes (PDs) based on InGaN / GaN multi-quantum-well layers, both of which have the same multi-quantum-well layer structure; The LED emits a wavelength-tunable light signal through current modulation, while the PD receives the reflected signal from the sample and outputs a photocurrent. The LED and PD are integrated on the same substrate, forming a monolithic structure that coordinates light source emission and reflection reception. Active spectral encoding is achieved through dynamic wavelength tuning of the LED. Furthermore, by increasing the injection current to drive multi-quantum-well carrier recombination, the emission spectrum continuously shifts from 575nm green light dominance to 400nm blue light dominance. A current-driven module is connected to the LED and regulates its injected current, enabling the LED to emit a light signal with an adjustable wavelength, thereby achieving dynamic wavelength tuning of the emission spectrum. The processing module reconstructs the sample reflectance spectrum based on the photocurrent sequence output by the PD and the LED emission spectrum data.

2. The spectrometer according to claim 1, characterized in that, The hierarchical structure of the monolithic integrated optoelectronic device includes: The gallium nitride layer, n-type gallium nitride layer, InGaN / GaN multi-quantum well layer, and p-type gallium nitride layer are stacked sequentially from bottom to top. Indium tin oxide (ITO) current diffusion layer covering the surface of a p-type gallium nitride layer; The PD area surrounding the LED; A distributed Bragg reflector (DBR) layer is applied to the device surface to enhance light extraction efficiency.

3. The spectrometer according to claim 1, characterized in that, The processing module includes a Physical Information Convolutional Neural Network (PI-CNN) processing module. Based on the photocurrent sequence and LED emission spectrum data output by the PD, it retrieves the sample reflectance spectrum by fusing physical model constraints and a data-driven reconstruction algorithm. The PI-CNN processing module includes: The feature extraction encoder, consisting of two layers of one-dimensional convolutional neural networks, is used to extract deep features from photocurrent sequences. The decoder, consisting of fully connected layers, maps feature vectors to reflectance spectra; The physical consistency constraint unit integrates the predicted spectrum with the system response function to generate the predicted photocurrent, and compares it with the measured photocurrent to calculate the loss.

4. The spectrometer according to claim 3, characterized in that, The training of the PI-CNN processing module adopts a two-stage strategy: In the first stage, the network parameters are jointly optimized based on the spectral mean square error loss, smoothing regularization loss, and total variational loss. In the second stage, a physical consistency loss is added on top of the loss in the first stage to constrain the physical matching relationship between the predicted spectrum and the measured photocurrent.

5. The spectrometer according to claim 4, characterized in that, The PI-CNN processing module also includes a multi-task classification branch: After completing the spectral reconstruction training, the encoder-decoder parameters are frozen, and the reconstructed spectrum is trained for a classification task using convolutional layers and fully connected layers.

6. The spectrometer according to claim 4, characterized in that, The loss function for the physical consistency loss is defined as: The integral of the product of the predicted spectrum and the system response kernel function is then added to the weighted mean square error of the measured photocurrent. The system response kernel function includes the LED emission spectrum and the PD response spectrum.

7. A method for acquiring and reconstructing reflectance spectra using the spectrometer described in any one of claims 1-6, characterized in that, include: S1. The injection current of the LED is controlled by the current driving module to emit a light signal with continuously adjustable wavelength to the sample under test; S2. Utilize the PD integrated on the same chip to receive the sample reflected light signal and output a photocurrent sequence corresponding to the reflection intensity; S3. Based on the photocurrent sequence and LED emission spectrum data, the sample reflectance spectrum is reconstructed by using a data-driven inversion algorithm that integrates the physical constraints of the system's optical response with the data of known reflectance spectrum training samples.

8. The method according to claim 7, characterized in that, In step S3, a PI-CNN model is established, and the sample reflectance spectrum is reconstructed using an inversion algorithm that combines physical constraints and data-driven approaches. The PI-CNN inversion algorithm includes: (a) Constructing a training set: Collect LED emission spectra and PD photocurrent responses corresponding to multiple sets of known standard reflection spectra; (b) First stage training: The network parameters are optimized by weighted sum of spectral reconstruction loss, smoothing regularization loss and total variational loss; (c) Second stage training: Physical consistency loss is added on the basis of the loss in stage (b), and the photocurrent matching degree is calibrated by the integral result of the predicted spectrum and the system response function; (d) Learning rate scheduling and gradient pruning techniques are used to improve training stability, and the reconstruction performance is evaluated by mean absolute error and spectral angle mapping.

9. The method according to claim 8, characterized in that, The calculation of the physical consistency loss specifically includes: The predicted photocurrent is generated by multiplying the predicted reflectance spectrum with the system response kernel function by wavelength. After normalizing the measured photocurrent, the weighted mean square error between the normalized measured photocurrent and the predicted photocurrent is calculated. The system response kernel function is composed of the product of the LED emission spectrum, the PD response spectrum, and the wavelength step.

Images