Calculation spectrometer collaborative design method and design architecture based on physical information deep learning

Abstract

The invention discloses a computing spectrometer collaborative design method and design architecture based on physical information deep learning. The architecture is composed of a spectrum sampling compression coding network, a physical submerged space and a spectrum decoding reconstruction network, and an end-to-end all-micro training network is formed. Wherein the coding network generates a filtering sampling spectrum corresponding to the response of the physical optical filter by using a pre-trained structure-spectrum predictor and trainable parameters (recorded as H), and maps an input spectrum to a submerged space with physical significance; and the decoding network reconstructs the original spectrum to be measured based on the submerged space. Physical information loss (such as diversity constraints of filter responses and limitations of overlapping among structural parameters) and spectrum reconstruction errors (such as mean square errors or other distance measures) are jointly adopted for training, collaborative design of hardware (optical filter parameters) and software (reconstruction algorithms) is achieved, channel redundancy is reduced, sampling efficiency is improved, and the system performance is improved. And the system is ensured to have good adaptability to different structure parameters or spectral bands.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of spectrometer and optical system design, and relates to a computational spectrometer collaborative design method and design architecture based on physical information deep learning. BACKGROUND

[0002] Traditional spectrometer design mainly relies on preset optical filter arrays to realize light spectrum reconstruction by capturing light intensity data of specific wavebands. However, due to the limitations of materials, processes and physical principles, these filter designs have problems such as low degree of freedom, channel redundancy, non-independent response and low reconstruction accuracy. In addition, the existing system usually adopts a series optimization mode of independent hardware design and then software remediation, which cannot realize the gradient bidirectional transmission between structure and algorithm by using the fully differentiable characteristics of neural network, thereby limiting the overall performance improvement. Therefore, there is an urgent need for a solution that can realize the collaborative design of hardware and software at the system level, optimize filter diversity and improve sampling efficiency. SUMMARY

[0003] The application aims to provide a computational spectrometer design method and design architecture based on physical information deep learning, to construct an end-to-end physical information deep learning framework, realize the joint optimization of optical filter structure parameters and spectrum reconstruction algorithm, and achieve the technical effects of soft and hardware co-evolution, reduction of channel redundancy and improvement of system universality.

[0004] I. Design architecture The computational spectrometer design architecture disclosed by the application comprises: Spectrum sampling compression encoding network: sequentially connected by trainable parameters H, pre-trained structure-spectrum predictor and spectrum sampling module, and outputting a physical latent space vector D; Physical latent space: receiving the output of the spectrum sampling compression encoding network, and the vector representing the actual detector encoding signal; Spectrum decoding reconstruction network: inputting the physical latent space vector D and outputting a reconstructed spectrum.

[0005] II. Collaborative design method The collaborative design method disclosed by the application comprises the following steps: S1: pre-training the structure-spectrum predictor to establish a differentiable mapping from the optical filter structure parameters to the spectrum response through electromagnetic simulation, optical experiment or first principle calculation data; S2: constructing a spectrum sampling compression encoding network, comprising: Trainable parameters H: representing the physical structure parameters of the optical filter, including at least one of length, width, periodicity or material combination; Pre-trained structure-spectrum predictor: inputting H and outputting a filter sampling spectrum S; Spectral sampling module: performing point multiplication operation of input spectrum I and filtered sampling spectrum S, mathematical form is D = (S x I), wherein D is physical latent space vector, f f is detector wavelength response function; S3: cascaded spectral decoding reconstruction network, inputting D and outputting reconstructed spectrum; S4: defining joint loss function, Loss = Recon (I, I rec ) + γ • Corr (S) , wherein Recon is spectral reconstruction loss between input spectrum I and reconstructed spectrum I rec , Corr is correlation loss between filtered sampling spectrum S; γ is balance coefficient; S5: based on joint loss function, synchronously optimizing trainable parameter H and decoding network weight through back propagation, gradient transmission path is: Recon loss→spectral decoding reconstruction network→physical latent space→spectral sampling module→structure-spectrum predictor→H, Corr loss→spectral sampling compression encoding network→H; S6: outputting optimized H as hardware manufacturing parameter, and spectral decoding reconstruction network as reconstruction algorithm.

[0006] The structure-spectrum predictor is one of full connection network, convolutional neural network or generative adversarial network, pre-training data is derived from electromagnetic simulation, optical experiment or first principle calculation, and weight thereof is frozen in collaborative training stage.

[0007] The trainable parameter H specifically comprises: Metasurface filter: unit period a: 1.0-3.0 μm, structure length l: 1.0-2.5 μm, width w: 0.5-1.0 μm, and satisfying constraint: a>l>w; Integrated optical path filter: waveguide spacing p, microdisk radius r, coupling spacing g or waveguide width w, and optical coupling is realized through inverted taper waveguide with insertion loss <2 dB or grating coupler with insertion loss <6 dB.

[0008] The spectral decoding reconstruction network is one of generative vector machine, full connection layer network or convolutional neural network.

[0009] The Corr loss is Pearson correlation coefficient, Spearman correlation coefficient or average value based on Gram matrix, for reducing spectral correlation of different channels; and the Recon loss is L1 loss, mean square error (MSE) or Huber loss.

[0010] ​The physical latent space is an output vector of a spectrum sampling compression encoding network, the dimension of which is consistent with the filter channel number of an actual spectrometer, and each component of the vector is physically equivalent to a detector sampling signal of the actual spectrometer.

[0011] The application is based on a physical information deep learning-based computing spectrometer collaborative design method, which uses a deep neural network to realize joint optimization of optical filter structure parameters and spectrum reconstruction algorithms, breaks through the limitations of traditional design in filter response independence, sampling efficiency and device compactness, and has strong structure and spectrum universality. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The overall schematic diagram of the general computing spectrometer design architecture based on physical information deep learning proposed in the application shows an end-to-end differentiable collaborative design process, including a spectrum sampling compression encoding network, a physical latent space and a spectrum decoding reconstruction network three parts; Figure 2 The internal structure diagram of the spectrum sampling compression encoding network in the application shows in detail the process of generating a filtered spectrum S from a trainable parameter H through a structure-spectrum predictor, and then obtaining a physical latent space vector D through a spectrum sampling module; Figure 3 The training process diagram of the pre-trained structure-spectrum predictor in the application shows the differentiable mapping process from structure parameters to spectrum responses through electromagnetic simulation or experimental data; Figure 4 The hardware structure and optical path schematic diagram of the infrared metasurface reflective portable spectrometer in embodiment 1 of the application, including a metasurface unit card, a thermal imager and the corresponding light intensity distribution (physical latent space D); Figure 5 The infrared metasurface design process and result schematic diagram in embodiment 1 of the application, including metasurface unit structure parameters (51), 25-channel reflective spectrum response (52) and 2D and 3D structure schematic diagram (53, 54); Figure 6 The spectrum reconstruction effect verification diagram of the integrated optical path scheme in embodiment 2 of the application on different materials (grass, glass, concrete and quartz) in the visible light band; Figure 7 The hardware structure schematic diagram of the integrated optical path scheme in embodiment 2 of the application, including the structure parameters of the waveguide disturbance filter and the micro-disk filter and the light coupling mode thereof. DETAILED DESCRIPTION

[0013] The design architecture of the application will be described below with reference to the accompanying drawings.

[0014] The overall design architecture of the application is shown in Figures 1-3 , specifically including: Spectral sampling compression encoding network (11): sequentially connected by trainable parameters H, pre-trained structure-spectrum predictor, spectral sampling module; Physical latent space (12): receives the vector D output by the encoding network, which has a physical meaning equivalent to the actual detector signal; Spectral decoding reconstruction network (13): input D and output reconstructed spectrum.

[0015] Note: Figure 1 The end-to-end architecture is shown, Figure 2 The internal modules of the encoding network (H→predictor→S→sampling module→D) are shown, Figure 3 The training process of the structure-spectrum predictor.

[0016] Embodiment 1 (infrared metasurface) This embodiment takes the mid-infrared waveband (wavelength 8-14 μm) as an example to construct a reflective portable spectrometer: Hardware structure: as shown in Figure 4 , including optical hardware 42, which includes a metasurface unit card and a thermal imager; the metasurface unit card is composed of N independent metasurface filters; the number of thermal imagers is one or more, and when the number of thermal imagers is more than one, the number is the same as the number of independent metasurface filters; The structural parameters (trainable parameters H) of the metasurface unit card are defined as: H = {unit period a, structure length l, structure width w} (a total of N×3 parameters) When the number of thermal imagers is one, the thermal imager uses an infrared area array detector; when the number of thermal imagers is more than one, the thermal imager uses a mid-infrared detector; Structure-spectrum predictor: A convolutional neural network (CNN) is used to obtain training data (such as FDTD simulation of reflected spectra of different (a, l, w) combinations) through electromagnetic simulation. The AdamW optimizer is used for pre-training, and the loss function is MSE.

[0017] Physical latent space implementation: After the incident spectrum of the material spectrum 41 is filtered by the metasurface unit card, the intensity distribution corresponding to each independent metasurface filter on the thermal imager is formed (43 in Figure 4 , the intensity vector is the physical latent space D (as shown in 43 in Figure 4 ), which satisfies: D= f (S⋅I) Where S is the metasurface reflectance spectrum, I is the input spectrum, f is the thermal imager response function.

[0018] Spectral decoding and reconstruction network: A three-layer fully connected network is used, with input dimension = N (number of filter channels) and output dimension = spectral resolution.

[0019] Training process: 1) Generate a simulated dataset: a mixture of Gaussian and Lorentz distribution spectra; 2) Joint loss function: Loss=Recon(I,I rec )+γ•Corr(S) Where Recon is the input spectrum I and the reconstructed spectrum I rec The spectral reconstruction loss between the filtered sampled spectra S is represented by Corr, which is the correlation loss between the two spectra; γ is the coefficient that balances the spectral reconstruction loss and the correlation loss. 3) Backpropagation optimization: Freeze the CNN predictor weights and only update H and the decoder network weights; 4) Optimizer: AdamW (learning rate 10) -3 ).

[0020] After training for practical applications, the trainable parameters H of the spectral sampling compression coding network are used to prepare the real physical structure. Taking 25 filter channels as an example, the specific schematic diagram and design results are shown below. Figure 5 As shown. Figure 5 Figure 51 shows a schematic diagram of the metasurface unit and design parameters, Figure 52 shows the reflection spectral response of 25 filters, and Figures 53 and 54 show the corresponding 2D and 3D structural schematic diagrams.

[0021] Figure 5 This demonstrates that the design method of this invention has the capability to practically design metasurface arrays, enabling the design of multi-channel filter structures. Figure 53 is a scaled-down diagram of the actual results, with the proportions of its period, length, width, and other features consistent with the proportions of the real structure.

[0022] Example 2 (Integrated Optical Path) This embodiment demonstrates an integrated optical solution for the visible light band: Hardware structure: Employs silicon-based optical chips, including: Waveguide perturbation filter ( Figure 7 (71-73): Trainable parameters H = {waveguide spacing p, structure length l, width w}; Microdisk filter ( Figure 7 (74): Trainable parameters H = {microdisk radius r, microdisk-optical waveguide coupling distance g, waveguide width w}; Optical coupling scheme: Input: Single-mode fiber (center wavelength 632nm); Coupler: Inverted taper waveguide (insertion loss < 2dB) or grating coupler (insertion loss < 6dB).

[0023] Key difference design: Structure-spectrum predictor: Physics-inspired 4-layer fully connected network (input H → output transmission spectrum S); Physical latent space: Photodetector current signal (needs linear response calibration); Decoding network: Light-weight CNN (parameter amount < 50k); Bypass compensation module: Add 3-layer fully connected network before the spectral sampling module; Use experimental data to train, compensate for the theoretical deviation of point multiplication operation.

[0024] 3. Reconstruction effect verification Test samples: Spectra of grass, glass, concrete, and quartz in the 400-700nm wavelength range Figure 6 ); Results: The average correlation coefficient between the reconstructed spectrum and the reference spectrum is > 0.95.

Claims

1. A method for collaborative design of a computational spectrometer based on physical information deep learning, characterized in that, Comprise: S1: pre-training structure-spectrum predictor, to establish a differentiable mapping from optical filter structure parameters to spectral response through electromagnetic simulation, optical experiment or first-principles calculation data; S2: construct a spectral sampling compression encoding network, including: Trainable parameters H: represent the physical structure parameters of the optical filter, including at least one of length, width, periodicity or material combination; Pre-training structure-spectrum predictor: input H, output filter sampling spectrum S; Spectral sampling module: performs a point-wise multiplication operation of the input spectrum I with the filtered sampled spectrum S, mathematically represented as D = (S x I), where D is the physical latent space vector, f f is the detector wavelength response function;​ S3: cascade spectral decoding reconstruction network, input D and output reconstructed spectrum; S4: define a joint loss function, Loss = Recon(I, I rec ) + γ • Corr(S) wherein Recon is the spectral reconstruction loss between the input spectrum I and the reconstructed spectrum I rec Corr is the correlation loss between the filtered sampled spectrum S; and γ is a balancing coefficient. S5: based on the joint loss function, the trainable parameters H and the decoding network weight are optimized synchronously by back propagation, and the gradient transmission path is: Recon loss→spectral decoding reconstruction network→physical latent space→spectral sampling module→structure-spectrum predictor→H, Corr loss→spectral sampling compression encoding network→H; S6: output the optimized H as the hardware manufacturing parameter, and the spectral decoding reconstruction network as the reconstruction algorithm.

2. The physical information deep learning based computational spectrometer co-design method of claim 1, wherein, The structure-spectrum predictor is one of a full connection network, a convolutional neural network or a generative adversarial network, the pre-training data is derived from electromagnetic simulation, optical experiment or first-principles calculation, and the weight is frozen in the collaborative training stage.

3. The physical information deep learning based collaborative design method of a computational spectrometer of claim 1, wherein, The trainable parameters H specifically include: Metasurface filter: unit period a: 1.0-3.0 μm, structure length l: 1.0-2.5 μm, width w: 0.5-1.0 μm, and satisfy the constraint: a>l>w; Integrated optical path filter: waveguide spacing p, microdisk radius r, coupling spacing g or waveguide width w, optical coupling is realized through an inverted taper waveguide with an insertion loss of <2 dB or a grating coupler with an insertion loss of <6 dB.

4. The physical information deep learning based computational spectrometer co-design method of claim 1, wherein, The spectral decoding reconstruction network is one of a generative vector machine, a full connection layer network or a convolutional neural network.

5. The physical information deep learning based computational spectrometer co-design method of claim 1, wherein, The Corr loss is a Pearson correlation coefficient, a Spearman correlation coefficient or an average value based on a Gram matrix, used to reduce the spectral correlation of different channels; the Recon loss is an L1 loss, a mean square error MSE or a Huber loss.

6. A computing spectrograph design architecture, characterized by, Comprise: Spectral sampling compression encoding network: sequentially connected by trainable parameters H, pre-training structure-spectrum predictor and spectral sampling module, outputting a physical latent space vector D; Physical latent space: receiving the output of the spectral sampling compression encoding network, the vector representing the actual detector signal; Spectral decoding reconstruction network: input the physical latent space vector, output the reconstructed spectrum.

7. The computing spectrograph design architecture of claim 6, wherein, The physical latent space is the output vector of the spectral sampling compression encoding network, the dimension of which is consistent with the number of filter channels of the actual spectrometer, and each component of the vector is physically equivalent to the detector sampling signal of the actual spectrometer.

Images