Non-spherical dust aerosol light scattering measurement inversion method and system based on physical constraint neural network

Abstract

The application discloses a non-spherical dust aerosol light scattering measurement inversion method and system based on a physically constrained neural network, and belongs to the technical field of atmospheric particulate optical measurement and inversion. The method comprises the following steps: modeling non-spherical dust particles by an optical model and calculating discrete single-particle optical properties; calculating the bulk optical properties of a particle group by a forward simulation module, constructing a training set and forming a physical constraint; based on the training set and the physical constraint, constructing a multi-task physically constrained neural network, and combining inversion parameter errors and physical constraint errors, so that the physically constrained neural network fits the observation data while maintaining physical consistency, high-precision and rapid inversion of the microphysical parameters of dust aerosol from multi-angle light scattering matrix measurement data is realized, and it is ensured that the inversion result meets the physical law. The application also provides a non-spherical dust aerosol light scattering measurement inversion system for realizing the non-spherical dust aerosol light scattering measurement inversion method.

Description

Technical Field

[0001] This invention belongs to the field of optical measurement and inversion technology of atmospheric particulate matter, and particularly relates to a method and system for measuring and inverting light scattering of non-spherical dust aerosols based on a physically constrained neural network. Background Technology

[0002] my country suffers from numerous sandstorms every year, causing severe losses and damage to industrial and agricultural production, transportation, and human life due to this extreme weather event. Therefore, conducting basic research on large-scale sandstorm observation and inversion methods is of significant scientific importance and will play a positive role in major national events such as sandstorm early warning and disaster reduction.

[0003] Dust aerosols are a significant component of atmospheric particulate matter, exerting a substantial influence on weather, climate, air quality, and radiation transport. Furthermore, during transport, dust aerosols increase pollution along their path and can mix with other pollutants to produce fog, haze, and other pollutants, significantly increasing atmospheric optical thickness and severely impacting the atmospheric environment.

[0004] The optical properties of dust aerosols are determined by microphysical parameters, such as particle shape, size distribution, and refractive index. Accurately obtaining these microphysical parameters is crucial in atmospheric science. Multi-angle light scattering measurements provide optical information with multi-angle resolution and are sensitive to particle size, refractive index, and shape, making them an important method for obtaining aerosol microphysical parameters.

[0005] However, traditional light scattering inversion methods have the following technical challenges: 1) Insufficient Non-spherical Characterization: Aerosols have complex shapes, and dust aerosols exhibit strong non-spherical characteristics. Traditional methods typically assume aerosol particles to be spherical. However, aerosol particles in the actual atmosphere often have diverse morphologies; for example, dust aerosol particles are typically non-spherical. Therefore, using spherical theory to calculate the particle spectral distribution of non-spherical aerosol particles will lead to inaccurate calculations, affecting the accuracy of aerosol particle spectral distribution calculations and further causing inversion errors.

[0006] 2) The physical model involves large computational loads, and the inversion method is inefficient: For example, the invention patent with publication number CN108932357A discloses a physical inversion method for calculating the microphysical properties of atmospheric particulate matter to optical scattering properties. This method obtains the microscopic morphology and composition of atmospheric aerosol particles with complex shapes and multi-component mixing characteristics through actual observation and data collection. It simulates the mixing mode and physicochemical properties of various aerosol components such as black carbon, organic matter, sulfates, soot, dust, and water droplets, and constructs a microphysical model of typical atmospheric particulate matter. Based on the calculation method and model for calculating the microphysical properties of atmospheric particulate matter to optical scattering properties provided by this invention, appropriate optical scattering calculation methods are selected for different atmospheric scenarios and different particulate matter microphysical properties to construct corresponding atmospheric particulate matter optical scattering models and calculate the optical scattering characteristics of aerosols with complex shapes and multi-component mixing. However, this physical model requires manual parameter tuning, the selection of parameters affects the inversion accuracy, and the cost of searching and calculating in high-dimensional parameter space is high, making it difficult to adapt to real-time measurement or large-scale experimental data processing. It is also difficult to jointly invert refractive index, particle size distribution and shape parameters.

[0007] 3) Insufficient physical constraints in deep learning inversion: Deep learning can learn the input-output relationship in a high-dimensional parameter space, but purely data-driven inversion methods are prone to producing physically unreasonable results and have poor interpretability. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for measuring and inverting light scattering of non-spherical dust aerosols based on a physically constrained neural network. This method involves modeling and calculating the discrete optical properties of individual non-spherical dust particles, and then calculating the bulk optical properties through a forward simulation module. A training set is constructed to form physical constraints. A multi-task physically constrained neural network is built, and the inversion parameter error is combined with the physical constraint error. This allows the network to maintain physical consistency while fitting the observed data, achieving high-precision and rapid inversion of microphysical properties such as particle size distribution, refractive index, and shape parameters of dust aerosols from multi-angle light scattering matrix measurement data, while ensuring that the inversion results satisfy physical laws.

[0009] To achieve the above-mentioned objectives, an embodiment provides a method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks, comprising: Step 1: Model the optical properties of non-spherical dust particles and calculate the discrete single-particle optical properties based on the original microphysical parameters of the non-spherical dust particles. Step 2: Based on the discrete optical properties of single particles, the particle size is discretized, the shape mixture average is calculated, and the particle size distribution is integrated through the forward simulation module to calculate the volume optical properties of the non-spherical dust particle group and output the initial optical scattering matrix. Step 3: Construct a neural network with an embedded forward simulation module as a physical constraint neural network. The physical constraint neural network takes the initial optical scattering matrix as input, extracts shared features from the initial optical scattering matrix through a shared layer, and then predicts preliminary microphysical parameters through a multi-task branch layer. The predicted preliminary microphysical parameters are input to the forward simulation module to apply physical constraints, reconstruct the optical scattering matrix, and correct the preliminary predicted microphysical parameters. A joint loss is established for backpropagation to optimize the physical constraint neural network, resulting in the trained physical constraint neural network. The SHAP method based on feature contribution is introduced to achieve interpretability analysis between the input optical scattering matrix and the predicted microphysical parameters. Step 4: In application, multi-angle scattering matrix measurement data of non-spherical dust particles are collected and preprocessed. The preprocessed multi-angle scattering matrix measurement data is then input into the trained physical constraint neural network for inversion, and microphysical parameters are output.

[0010] Preferably, the original microphysical parameters of non-spherical dust particles include the real and imaginary parts of the refractive index, shape parameters, size parameters, and particle size distribution parameters.

[0011] Preferably, in step 1, a hyperellipsoidal model is used to model the optical model of the non-spherical dust particles. The hyperellipsoidal model describes the basic morphology of the non-spherical dust particles by including shape parameters such as aspect ratio and smoothness.

[0012] Preferably, in step 1, when calculating the optical properties of a single particle based on the original microphysical parameters of the non-spherical dust particles, a hybrid calculation strategy is adopted, including: when the scale parameter When the scale parameter is ≤50, the optical properties of a single particle are calculated using the invariant embedding T-matrix method; when the scale parameter is <50, the optical properties of a single particle are calculated using the invariant embedding T-matrix method. When the scale parameter is less than 200, a deep neural network model is used to calculate the optical properties of a single particle; when the scale parameter is less than or equal to 200... When the value is ≤1000, the optical properties of a single particle are calculated using an improved geometric optics method.

[0013] Preferably, in step 2, particle size discretization is performed using a forward simulation module, including: The scale parameters of non-spherical dust particles Convert to radius of an equal-volume sphere The conversion formula is: , In the formula, The scale parameter is represented as The volume corresponding to non-spherical dust particles; The radius of an equal-volume sphere is constrained within the range of 0.05–15 μm, and in logarithmic space... The upper part is divided into multiple discrete particle size intervals to complete the discretization expression of continuous particle size distribution.

[0014] For cases where different non-spherical particles have different radii of equal volume under the same scale parameters, this invention can solve the problem of inconsistent size definitions for non-spherical dust particles by discretizing the particle size, establish a discrete grid for integral calculation, and fix all subsequent calculations within this particle size range.

[0015] More preferably, the discrete particle size range is 22.

[0016] Preferably, in step 2, shape blending averaging is performed using a forward simulation module, including: Given a smoothness, the aspect ratio is uniformly sampled within the particle size range, and for each sampling point, the corresponding discrete single-particle optical properties are extracted. Within each interval, trigonometric integration is performed on the discrete single-particle optical properties of all sampling points to obtain the normalized optical properties within each particle size interval, including the scattering coefficient. , and optical scattering matrix .

[0017] By averaging the shapes, the randomness of the morphology (aspect ratio) of non-spherical dust particles is processed, and representative optical properties of each shape category (defined by smoothness) are obtained.

[0018] More preferably, the particle size range of the uniform sampling is in the range of aspect ratio 0.5-2.0.

[0019] Preferably, in step 2, the particle size distribution integration is performed using a forward simulation module, including: A particle size distribution model was established to describe the particle size distribution of non-spherical dust particle groups; Based on different particle size distributions, the normalized optical properties within each particle size range are weighted and summed to obtain the volume scattering coefficient. Volume extinction coefficient Volume scattering matrix The calculation formula is as follows: , , In the formula, For the total discrete particle size range, For the first The representative volume sphere radius for each particle size range This represents the particle volume concentration distribution value within a unit logarithmic radius interval. The scattering coefficient is... Extinction coefficient, For the optical scattering matrix, For the first The real part of the refractive index , No. The imaginary part of the refractive index , For the first A degree of smoothness.

[0020] More preferably, the established particle size distribution model includes at least one of the following: bimodal log-normal distribution, unimodal log-normal distribution, trimodal log-normal distribution, gamma distribution, and Weiber distribution.

[0021] Preferably, the shared layer in step 3 includes at least one fully connected layer, which employs the ReLU activation function to extract shared features from the optical scattering matrix.

[0022] Preferably, the multi-task branch layer in step 3 includes three parallel task branches, each task branch being connected to a shared layer and including at least one fully connected layer for predicting preliminary microphysical parameters; wherein, the first task branch includes two parallel sub-task branches for inverting the real and imaginary parts of the refractive index, respectively, using the shared features as input; the second task branch is used to invert the shape parameters using the shared features as input; and the third task branch is used to invert the particle size distribution parameters using the shared features as input.

[0023] In one embodiment, the joint loss established in step 3 includes microphysical parameter loss and physical constraint loss, calculated as follows: , In the formula, For joint losses, For the loss of microphysical parameters, For physical constraint loss, These are the weighting coefficients corresponding to the physical constraint loss; Among them, the microphysical parameter loss The calculation formula is: , , In the formula, For the first The microphysical parameter prediction loss for each task. To predict the loss for the microphysical parameters used in the inversion of the real part of the refractive index, To predict the loss for the microphysical parameters of the inverted imaginary part of the refractive index, To predict the loss of microphysical parameters for inverting shape parameters, To predict the loss for the microphysical parameters of the inverted particle size distribution parameters, The number of training samples, For the first The task is The true microphysical parameters, For the first The task is Microphysical parameters predicted by physical constraint neural networks. The square Euclidean norm; The physical constraint loss The calculation formula is: , In the formula, For the first The input scattering matrix of each sample For the first Microphysical parameters predicted by a neural network based on physical constraints for a single sample. This is an embedded forward simulation module.

[0024] Preferably, in step 4, the preprocessing includes: Based on the trained physical constraint neural network, the missing scattering angles in the multi-angle scattering matrix measurement data of the collected non-spherical dust particles are interpolated, and the data is resampled onto the fixed-angle grid used by the trained physical constraint neural network and normalized.

[0025] This invention also provides a physically constrained neural network-based system for measuring and inverting the light scattering of non-spherical dust aerosols, used to achieve the aforementioned physically constrained neural network-based system for measuring and inverting the light scattering of non-spherical dust aerosols, comprising: The non-spherical optical model module is used to model the optical properties of non-spherical dust particles and calculate the discrete single-particle optical properties based on the original microphysical parameters of the non-spherical dust particles. The forward simulation module is used to discretize the particle size, average the shape mixture and integrate the particle size distribution based on the discrete single particle optical properties, calculate the volume optical properties of the non-spherical dust particle group, and output the initial optical scattering matrix. The physical constraint neural network construction and training module takes an initial optical scattering matrix as input, extracts shared features from the initial optical scattering matrix through a shared layer, and then predicts preliminary microphysical parameters through a multi-task branch layer. The predicted preliminary microphysical parameters are input to the forward simulation module to apply physical constraints, reconstruct the optical scattering matrix, and correct the preliminary predicted microphysical parameters, thus completing the construction of the physical constraint neural network. A joint loss is established and backpropagation is used to optimize the physical constraint neural network, resulting in the trained physical constraint neural network. The feature contribution-based SHAP method is introduced to achieve interpretability analysis between the input optical scattering matrix and the predicted microphysical parameters. The inversion and application module is used to collect and preprocess multi-angle scattering matrix measurement data of non-spherical dust particles. The preprocessed multi-angle scattering matrix measurement data is then input into the trained physical constraint neural network for inversion to obtain the actual microphysical parameters of the dust aerosol.

[0026] Compared with the prior art, the beneficial effects of the present invention include at least the following: The present invention provides a method and system for measuring and inverting light scattering of aerosols in non-spherical dust particles based on a physically constrained neural network. Compared with existing technologies, this method performs optical modeling of aerosol particles, which can accurately characterize the scattering characteristics of aerosol particles. Then, through a forward simulation module, it performs integrated calculations on the particle swarm, completes the dataset construction and forms physical constraints, and constitutes a complete optical calculation chain from aerosol modeling to multi-distributed system simulation, supporting end-to-end differentiable calculations. Furthermore, it constructs a multi-task physically constrained neural network and establishes a joint loss function, which can simultaneously take into account the complex scattering characteristics of aerosol particles, physical consistency, and output interpretable results, which helps to rapidly, stably, and accurately invert the microphysical parameters of dust aerosols. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0028] Figure 1 A schematic diagram of the process for measuring and inverting light scattering of non-spherical dust aerosols based on a physically constrained neural network, provided by the present invention.

[0029] Figure 2 This is a schematic diagram of optical models for various non-spherical dust particles.

[0030] Figure 3 This is a schematic diagram of the forward simulation module.

[0031] Figure 4 This is a schematic diagram of a multi-task neural network structure.

[0032] Figure 5 This is a schematic diagram of the structure of a physically constrained neural network.

[0033] Figure 6 A schematic diagram of the structure of the non-spherical dust aerosol light scattering measurement and inversion system based on a physically constrained neural network provided by the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the accompanying drawings and... The embodiments further illustrate the present invention in detail. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0035] To achieve high-precision and rapid inversion of microphysical properties such as particle size distribution, refractive index, and shape parameters of dust aerosols from multi-angle light scattering matrix measurement data, while ensuring that the inversion results conform to physical laws, such as... Figure 1 As shown in the embodiment, an inversion method for measuring light scattering of non-spherical dust aerosols based on physically constrained neural networks includes the following steps: Step 1: Model the optical properties of non-spherical dust particles and calculate the discrete single-particle optical properties based on the original microphysical parameters of the non-spherical dust particles.

[0036] In the embodiments, such as Figure 2 As shown, a parametric geometric model is used to describe particle morphology, which can flexibly represent various non-spherical structures (such as near-ellipsoidal, near-spherical, or angular structures) to facilitate optical scattering characteristic calculations. For ease of calculation, this embodiment uses a hyperellipsoidal model, with aspect ratio... and smoothness Describe the particle morphology. Schematic diagrams of different particle morphologies are shown below. Figure 1 As shown.

[0037] When calculating the optical properties of a single particle based on its original microphysical parameters, a hybrid calculation strategy is employed, considering scale parameters. ( ,in For wave number, , Let be the semi-axis length of the hyperellipsoid along the axis of rotational symmetry. The semi-axis length of the hyperellipsoid on the equatorial plane perpendicular to the rotational symmetry axis is relatively small. For parameters ≤50, the Invariant Embedding T Matrix (IITM) is used for calculation. For large-scale parameters (200≤50≤50), the Invariant Embedding T Matrix (IITM) is used. ≤1000), the improved geometrical optics method (IGOM) was used for calculation, and the neural network method (DNN) was used to calculate the parameters at medium scales (50< Optical properties of a single particle (<200).

[0038] By employing optical solution methods for non-spherical particles, different optical property solution methods are used based on the scale parameter range of non-spherical dust particles to balance computational accuracy and efficiency. The generated data is further used to generate a neural network training set and to form physical constraints.

[0039] Step 2: Based on the discrete optical properties of single particles, the particle size is discretized, the shape is mixed and averaged, and the particle size distribution is integrated through the forward simulation module. The volume optical properties of the non-spherical dust particle group are calculated, and the initial optical scattering matrix is ​​output.

[0040] In the embodiments, such as Figure 3 As shown, based on the input microphysical parameters, such as the real part of the refractive index... and the virtual part Aspect Ratio Smoothness And particle size distribution parameters (such as the geometric standard deviation of the coarse mode in a bimodal log-normal distribution). and mean Geometric standard deviation of fine modes and mean coarse mode weights w This module calculates the optical scattering matrix of non-spherical dust particles. It is also used to generate the training set and to embed it as a physical constraint into the neural network framework.

[0041] 2.1. Calculation of the radius of a sphere with equal volume: For non-spherical particles, for a given wavelength... (Taking 632nm as an example) and scale parameters sphere radius of equal volume The conversion formula is: , In the formula, The scale parameter is represented as The volume corresponding to non-spherical dust particles; 2.2. Particle Size Unification Strategy: Because the volume radii of different non-spherical particles differ under the same scale parameter, this module unifies the particle size representation for different particle shapes. First, the particle size range is uniformly constrained to 0.05~15 μm, and then... The particle size distribution is divided into 22 discrete particle size intervals. Then, within each interval, particles of different shapes that meet the specified conditions are uniformly processed, and their optical properties, including the scattering coefficient, are obtained through integrated calculations. , and optical scattering matrix This strategy ensures that particles of different shapes have consistent particle size characterization and optical properties within the same particle size range.

[0042] 2.3. Optical Property Calculation: First, a mixed average is applied to the shape. For each smoothness... Uniform sampling is performed within an aspect ratio range of 0.5 to 2.0, and optical properties are averaged to obtain optical properties with single smoothness control. Simultaneously, the refractive index range is considered. , Given In the kernel function, the normalized optical properties within each particle size range are obtained by matching the discrete parameter grid through the nearest neighbor search.

[0043] 2.4. Integrating to determine volume scattering characteristics: based on different particle size distributions The volume scattering coefficient is obtained by weighted summation of the optical properties for each particle size range. Volume extinction coefficient Volume scattering matrix : , , In the formula, For the total discrete particle size range, For the first The representative volume sphere radius for each particle size range This represents the particle volume concentration distribution value within a unit logarithmic radius interval. The scattering coefficient is... Extinction coefficient, For the optical scattering matrix, For the first The real part of the refractive index , No. The imaginary part of the refractive index , For the first A degree of smoothness.

[0044] 2.5. Particle Size Distribution Modeling: In this embodiment, the particle size distribution is assumed to be a bimodal log-normal distribution, including coarse and fine modes. The specific formula is as follows: , In the formula, For the first The volume corresponding to a non-spherical particle and These are the geometric standard deviation and mean of the coarse and fine modes, respectively. The weights are for the coarse modes, with the total volume concentration fixed at 1 μm. -3 cm -3 .

[0045] It should be emphasized that the bimodal log-normal distribution is a preferred, not a limiting, approach of this invention. The particle size distribution model can also be extended to unimodal, trimodal, or other modal log-normal distributions, or other distribution functions such as gamma distribution and Weiber distribution that can be used to describe particle spectra can be employed.

[0046] 2.6. Module Output: The final output of the module is a volume scattering matrix. .

[0047] 2.7. Differentiable Implementation: In this process, the relevant calculations are based on the Tensorflow framework to ensure that the calculations are differentiable, thereby realizing gradient backpropagation.

[0048] Step 3: Construct a neural network with an embedded forward simulation module as a physical constraint neural network. The physical constraint neural network takes the initial optical scattering matrix as input, extracts shared features from the initial optical scattering matrix through a shared layer, and then predicts preliminary microphysical parameters through a multi-task branch layer. The predicted preliminary microphysical parameters are input to the forward simulation module to apply physical constraints, reconstruct the optical scattering matrix, and correct the preliminary predicted microphysical parameters. A joint loss is established for backpropagation optimization of the physical constraint neural network to obtain the trained physical constraint neural network. The SHAP method based on feature contribution is introduced to achieve interpretability analysis between the input optical scattering matrix and the predicted microphysical parameters.

[0049] In this embodiment, constructing a physically constrained neural network is the core component of this patent, used for high-precision and rapid inversion of the microphysical parameters of non-spherical dust particles based on multi-angle scattering matrix data. The physically constrained neural network achieves physical consistency of the inversion results by embedding a differentiable forward simulation module, and also possesses interpretability. Specifically: 3.1. Multi-task Neural Network MT-FCNN: Constructing a multi-task fully connected neural network for the simultaneous inversion of multiple microphysical parameters, one preferred embodiment of which has the following network structure: Figure 4 As shown, the network input consists of observations of six scattering matrix elements at 42 scattering angles, flattened into a 252-dimensional vector. The network includes a shared layer and task-specific branches. The shared layer is composed of multiple stacked fully connected layers, serving as a general feature extractor. In this embodiment, the shared layer contains four fully connected layers with the following numbers of neurons: 512, 256, 128, and 64, respectively. Each layer uses the ReLU activation function.

[0050] Following the shared layer, the network has three branches, each composed of stacked fully connected layers, each handling a different task and learning specific parameters. Branch one is further divided into two parallel sub-branches, one for retrieving the real part of the refractive index. (Task 1) and the virtual part (Task 2), Branch 2 is used to invert the shape parameter smoothness. (Task 3), Branch 3 is used to invert particle size distribution parameters. (Task 4). The specific number of layers and the size of neurons in each branch can be designed and optimized according to the complexity of the actual task. Through the above multi-task learning architecture, this framework can simultaneously and efficiently obtain multiple microphysical parameters such as the refractive index, shape characteristics, and particle size distribution of particles from the aforementioned multi-angle scattering matrix, effectively improving data utilization efficiency and model generalization ability.

[0051] 3.2. Physical Constraint Embedding: Incorporating the forward simulation module ( Figure 3 ) are embedded in MT-FCNN in the form of differentiable structures ( Figure 4 This constitutes an end-to-end, trainable physical constraint neural network. Specifically, the microphysical parameters predicted by MT-FCNN are first input into the forward simulation module to reconstruct the optical scattering matrix, and then compared with the input real measured optical scattering matrix to ensure that the microphysical parameters inverted by the network automatically satisfy physical laws, thereby correcting the initially predicted microphysical parameters.

[0052] 3.3. Joint Loss and Training Optimization: such as Figure 5 As shown, the joint loss for the entire training process consists of two parts. First, Output the micro-physics parameter prediction error for MT-FCNN: , , In the formula, For the first The microphysical parameter prediction loss for each task. To predict the loss for the microphysical parameters used in the inversion of the real part of the refractive index, To predict the loss for the microphysical parameters of the inverted imaginary part of the refractive index, To predict the loss of microphysical parameters for inverting shape parameters, To predict the loss for the microphysical parameters of the inverted particle size distribution parameters, The number of training samples, For the first The task is The true microphysical parameters, For the first The task is Microphysical parameters predicted by physical constraint neural networks. It is the square Euclidean norm.

[0053] In addition, physical constraint loss The formula for calculating the squared difference between the reconstructed optical scattering matrix and the input true measured optical scattering matrix is ​​as follows: , In the formula, For the first The input scattering matrix of each sample For the first Microphysical parameters predicted by a neural network based on physical constraints for a single sample. For embedded positive modules.

[0054] The corresponding joint loss function is: , In the formula, For joint losses, For the loss of microphysical parameters, For physical constraint loss, This represents the weighting coefficient corresponding to the physical constraint loss. In this embodiment, we set φ = 1.

[0055] During the training process, The gradient is backpropagated along the MT-FCNN to update the weight parameters of the MT-FCNN. The gradient propagates simultaneously along the forward simulation module and the backward propagation of MT-FCNN, thus jointly updating the parameters of the entire physical constraint neural network. The training process is as follows: Figure 3 As shown. During training, gradient descent-like optimization methods can be used, combined with strategies such as learning rate decay and early stopping, to improve stability and avoid overfitting.

[0056] 3.4. Interpretability Implementation: To improve interpretability, after the physical constraint neural network training is completed, this invention further introduces the SHAP method based on feature contribution.

[0057] First, the pre-trained physical constraint neural network to be interpreted is loaded, and a representative subset of the training data is extracted as the background dataset. Then, for any input sample (multi-angle scattering matrix vector), the SHAP method is used to calculate the SHAP value of each input feature in the physical constraint neural network for each microphysical parameter (such as refractive index, shape parameter, particle size distribution, etc.) in the final output of the physical constraint neural network. Finally, by analyzing the obtained SHAP values, the dominant scattering matrix elements of each key inversion microphysical parameter can be identified globally, and the specific influence of each input scattering matrix on the final inversion parameter values ​​can be shown. This helps to evaluate the contribution of different scattering matrices to the inversion parameters, verify whether the model predictions conform to physical priors, and effectively improve the interpretability and reliability of the model.

[0058] Step 4: In application, multi-angle scattering matrix measurement data of non-spherical dust particles are collected and preprocessed. The preprocessed multi-angle scattering matrix measurement data is then input into the trained physical constraint neural network for inversion, and microphysical parameters are output.

[0059] In this embodiment, this step is used to apply the trained inversion framework to actual observation data to achieve the inversion of microphysical parameters of dust aerosols. Specifically: 4.1. Data Acquisition: Obtain multi-angle scattering matrix measurement data of non-spherical dust particles, and extract the six required scattering matrix elements: P 11 P 12 / P11 P 22 / P11 P 33 / P11 P 34 / P11 P 44 / P11 And the corresponding scattering angle θ.

[0060] 4.2. Data Processing: Based on the input angle required by the neural network, the missing scattering angle is interpolated, and the data is resampled onto the fixed-angle grid used by the neural network.

[0061] 4.3. Normalization: For P 11 Normalization is performed at θ=30°, and the six elements are concatenated in sequence to form 6× The input vector, where This represents the number of scattering angles.

[0062] 4.4. Physically Constrained Neural Network Inversion: The processed vectors are input into the trained physically constrained neural network, which outputs the inverted microphysical parameters of the dust aerosol, including particle size distribution parameters, refractive index, and particle shape parameters.

[0063] Through the above embodiments, the present invention can achieve high-precision and rapid inversion of microphysical parameters of non-spherical dust particles, with both physical consistency and model interpretability, and is suitable for inversion of multi-angle light scattering measurement data.

[0064] Based on the same inventive concept, the embodiment also provides a non-spherical dust aerosol light scattering measurement and inversion system based on a physically constrained neural network, used to realize the aforementioned non-spherical dust aerosol light scattering measurement and inversion based on a physically constrained neural network, such as... Figure 6 As shown, it includes: The non-spherical optical model module is used to model the optical properties of non-spherical dust particles and calculate the discrete single-particle optical properties based on the original microphysical parameters of the non-spherical dust particles. The forward simulation module is used to discretize the particle size, average the shape mixture and integrate the particle size distribution based on the discrete single particle optical properties, calculate the volume optical properties of the non-spherical dust particle group, and output the initial optical scattering matrix. The physical constraint neural network construction and training module takes an initial optical scattering matrix as input, extracts shared features from the initial optical scattering matrix through a shared layer, and then predicts preliminary microphysical parameters through a multi-task branch layer. The predicted preliminary microphysical parameters are input to the forward simulation module to apply physical constraints, reconstruct the optical scattering matrix, and correct the preliminary predicted microphysical parameters, thus completing the construction of the physical constraint neural network. A joint loss is established and backpropagation is used to optimize the physical constraint neural network, resulting in the trained physical constraint neural network. The feature contribution-based SHAP method is introduced to achieve interpretability analysis between the input optical scattering matrix and the predicted microphysical parameters. The inversion and application module is used to collect and preprocess multi-angle scattering matrix measurement data of non-spherical dust particles. The preprocessed multi-angle scattering matrix measurement data is then input into the trained physical constraint neural network for inversion to obtain the actual microphysical parameters of the dust aerosol.

[0065] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for measuring and inverting light scattering in non-spherical dust aerosols based on physically constrained neural networks, characterized in that, Includes the following steps: Step 1: Model the optical properties of non-spherical dust particles and calculate the discrete single-particle optical properties based on the original microphysical parameters of the non-spherical dust particles. Step 2: Based on the discrete optical properties of single particles, the particle size is discretized, the shape mixture average is calculated, and the particle size distribution is integrated through the forward simulation module to calculate the volume optical properties of the non-spherical dust particle group and output the initial optical scattering matrix. Step 3: Construct a neural network with an embedded forward simulation module as a physical constraint neural network. The physical constraint neural network takes the initial optical scattering matrix as input, extracts shared features from the initial optical scattering matrix through a shared layer, and then predicts preliminary microphysical parameters through a multi-task branch layer. The predicted preliminary microphysical parameters are input to the forward simulation module to apply physical constraints, reconstruct the optical scattering matrix, and correct the preliminary predicted microphysical parameters. A joint loss is established and used for backpropagation to optimize the physical constraint neural network, resulting in the trained physical constraint neural network. Furthermore, the SHAP method based on feature contribution is introduced to achieve interpretability analysis between the input optical scattering matrix and the predicted microphysical parameters; Step 4: In application, multi-angle scattering matrix measurement data of non-spherical dust particles are collected and preprocessed. The preprocessed multi-angle scattering matrix measurement data is then input into the trained physical constraint neural network for inversion, and microphysical parameters are output.

2. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 1, characterized in that, The original microphysical parameters of non-spherical dust particles include the real and imaginary parts of the refractive index, shape parameters, scale parameters, and particle size distribution parameters.

3. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 2, characterized in that, In step 1, a hyperellipsoidal model is used to model the optical model of non-spherical dust particles. The hyperellipsoidal model describes the basic morphology of non-spherical dust particles by including shape parameters such as aspect ratio and smoothness.

4. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 2, characterized in that, In step 2, particle size discretization is performed using a forward simulation module, including: The scale parameters of non-spherical dust particles Convert to radius of an equal-volume sphere The conversion formula is: ， In the formula, The scale parameter is represented as The volume corresponding to non-spherical dust particles; The radius of an equal-volume sphere is constrained within the range of 0.05–15 μm, and in logarithmic space... The upper part is divided into multiple discrete particle size intervals to complete the discretization expression of continuous particle size distribution.

5. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 3, characterized in that, In step 2, shape blending averaging is performed using the forward simulation module, including: Given a smoothness, the aspect ratio is uniformly sampled within the particle size range, and for each sampling point, the corresponding discrete single-particle optical properties are extracted. Within each interval, trigonometric integration is performed on the discrete single-particle optical properties of all sampling points to obtain the normalized optical properties within each particle size interval, including the scattering coefficient. , and optical scattering matrix .

6. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 5, characterized in that, In step 2, the particle size distribution is integrated using the forward simulation module, including: A particle size distribution model was established to describe the particle size distribution of non-spherical dust particle groups; Based on different particle size distributions, the normalized optical properties within each particle size range are weighted and summed to obtain the volume scattering coefficient. Volume extinction coefficient Volume scattering matrix The calculation formula is as follows: ， ， In the formula, For the total discrete particle size range, For the first The representative volume sphere radius for each particle size range This represents the particle volume concentration distribution value within a unit logarithmic radius interval. The scattering coefficient is... Extinction coefficient, For the optical scattering matrix, For the first The real part of the refractive index , No. The imaginary part of the refractive index , For the first A degree of smoothness.

7. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 1, characterized in that, The shared layer in step 3 includes at least one fully connected layer, which employs the ReLU activation function to extract shared features from the initial optical scattering matrix.

8. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 1, characterized in that, The multi-task branch layer in step 3 includes three parallel task branches, each of which is connected to a shared layer and includes at least one fully connected layer for predicting preliminary microphysical parameters. Specifically, the first task branch contains two parallel sub-task branches, which are used to invert the real and imaginary parts of the refractive index, respectively, using the shared features as input. The second task branch is used to invert the shape parameters using the shared features as input. The third task branch is used to invert the particle size distribution parameters using the shared features as input.

9. The method for measuring and inverting light scattering of non-spherical dust aerosols based on physically constrained neural networks according to claim 7 or 8, characterized in that, The joint loss established in step 3 includes microphysical parameter loss and physical constraint loss, and the calculation formula is as follows: ， In the formula, For joint losses, For the loss of microphysical parameters, For physical constraint loss, These are the weighting coefficients corresponding to the physical constraint loss; Among them, the microphysical parameter loss The calculation formula is: ， ， In the formula, For the first The microphysical parameter prediction loss for each task. To predict the loss for the microphysical parameters used in the inversion of the real part of the refractive index, To predict the loss for the microphysical parameters of the inverted imaginary part of the refractive index, To predict the loss of microphysical parameters for inverting shape parameters, To predict the loss for the microphysical parameters of the inverted particle size distribution parameters, The number of training samples, For the first The task is The true microphysical parameters, For the first The task is Microphysical parameters predicted by physical constraint neural networks. The square Euclidean norm; The physical constraint loss The calculation formula is: ， In the formula, For the first The input scattering matrix of each sample For the first Microphysical parameters predicted by a neural network based on physical constraints for a single sample. This is an embedded forward simulation module.

10. A measurement and inversion system for light scattering of non-spherical dust aerosols based on physically constrained neural networks, characterized in that, For implementing the light scattering measurement and inversion of non-spherical dust aerosols based on physically constrained neural networks as described in any one of claims 1-9, the method includes: The non-spherical optical model module is used to model the optical properties of non-spherical dust particles and calculate the discrete single-particle optical properties based on the original microphysical parameters of the non-spherical dust particles. The forward simulation module is used to discretize the particle size, average the shape mixture and integrate the particle size distribution based on the discrete single particle optical properties, calculate the volume optical properties of the non-spherical dust particle group, and output the initial optical scattering matrix. The physical constraint neural network construction and training module takes an initial optical scattering matrix as input, extracts shared features from the initial optical scattering matrix through a shared layer, and then predicts preliminary microphysical parameters through a multi-task branch layer. The predicted preliminary microphysical parameters are input to the forward simulation module to apply physical constraints, reconstruct the optical scattering matrix, and correct the preliminary predicted microphysical parameters, thus completing the construction of the physical constraint neural network. A joint loss is established and backpropagation is used to optimize the physical constraint neural network, resulting in the trained physical constraint neural network. The feature contribution-based SHAP method is introduced to achieve interpretability analysis between the input optical scattering matrix and the predicted microphysical parameters. The inversion and application module is used to collect and preprocess multi-angle scattering matrix measurement data of non-spherical dust particles. The preprocessed multi-angle scattering matrix measurement data is then input into the trained physical constraint neural network for inversion to obtain the actual microphysical parameters of the dust aerosol.

Images