A machine learning-based passive radiative cooling film inverse design method

Abstract

The application discloses a passive radiation cooling film reverse design method based on machine learning, which collects optical characteristic data of a base material through experiments, establishes a basic data set, and divides a training set and a test set. On this basis, a prediction model of components and thickness to optical characteristic data is established, and a reverse design model of optical characteristic data to components and thickness is established. The prediction model and the reverse design model are trained, the initial design value of the components and the thickness under the target optical characteristics is calculated through the reverse design model, and the optical characteristic data obtained through optical simulation is compared, when the mean square error meets the requirement, the design of the passive radiation cooling film is completed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of passive energy utilization, and particularly relates to a passive radiation refrigeration film reverse design method based on machine learning BACKGROUND

[0002] With the increasingly serious global warming and energy crisis, energy saving and emission reduction has gradually become an important research direction in various fields. Traditional active refrigeration methods face problems such as low efficiency and environmental pollution. Passive radiation refrigeration technology utilizes the high emissivity of materials in the mid-infrared range to transfer heat to space through radiation, thereby achieving natural cooling. As a new type of energy utilization technology, passive radiation refrigeration does not rely on traditional power-driven systems, has significant energy-saving advantages and environmental protection characteristics, and has shown broad application prospects in building energy saving, cold chain transportation, electronic device heat dissipation and other fields. However, the manufacturing of passive radiation refrigeration films still faces many challenges. In the manufacturing process of traditional passive radiation refrigeration films, the selection of film substrates and the design of films mainly depend on experience and experiments, and lack of systematic theoretical guidance. Existing design methods usually rely on a single material property, such as emissivity or refractive index, ignoring the interaction of multiple optical properties and material thickness, resulting in a large uncertainty in the design process. In addition, the selection and optimization of substrates in traditional design methods often rely on a large amount of experimental data and repeated verification, which is low in design efficiency and difficult to cope with changing environmental conditions and changing application requirements.

[0003] In the existing disclosed technology, Su et al. [1] Two structures of quadrilateral prism metasurface and circular cone metasurface are constructed, and K-neighbor algorithm is used to predict the optimal structure parameters and corresponding absorption / emission in the solar spectrum and atmospheric window, providing a new idea for the design of passive radiation coolers. However, the data preprocessing process in the construction of the model is relatively simple, and problems such as data noise and data imbalance may not be thoroughly discussed, which may affect the generalization ability of the model. Ding et al. [2] Deep learning models reveal the complex relationship between biomimetic metamaterials and their spectral responses, accelerating the design and optimization of robust biomimetic radiation cooling metamaterials and making significant progress in standardized passive radiation cooling applications.

[0004] To solve the above problems, the present application provides a passive radiative cooling film reverse design method based on machine learning. This method combines machine learning technology with the design process of passive radiative cooling film, optimally selects appropriate material types and thicknesses according to specific spectral emissivity requirements, and verifies its cooling effect through optical simulation. Compared with traditional design methods, the present application significantly improves the efficiency and accuracy of design by establishing efficient performance prediction models and reverse design models, while effectively reducing experimental and development costs.

[0005] REFERENCES

[0006] [1]Wei Su;Zhipeng Ding;Yinlong Luo;Lipengan Ye;Hong Wu;Hongbing Yao. Machine learning-enabled design of metasurface based near-perfect daytimeradiative cooler[J]. Solar Energy Materials and Solar Cells, 2023, 260:112488.

[0007] [2]Zhenmin Ding;Xin Li;Qingxiang Ji;Yunce Zhang;Honglin Li;Hulin Zhang;Lorenzo Pattelli;Yao Li;Hongbo Xu;Jiupeng Zhao. Machine-learning-assisted design of a robust biomimetic radiative cooling metamaterial[J]. ACS Materials Letters, 2024, 6(6):2416-2424. SUMMARY

[0008] The present application aims to provide a passive radiative cooling film reverse design method based on machine learning, which optimizes the design process of passive radiative cooling film by combining machine learning technology with the design requirements of passive radiative cooling film, thereby reducing the time and cost of the design process.

[0009] To achieve the purpose of the present application, the present application adopts the following technical solutions:

[0010] S10 collects optical property data of passive radiative cooling film substrates through experiments to establish a basic data set;

[0011] S20 pre-processes the optical property data collected in S10, and divides the training set and the test set using K-fold cross-validation method;

[0012] S30 establishes a prediction model of the composition and thickness of the passive radiative cooling film to the optical property data according to the basic data set, and trains the model;

[0013] S40 establishes a reverse design model of the optical property data to the composition and thickness of the passive radiative cooling film according to the basic data set, and trains the model;

[0014] S50 inputs the target optical property data into the reverse design model of the optical property data to the composition and thickness of the passive radiative cooling film trained in S40, to obtain the initial design value of the composition and thickness of the passive radiative cooling film;

[0015] S60 performs optical simulation on the initial design value of the composition and thickness of the passive radiative cooling film obtained in S50, to predict the emissivity of the passive radiative cooling film in the atmospheric window wave band and the reflectivity of the passive radiative cooling film in the solar wave band;

[0016] S70 calculates the mean square error of the emissivity of the passive radiative cooling film in the atmospheric window wave band and the reflectivity of the passive radiative cooling film in the solar wave band obtained in S60 and the target optical property data in S50;

[0017] S80 judges whether the mean square error is less than 0.1, if the mean square error is less than 0.1, the design of the passive radiative cooling film is completed, if the mean square error is greater than or equal to 0.1, steps S30-S80 are repeated.

[0018] In S10, the passive radiative cooling film substrate is one or more of Al, Al2O3, Au, BaF2, Be, CaF2, Cr, Cu, Fe, Mn, KBr, SiO2, Si3N4, SiC, Ti, TiO2, ZnO, ZnS, PDMS, PE, PVA, and PVDF.

[0019] The optical property data includes one or more of the refractive index, the reflectivity, the emissivity, and the extinction coefficient of the passive radiative cooling film substrate at the corresponding wavelength.

[0020] The collected basic data set is divided into the following two matrices:

[0021] X i ={element i ,h i};

[0022] Y i ={n i ,R i ,ε i ,ki ,}

[0023] wherein h i is the thickness of the passive radiative cooling film substrate; n i is the refractive index of the passive radiative cooling film substrate; R i is the reflectivity of the passive radiative cooling film substrate; ε i is the emissivity of the passive radiative cooling film substrate; k i is the extinction coefficient of the passive radiative cooling film substrate.

[0024] In S20, the optical characteristic data collected in S10 is preprocessed, data cleaning, standardization and normalization processing are performed, abnormal values and missing data are removed, and data quality is ensured.

[0025] First, the missing values of a certain feature d j are filled by the mean filling method:

[0026]

[0027] wherein M is the number of non-missing values.

[0028] The abnormal value of a certain feature value d j is defined as a value greater than 3 times the mean μ j 3 standard deviation σ j , and the abnormal value is deleted:

[0029] d j -μ j |>3σ j

[0030] Robust standardization method is adopted to standardize the variable, and the data is compressed to the range of [-1.1];

[0031]

[0032] wherein Q1 and Q3 are the 25th and 75th percentiles of the data, respectively, and Q2 is the median.

[0033] Assuming that the minimum and maximum values of a feature d j are min j and max j , respectively, the normalization formula is:

[0034]

[0035] The data set is divided into training set and test set by K-fold cross-validation method, and K=10 is taken. In order to reduce the risk of overfitting of the model on the basic data set, so as to improve the generalization ability of the model, so as to ensure the efficiency and accuracy of the machine learning model.

[0036] The step S30 specifically comprises:

[0037] S31 takes the composition and thickness of the passive radiative cooling film in the basic data set as input, and the optical property data as output, and constructs a prediction model of the composition and thickness of the passive radiative cooling film to the optical property data;

[0038] X i is taken as the input feature vector of composition and thickness, and Y i is taken as the corresponding optical property data as the output vector.

[0039] The convolutional neural network model is used for prediction. For one-dimensional data, one-dimensional convolution operation is used:

[0040] F1=X*W1+b1

[0041] Where, X is the input data, * is the convolution operation, W1 is the convolution kernel, b1 is the bias term, and F1 is the output feature map of the convolution layer.

[0042] The ReLU activation function is used for nonlinear conversion:

[0043] A1=ReLU(F1)

[0044] The definition of ReLU function is:

[0045] ReLU(x)=max(0,x)

[0046] Max pooling is used to select the maximum value in the local area, so as to reduce the dimension of data and improve the robustness of the model:

[0047] P1=MaxPool(A1)

[0048] After convolution and pooling operation, the extracted features are mapped to the final output through full connection layer. Each node is connected with all nodes of the previous layer:

[0049]

[0050] Where, is the predicted optical property data of the model, W2 and b2 are the weights and bias of the full connection layer.

[0051] Finally, a prediction model of the composition and thickness of the passive radiative cooling film to the optical property data is constructed:

[0052]

[0053] wherein, is the optical property of the model output, x is the input of the composition and thickness, f is the mapping function learned by the convolutional neural network, and θ is the parameters (including weights and biases) of all convolutional layers, pooling layers, and fully connected layers.

[0054] S32 trains the prediction model of the composition and thickness of the passive radiative cooling film to the optical property data constructed in S31 using the training set, and tests the accuracy of the trained prediction model of the composition and thickness of the passive radiative cooling film to the optical property data using the test set.

[0055] S33 selects the mean square error as the loss function to measure the gap between the predicted value and the actual value:

[0056]

[0057] The Adam optimizer is selected to adaptively adjust the learning rate, and the update formula of Adam is:

[0058]

[0059]

[0060] wherein, α is the learning rate, β1 and β2 are the decay coefficients of the first and second moments, and ε is a constant to prevent division by zero.

[0061] S34 selects one of forward propagation, loss function calculation, back propagation, and parameter update as the training method, and trains the prediction model of the composition and thickness of the passive radiative cooling film to the optical property data to an accuracy of not less than 0.9.

[0062] wherein, the forward propagation transmits the input optical property data y i The input optical property data y is transmitted into the convolutional neural network, and after passing through each layer (convolutional layer, pooling layer, and fully connected layer), the predicted value is generated, i.e., the composition and thickness data.

[0063] The back propagation algorithm calculates the gradient of the loss function with respect to each parameter, transmits the error from the output layer back to the input layer through the chain rule, and updates each weight in the network.

[0064] The Adam optimizer is used to update the parameters of the convolutional neural network according to the calculated gradient.

[0065] The above steps are repeated until the loss function converges and the training accuracy reaches the required accuracy (not less than 0.9). The training goal is to make the inverse design model of the optical property data to the composition and thickness data perform well on the test set.

[0066] The step S40 specifically comprises:

[0067] S41, taking the optical property data in the basic data set as input, and the composition and thickness of the passive radiative cooling film as output, constructing an inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film;

[0068] Y i as the input feature vector of the optical property data, X i as the corresponding composition and thickness as the output vector.

[0069] The convolutional neural network model is used for prediction, and the model construction process is similar to S31.

[0070] S42, the training set is used to train the inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film constructed in S41; and the test set is used to test the precision of the trained inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film. The model training process is similar to S32;

[0071] S43, selecting mean square error as the loss function, and selecting Adam optimizer to adaptively adjust the learning rate;

[0072] S44, the selected training method is one of forward propagation, loss function calculation, back propagation and parameter update; and the inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film is trained to a precision not less than 0.9.

[0073] In S50, the target optical property data is the emissivity of the passive radiative cooling film in the atmospheric window band and the reflectivity in the solar band.

[0074] As a preferred example, the emissivity of the passive radiative cooling film in the atmospheric window band is not less than 0.9, and the reflectivity in the solar band is not less than 0.9.

[0075] In S60, the optical simulation uses the transfer matrix method to simulate the optical response of the passive radiative cooling film to calculate the optical property data of the passive radiative cooling film substrate at different wavelengths.

[0076] Suppose that light waves are incident on the passive radiative cooling film from the air, and the film is below another medium (such as air or substrate). The whole system can be regarded as a three-layer structure.

[0077] For the passive radiative cooling film, the core of the transfer matrix method for calculating reflectivity and transmittance lies in the optical impedance and phase change of each layer.

[0078] Optical impedance Z0 in air: The impedance of air is usually known and approximately constant Z0 = 1 (in light wave propagation).

[0079] Optical impedance Z in passive radiative cooling film film Impedance of the film is related to its refractive index n film and the incident angle θ film , expressed as:

[0080]

[0081] Transmission matrix M of passive radiative cooling film film is:

[0082]

[0083] where δ = k film d film is the phase change of the passive radiative cooling film, is the wave number in the film, d film is the thickness of the passive radiative cooling film, and λ is the wavelength of the incident light, Z film is the optical impedance of the film, and n film is the refractive index of the film.

[0084] Reflectivity refers to the proportion of incident light reflected back at the surface of the film. Reflection between the film and air can be calculated by the transmission matrix method:

[0085]

[0086] where Z0 is the optical impedance of air, Z film is the optical impedance of the film, θ air and θ film are the incident angles of light in air and the film, respectively.

[0087] Transmittance refers to the proportion of light transmitted through the film, and the calculation formula is:

[0088]

[0089] Emissivity reflects the ability of the film material to radiate thermal energy in a specific waveband. It can be calculated by the following formula:

[0090] ε = 1 - R - A

[0091] A = 1 - R - T

[0092] where A is the absorption rate of the film, and ε is the emissivity of the film.

[0093] Extinction coefficient is related to the refractive index of the film and the absorption characteristics of the material. Its calculation formula is:

[0094]

[0095] where Im(n film ) is the imaginary part of the complex refractive index of the thin film.

[0096] The mean square error between the predicted emissivity of the passive radiative cooling thin film in the atmospheric window band and the target optical property data is calculated by optical simulation. If the mean square error is less than 0.1, the design is complete; if the error is larger, repeat steps S30-S80 to re-adjust the composition and thickness, and further optimize the design.

[0097] The present application has the following advantages and beneficial effects:

[0098] The present application combines machine learning with the design of passive radiative cooling thin films, achieving accurate optimization of the structure and composition of passive radiative cooling thin films according to target spectral emissivity. Unlike traditional empirical design, this method automatically adjusts the composition of the thin film substrate and the thickness of the thin film using experimental data and numerical simulation, thereby optimizing the cooling performance of the radiative thin film, significantly shortening the design cycle and reducing development costs. This technology not only improves design efficiency and accuracy, but also allows for customized design according to different application requirements, has wide application potential, especially in the fields of building energy saving, electronic device heat dissipation, photovoltaic systems and air conditioning, and promotes the development of low-carbon and green technology. BRIEF DESCRIPTION OF DRAWINGS

[0099] Figure 1 is a flowchart of the present application;

[0100] Figure 2 is an implementation flowchart of an embodiment of the present application;

[0101] Figure 3 is an optical simulation diagram constructed by the present application. DETAILED DESCRIPTION

[0102] The technical solutions of the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0103] This embodiment uses a machine learning algorithm to optimize the design of passive radiative cooling thin films. The research goal is to achieve efficient cooling effect by optimizing the emissivity of the thin film in the atmospheric window band and the reflectivity in the solar band, especially during the day and at night. The optimization method based on machine learning helps to automatically and accurately find the most suitable substrate composition and combination, as well as the thickness of the thin film.

[0104] Al2O3 (alumina), PDMS (polydimethylsiloxane), SiO2 (silicon dioxide), and Al (aluminum) are selected as the thin film substrate, and the optical property data of the passive radiative cooling thin film substrate is collected through experiments to establish a basic data set;

[0105] The basic dataset is preprocessed to interpolate missing values in the experimental data, ensuring data integrity. Robust standardization is used to normalize the optical property data, compressing it to the range [-1, 1] and reducing the impact of dimensional differences on the machine learning model. The data is divided into 10 subsets, and K-fold cross-validation (K=10) is used to ensure the model's generalization ability.

[0106] Machine learning model training:

[0107] 1) Predictive model from composition and thickness to optical property data:

[0108] A predictive model is constructed using composition and thickness as input and optical property data as output. The model is trained using a neural network, with an Adam optimizer and mean squared error as the loss function. The model's training accuracy is above 0.9, ensuring its good predictive ability for the optical properties of passive radiative cooling film substrates.

[0109] 2) Reverse design model from optical property data to composition and thickness:

[0110] A reverse design model is constructed based on optical property data. This reverse design model inputs the optical property data of the substrate and outputs the composition and thickness of the thin film. The reverse design model is trained through backpropagation, achieving a prediction mean squared error of less than 0.1.

[0111] The target optical property data (emissivity of passive radiative cooling film in the atmospheric window band and reflectivity in the solar band) is input into the trained reverse design model to obtain the composition and thickness of the passive radiative cooling film substrate. The results are shown in the table below.

[0112] Table 1 Initial design results of passive radiative cooling film

[0113]

[0114] The transmission matrix method is used to simulate the optical properties of the initially designed thin film, predicting the reflectivity and emissivity of the passive radiative cooling film at different wavelengths.

[0115] The mean squared error between the simulated emissivity of the passive radiative cooling film in the atmospheric window band and the target value is calculated. If the error is greater than 0.1, the composition and thickness are adjusted, and the next round of training and optimization is performed.

[0116] After multiple iterations, the final designed thin film has an emissivity in the atmospheric window band not less than 0.9, and a reflectivity in the solar band not less than 0.9, meeting the target optical property requirements.

[0117] The above-described specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application, and it should be understood that the above-described specific embodiments are merely examples of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for inverse design of passive radiative cooling film based on machine learning, characterized in that, The method comprises the following steps: S10, collecting optical property data of the passive radiative cooling film substrate through experiments to establish a basic data set; S20, preprocessing the optical property data collected in S10, and dividing a training set and a test set using K-fold cross-validation method; S30, establishing a prediction model of the composition and thickness of the passive radiative cooling film to the optical property data according to the basic data set, and training the prediction model; The specific steps of S30 comprise: S31, taking the composition and thickness of the passive radiative cooling film in the basic data set as input and taking the optical property data as output to construct a prediction model of the composition and thickness of the passive radiative cooling film to the optical property data; S32, training the prediction model of the composition and thickness of the passive radiative cooling film to the optical property data constructed in S31 using the training set, and testing the accuracy of the trained prediction model of the composition and thickness of the passive radiative cooling film to the optical property data using the test set; S33, selecting mean square error as a loss function and selecting Adam optimizer to adaptively adjust a learning rate; S34, selecting one of forward propagation, loss function calculation, back propagation and parameter update as a training method, and training the prediction model of the composition and thickness of the passive radiative cooling film to the optical property data to an accuracy of not less than 0.9; S40, establishing an inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film according to the basic data set, and training the inverse design model; The specific steps of S40 comprise: S41, taking the optical property data in the basic data set as input and taking the composition and thickness of the passive radiative cooling film as output to construct an inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film; S42, training the inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film constructed in S41 using the training set, and testing the accuracy of the trained inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film using the test set; S43, selecting mean square error as a loss function and selecting Adam optimizer to adaptively adjust a learning rate; S44, selecting one of forward propagation, loss function calculation, back propagation and parameter update as a training method, and training the inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film to an accuracy of not less than 0.9; S50, inputting target optical property data into the trained inverse design model of the optical property data to the composition and thickness of the passive radiative cooling film in S40 to obtain initial design values of the composition and thickness of the passive radiative cooling film; S60, performing optical simulation on the initial design values of the composition and thickness of the passive radiative cooling film obtained in S50 to predict the emissivity of the passive radiative cooling film in the atmospheric window wave band and the reflectivity of the passive radiative cooling film in the solar wave band; The optical simulation uses a transfer matrix method to simulate the optical response of the passive radiative cooling film to calculate the optical property data of the passive radiative cooling film substrate at different wavelengths; S70, calculating the mean square error between the emissivity in the atmospheric window wave band and the reflectivity in the solar wave band obtained in S60 and the target optical property data in S50. S80 judges whether the mean square error is less than 0.1, if the mean square error is less than 0.1, the design of the passive radiative cooling film is completed, if the mean square error is greater than or equal to 0.1, steps S30-S80 are repeated.

2. The machine learning-based inverse design method of a passive radiative cooling film according to claim 1, wherein, The passive radiative cooling film substrate is one or more of Al, Al2O3, Au, BaF2, Be, CaF2, Cr, Cu, Fe, Mn, KBr, SiO2, Si3N4, SiC, Ti, TiO2, ZnO, ZnS, PDMS, PE, PVA, PVDF.

3. The machine learning based inverse design method of a passive radiative cooling film according to claim 1, wherein, The optical property data includes one or more of the refractive index, reflectivity, emissivity and extinction coefficient of the passive radiative cooling film substrate.

4. The machine learning based inverse design method of a passive radiative cooling film according to claim 1, wherein, The basic data set includes the composition and thickness of the passive radiative cooling film and the optical property data of the passive radiative cooling film substrate.

5. The machine learning based inverse design method of a passive radiative cooling film according to claim 1, wherein, The preprocessing method includes: S21 processes missing values or abnormal values in the data to ensure data quality; S22 adopts a Robust standardization method to standardize the variables and compress the data to the range of [-1, 1].

6. The machine learning based inverse design method of a passive radiative cooling film according to claim 1, wherein, K in the K-fold cross-validation method is 10.

7. The machine learning based inverse design method of a passive radiative cooling film according to claim 1, wherein, The target optical property data is the emissivity of the passive radiative cooling film in the atmospheric window waveband and the reflectivity in the solar waveband.

Images