A method and application for intelligent screening of radiation-cooled ceramic formulations based on Transformer model and genetic algorithm.

Abstract

The present invention is entitled "A Method and Application for Intelligent Screening of Radiation-Cooled Ceramic Formulas Based on Transformer Model and Genetic Algorithm", belonging to the field of machine learning technology. The technical problems to be solved are to reduce the complexity of system implementation, improve the overall operational stability, avoid generating invalid formulas, and reduce experimental costs and R&D cycle. The key points of the technical solution are: (1) Parametrically describe the radiation-cooled ceramic formulas and establish a physical and chemical constraint model of oxide composition ratio to construct a feasible formula solution space that meets the constraint conditions; (2) Construct an optical performance evaluation model; (3) In the feasible formula solution space constructed in step (1), generate an initial candidate ceramic formula population that meets the constraint conditions by random initialization, which serves as the initial individuals of the genetic algorithm; (4) Define and evaluate the fitness function; (5) Genetic evolution and formula screening output.

Description

Technical Field

[0001] This invention provides a method and application for intelligent screening of radiation-cooled ceramic formulations based on the Transformer model and genetic algorithm, belonging to the field of machine learning technology. Background Technology

[0002] Radiation-cooled ceramics are a class of functional ceramic materials that achieve high reflectivity in the solar radiation band and high emissivity in the mid-to-far infrared atmospheric window band (8-13 μm), thereby reducing the surface temperature of the material without external energy input. These materials have significant application value in fields such as building energy conservation, passive refrigeration, and spacecraft thermal control.

[0003] Radiation-cooled ceramics are typically prepared by mixing various inorganic oxide raw materials in a certain proportion and then sintering them at high temperatures. Their optical properties are highly dependent on the types and proportions of each oxide component. For multi-component oxide systems, there are complex nonlinear coupling relationships between different components, and even small changes in the proportions can lead to significant differences in the optical properties of the material.

[0004] Existing methods for designing radiation-cooled ceramic formulations mainly include empirical and experimental manual design methods and computationally assisted methods based on forward prediction. Empirical and experimental manual design methods rely on researchers' literature experience or prior knowledge to select several oxides and repeatedly adjust their proportions, screening suitable formulations through extensive experimental testing. This method has a long development cycle, high experimental costs, and difficulty in systematically exploring large-scale formulation spaces. Some studies have introduced machine learning models to predict and model the "known ceramic formulation → corresponding optical properties," while computationally assisted methods based on forward prediction are used to quickly evaluate the performance of existing formulations. However, these methods can only evaluate given formulations and cannot actively generate new formulation schemes under target performance constraints.

[0005] In practical engineering applications, materials designers are more concerned with the technical problem of how to automatically screen and determine the required chemical composition ratios within a continuous, multi-component ceramic formulation space under the condition of pre-setting target optical performance indicators.

[0006] Relevant patent documents retrieved: This document, published in China (CN118335217A) on July 12, 2024, discloses a method for obtaining the proportion of radiation shielding materials based on a genetic algorithm. The method includes the following steps: First, an initial population is randomly generated based on the composition range of the radiation shielding material. The performance parameters of the initial population are calculated, and the objective function of the genetic algorithm is established. Next, non-dominated sorting is performed, and the genetic algorithm is used to perform crossover and mutation on the radiation shielding material. The process is iterated according to constraints until a given number of iterations N is reached, resulting in an optimized proportion of the radiation shielding material. However, this invention does not specifically address radiation-cooled ceramics.

[0007] Relevant non-patent literature retrieved: The journal or book title is *Automation Applications*, and the document title is "Application of Improved Genetic Algorithm in Ceramic Formulation Design," Volume 66, Issue 10, published in May 2025. This document discloses a study on ceramic formulation design using a standard genetic algorithm with relative error as the objective function. To improve algorithm performance, a penalty function is introduced to handle constraints, and the crossover and mutation operations are improved to construct an optimized genetic algorithm. In ceramic formulation experiments, the improved genetic algorithm is compared with the standard genetic algorithm. However, this method does not address the issue of avoiding the generation of unpreparable and invalid formulations.

[0008] The prior art represented by the aforementioned documents has at least the following unresolved technical problems or defects: Existing technologies have many shortcomings in addressing the above problems. First, the formulation space of multi-component oxide systems has high dimensionality and strong continuity, making manual search or simple enumeration methods difficult to apply. Second, forward prediction models cannot directly solve the problem of deriving material formulations from target performance. Third, there is a lack of an intelligent method that can automatically complete formulation search and optimization under performance targets and formulation constraints.

[0009] Therefore, it is necessary to propose a new technical solution to achieve automated and intelligent formulation screening of radiation-cooled ceramics under the constraint of target optical performance. Summary of the Invention

[0010] The purpose of this invention is to provide: A method and application for intelligent screening of radiation-cooled ceramic formulations based on the Transformer model and genetic algorithm, and related technologies, are proposed to address technical issues such as reducing system implementation complexity, improving overall operational stability, avoiding the generation of invalid formulations, and reducing experimental costs and R&D cycles, or combinations thereof.

[0011] Terminology Explanation: Unless otherwise defined, all technical terms in this document have the same meanings as commonly understood by one of ordinary skill in the art to which the subject matter of the claims pertains. Unless otherwise stated, all patents, patent inventions, and publications cited in this document are incorporated herein by reference in their entirety. If multiple definitions exist for terms in this document, the definitions in this chapter shall prevail.

[0012] It should be understood that the above brief description and the following detailed description are exemplary and for illustrative purposes only, and do not limit the subject matter of the invention in any way. In this invention, the singular is used in conjunction with the plural unless otherwise specifically stated. It should also be noted that, unless otherwise stated, the use of “or” or “or” means “and / or”. Furthermore, the use of the term “comprising” and other forms such as “including,” “containing,” and “contains” are not limiting.

[0013] The definition of the standard chemical term can be found in the reference "Radiative Cooling: Novel Micro-Nano Radiative Thermal Management Technology", Qiu Jun, Harbin Institute of Technology Press.

[0014] Unless specifically defined herein, the use of all commercially available products herein employs standard techniques. For example, it may be carried out using the manufacturer's instructions for use with the kit, or in accordance with methods known in the art or the description of this invention. The techniques and methods described herein can generally be implemented according to conventional methods well known in the art, based on the descriptions in the various summary and more specific documents cited and discussed in this specification.

[0015] The terms “optional / arbitrary” or “optionally / arbitrarily” mean that the event or situation described below may or may not occur, including both the occurrence and non-occurrence of the event or situation.

[0016] In a first aspect, the present invention provides: a method for intelligent screening of radiation-cooled ceramic formulations based on the Transformer model and genetic algorithm, comprising the following steps: (1) The radiation-cooled ceramic formulation is parameterized and a physical and chemical constraint model of the oxide composition ratio is established to construct a feasible formulation solution space that satisfies the constraint conditions. (2) Construction of optical performance evaluation model; (3) Within the feasible solution space constructed in step (1), an initial candidate ceramic formula population that satisfies the constraints is generated using a random initialization method, which serves as the initial individuals for the genetic algorithm; (4) Definition and evaluation of fitness function; (5) Genetic evolution and formula screening output.

[0017] The parameterization described in step (1) is as follows: the chemical composition of the radiation-cooled ceramic is represented as a vector of percentage content of various oxides.

[0018] The percentage content vector of the components composed of the oxides is a continuous variable.

[0019] Among them, the constraint modeling in step (1) is: constructing a feasible solution space of the oxide component percentage content vector, so that it at least satisfies the constraint that the percentage content of each oxide component is not less than zero, and the sum of the percentage content of all oxide components is 100%. The percentage content vector of the oxide components is a continuous variable, meaning that the percentage content of each oxide component can take any real value within its range, and the sum of these values ​​is 100% through normalization constraints.

[0020] The oxide is one or more of the metal oxides or non-metal oxides constituting the radiation-cooled ceramic, including but not limited to Al2O3, SiO2, TiO2, MnO, K2O, Na2O, CaO, MgO, Fe2O3, ZnO, BaO, SrO, Li2O, P2O5, and CuO.

[0021] In the initialization stage of step (3), an initial candidate ceramic formula that satisfies the component percentage content constraint is generated by randomly sampling within the feasible formula solution space.

[0022] In the genetic evolution process of step (5), the candidate individuals generated by crossover and mutation are subject to constraint correction processing so that they always meet the constraint conditions that the percentage content of each oxide component is not less than zero and the total is 100%.

[0023] The method described in step (2) is as follows: a neural network model based on the Transformer structure is used as input, which is the percentage content vector of the oxide components, and the output is the corresponding optical performance parameters.

[0024] The optical performance parameters include at least one of solar reflectivity and infrared emissivity.

[0025] The neural network model of the Transformer structure is trained offline using existing experimental sample data to learn the nonlinear mapping relationship between the chemical composition and optical properties of ceramics.

[0026] The neural network model with the Transformer structure performs forward inference computation to predict and evaluate the optical performance of candidate formulations, without participating in the evolutionary strategy formulation process of the genetic algorithm.

[0027] The random initialization method in step (3) is as follows: under the constraints, a genetic algorithm is used to model the initial candidate formula, and a continuous numerical encoding method is used to represent the ceramic formula, so as to generate an initial candidate formula population that meets the constraints.

[0028] The genetic algorithm uses a continuous numerical encoding method for gene coding, with each gene coding corresponding to the percentage content of an initial candidate ceramic oxide component.

[0029] The preset target optical performance indicators are target values ​​pre-set according to the specific application scenario requirements or design indicators, including target solar reflectivity and / or target infrared emissivity.

[0030] The optical performance indicators are obtained through experimental measurement or statistical analysis based on existing experimental data.

[0031] The method for constructing the fitness function in step (4) is: based on the error between the predicted optical performance parameters and the preset target optical performance index; the error includes at least one of the following: squared error, absolute error, and distance metric.

[0032] The fitness function can be in the form of a single objective or a multi-objective, and different weights can be assigned to performance indicators such as solar reflectivity and infrared emissivity according to different application requirements, so as to guide the genetic algorithm to converge toward a ceramic formula that meets the needs of a specific application scenario.

[0033] The evaluation in step (4) is as follows: calculate the fitness value of each candidate ceramic formula, sort the candidate formulas according to the fitness function value, and select the top N candidate formulas with the smallest fitness function value to enter the next generation of genetic evolution process, where N is the preset retention quantity; the fitness function value is the weighted sum of the errors between the predicted optical performance parameters and the preset target optical performance index, thereby guiding the formula vector to meet the direction optimization of the target optical performance index.

[0034] The weights of the weighted sum are derived from the different weights assigned to performance indicators such as solar reflectivity and infrared emissivity according to different application requirements.

[0035] Wherein, the genetic evolution in step (5) is: based on the fitness function, through genetic operations, the initial candidate formula population obtained in step (3) is subjected to multiple generations of evolutionary iteration.

[0036] The genetic operations include at least one of selection, crossover, and mutation.

[0037] In each generation of the multi-generational evolution, candidate ceramic formulations are evaluated and screened based on the fitness function, so that the formulations gradually converge in the continuous formulation space toward meeting the target optical performance indicators.

[0038] The formula screening output in step (5) is as follows: when the genetic algorithm meets the preset termination condition, it outputs one or more sets of radiation-cooled ceramic formula schemes that meet the target optical performance requirements. The formula schemes are given in the form of oxide component percentage content and can be further converted into corresponding raw material ratio schemes for experimental verification or engineering applications.

[0039] The preset termination condition is: reaching a preset number of iterations or the fitness function converging to a set threshold.

[0040] Wherein, the number of iterations is any integer within a preset range of 50-1000; The set threshold is any preset convergence criterion in which the rate of change of the fitness function is less than 0.1%-5% over several consecutive generations.

[0041] Without departing from the core idea of ​​this invention, the following alternative designs can be made: (1) Replace the genetic algorithm (GA) with other intelligent optimization algorithms; For example, the genetic algorithm can also be replaced by particle swarm optimization algorithm, simulated annealing algorithm, differential evolution algorithm, ant colony algorithm, artificial bee colony algorithm, or other swarm intelligence optimization algorithm.

[0042] (2) Replace the performance evaluation model with other machine learning or deep learning models; For example, the performance evaluation model can also be replaced by a multilayer perceptron model, a support vector machine model, a random forest model, a gradient boosting tree model, a convolutional neural network model, a recurrent neural network model, or a long short-term memory network model.

[0043] (3) Expand the input formula from the percentage content of oxide components to the raw material-level formula parameters.

[0044] All of the above modifications should be considered to fall within the protection scope of this invention.

[0045] Secondly, the present invention provides the application of the above method in the preparation of radiation-cooled ceramic materials.

[0046] The present invention has at least the following beneficial effects: Compared with existing technologies, this invention has better technical effects in reducing system implementation complexity, improving overall operational stability, avoiding the generation of invalid formulas, and reducing experimental costs and R&D cycles.

[0047] (1) By separating the optical performance prediction model from the formula screening process, the present invention makes the radiation-cooled ceramic formula screening process independent of the model’s reverse training or gradient information, thereby reducing the complexity of system implementation and improving the overall operational stability.

[0048] (2) By adopting a continuous encoding and physical constraint correction mechanism, the genetic algorithm always satisfies the physical and chemical constraints of the percentage content of oxide components during the search process in the continuous ceramic formula space, thus avoiding the generation of invalid formulas that cannot be prepared.

[0049] (3) By using a fitness evaluation strategy based on predicted optical performance, efficient screening of radiation-cooled ceramic formulations can be achieved without the need for extensive physical experiments, which significantly reduces experimental costs and R&D cycle.

[0050] (4) By outputting multiple sets of candidate ceramic formulation schemes, it provides flexible space for subsequent process verification and performance trade-offs, and improves the feasibility and applicability of this method in practical engineering applications.

[0051] Furthermore, based on the present invention: Furthermore, this invention combines an optical performance evaluation model with a genetic algorithm optimization process, achieving efficient search and screening within a continuous formulation space. Compared to methods that rely solely on performance prediction models or optimization algorithms, this invention can quickly obtain ceramic formulations that meet target optical performance requirements while satisfying constraints, thus offering significant advantages in reducing search complexity, improving screening efficiency, and avoiding the generation of invalid formulations.

[0052] In some implementations, the technical effectiveness of the present invention can be verified by comparing methods that use only genetic algorithms for optimization or only performance prediction models for recipe screening. The comparison results show that the present invention outperforms other methods in terms of search efficiency, recipe feasibility, and target performance matching. Detailed Implementation

[0053] The following non-limiting embodiments are intended to enable those skilled in the art to gain a more comprehensive understanding of the present invention, but do not limit the invention in any way. The following content is merely an exemplary description of the scope of protection claimed by the present invention, and those skilled in the art can make various changes and modifications to the present invention based on the disclosed content, and such changes should also fall within the scope of protection claimed by the present invention.

[0054] The present invention will be further described below by way of specific embodiments. Unless otherwise specified, all instruments, devices, equipment, reagents, products, etc., used in the embodiments of the present invention are obtained through conventional commercial means.

[0055] Example 1: A method for intelligent screening of radiation-cooled ceramic formulations based on Transformer model and genetic algorithm (1) The radiation-cooled ceramic formulation is parameterized and a physical and chemical constraint model of the oxide composition ratio is established to construct a feasible formulation solution space that satisfies the constraint conditions. (2) Construction of optical performance evaluation model; (3) Within the feasible solution space constructed in step (1), an initial candidate ceramic formula population that satisfies the constraints is generated using a random initialization method, which serves as the initial individuals for the genetic algorithm; (4) Definition and evaluation of fitness function; (5) Genetic evolution and formula screening output.

[0056] In some embodiments, the parameterization described in step (1) is as follows: the chemical composition of the radiation-cooled ceramic is represented as a vector of percentage contents of various oxides.

[0057] In some embodiments, the percentage content vector of the components composed of the oxides is a continuous variable.

[0058] In some embodiments, the constraint modeling in step (1) is: constructing a feasible solution space for the oxide component percentage content vector, such that it at least satisfies the constraint that the percentage content of each oxide component is not less than zero, and the sum of the percentage contents of all oxide components is 100%. In some embodiments, the percentage content vector of the oxide components is a continuous variable, that is, the percentage content of each oxide component can take any real value within its range, and the sum of them is made up to 100% by normalization constraint.

[0059] In some embodiments, the oxide is one or more of the metal oxides or non-metal oxides constituting the radiation-cooled ceramic, including but not limited to Al2O3, SiO2, TiO2, MnO, K2O, Na2O, CaO, MgO, Fe2O3, ZnO, BaO, SrO, Li2O, P2O5, and CuO.

[0060] In some embodiments, during the initialization phase of step (3), an initial candidate ceramic formula that satisfies the component percentage content constraint is generated by randomly sampling within the feasible formula solution space.

[0061] In some embodiments, during the genetic evolution process of step (5), the candidate individuals generated by crossover and mutation are subject to constraint correction processing so that they always meet the constraint conditions that the percentage content of each oxide component is not less than zero and the total is 100%.

[0062] In some embodiments, the construction method in step (2) is as follows: a neural network model based on the Transformer structure is input as a vector of the percentage content of oxide components, and the output is the corresponding optical performance parameters.

[0063] In some embodiments, the optical performance parameters include at least one of solar reflectivity and infrared emissivity.

[0064] In some embodiments, the neural network model of the Transformer structure is trained offline using existing experimental sample data to learn the nonlinear mapping relationship between the chemical composition and optical properties of ceramics.

[0065] In some embodiments, the neural network model of the Transformer structure performs forward inference computation to predict and evaluate the optical performance of candidate formulations, without participating in the evolutionary strategy formulation process of the genetic algorithm.

[0066] In some embodiments, the random initialization method in step (3) is as follows: a genetic algorithm is used to search and optimize the initial candidate formulas, and the percentage content of oxide components in each optimized initial candidate formula is regarded as a separate individual in the genetic algorithm.

[0067] In some embodiments, the genetic algorithm uses a continuous numerical encoding method for gene coding, with each gene coding corresponding to an initial candidate ceramic oxide component percentage content.

[0068] In some embodiments, the fitness function in step (4) is defined as follows: the candidate ceramic formula obtained in step (3) is input into the optical performance evaluation model constructed in step (2) to obtain the corresponding predicted optical performance parameters, and the fitness function is constructed based on the difference between the predicted optical performance parameters and the preset target optical performance index.

[0069] In some embodiments, the preset target optical performance index is a target value pre-set according to the specific application scenario requirements or design specifications, including target solar reflectivity and / or target infrared emissivity.

[0070] In some embodiments, the optical performance indicators are obtained through experimental measurement or statistical analysis based on existing experimental data.

[0071] In some embodiments, the method for constructing the fitness function is: based on the error between the predicted optical performance parameters and the preset target optical performance index; the error includes at least one of the following: squared error, absolute error, and distance metric.

[0072] In some embodiments, the fitness function may take the form of a single objective or a multi-objective, and different weights may be assigned to performance indicators such as solar reflectivity and infrared emissivity according to different application requirements, so as to guide the genetic algorithm to converge toward a ceramic formula that meets the needs of a specific application scenario.

[0073] In some embodiments, the evaluation in step (4) is as follows: calculate the fitness value of each candidate ceramic formulation, sort the candidate formulations according to the fitness function value, and select the top N candidate formulations with the smallest fitness function value to enter the next generation of genetic evolution process, where N is the preset retention quantity; the fitness function value is the weighted sum of the errors between the predicted optical performance parameters and the preset target optical performance index, thereby guiding the formulation vector to meet the direction optimization of the target optical performance index.

[0074] In some embodiments, the weights of the weighted sum are derived from the different weights assigned to performance indicators such as solar reflectivity and infrared emissivity according to different application requirements.

[0075] In some embodiments, the genetic evolution in step (5) is: based on the fitness function, the initial candidate formula population obtained in step (3) is subjected to multiple generations of evolutionary iterations through genetic operations.

[0076] In some embodiments, the genetic operation includes at least one of selection, crossover, and mutation.

[0077] In some embodiments, during each generation of the multi-generational evolution, candidate ceramic formulations are evaluated and screened based on a fitness function, so that the formulations gradually converge in the continuous formulation space toward the direction of satisfying the target optical performance indicators.

[0078] In some embodiments, the formula screening output in step (5) is: when the genetic algorithm meets the preset termination condition, it outputs one or more sets of radiation-cooled ceramic formula schemes that meet the target optical performance requirements. The formula schemes are given in the form of oxide component percentage content and can be further converted into corresponding raw material ratio schemes for experimental verification or engineering applications.

[0079] In some embodiments, the preset termination condition is: reaching a preset number of iterations or the fitness function converging to a set threshold.

[0080] In some embodiments, the number of iterations is any integer within a preset range of 50-1000; In some embodiments, the set threshold is any preset convergence criterion in which the rate of change of the fitness function is less than 0.1%-5% over several consecutive generations.

[0081] Example 2 Based on existing radiation-cooled ceramic sample data, a dataset containing the percentage content of oxide components and their corresponding optical performance parameters was constructed. The Transformer neural network model was then trained to obtain a stable optical performance evaluation model.

[0082] Set target optical performance indicators and initialize the ceramic formulation population. Input each generation of candidate formulations into the performance evaluation model to calculate fitness. Through multiple generations of genetic evolution, select ceramic formulations that meet the target performance requirements.

[0083] Based on the Transformer model established in Example 1, the target optical performance is input, data is processed, and then a GA (Genetic Algorithm) model is used to generate candidate formulations. The Transformer model is then used again for evaluation, followed by a series of evolutionary operations to optimize the formulation and ultimately select the final formulation. First, the ceramic formulations to be screened are represented in parameterized form and input into the genetic algorithm optimization module. The parameterized formulation consists of the mass or molar percentage of multiple oxide components, and non-negativity constraints and total quantity constraints are introduced during the generation process to ensure that the candidate formulations meet physical feasibility requirements. The genetic algorithm module performs selection, crossover, and mutation operations on the candidate formulations to generate a new candidate formulation population. Each candidate formulation undergoes optical performance prediction via a performance evaluation model, which uses a deep learning-based Transformer network structure and is only used for numerical prediction of formulation performance, without participating in the formulation generation and decision-making process. Based on the predicted optical performance results and combined with the preset target optical performance, the corresponding fitness evaluation index is calculated, and this evaluation result is fed back to the genetic algorithm module to guide the evolutionary optimization of the next round of candidate formulations. When the preset number of iterations or fitness convergence conditions are met, the ceramic formula parameters that best match the target optical performance are output, realizing intelligent automatic screening of ceramic formulas in a continuous parameter space.

[0084] Candidate ceramic formulations generated by a genetic algorithm are first fed into an optical performance evaluation model. The performance evaluation module employs a deep learning model based on a Transformer architecture. This model establishes a mapping relationship between ceramic formulation parameters and optical performance through training on historical sample data. The performance evaluation model performs forward inference on the input candidate formulations and outputs corresponding optical performance prediction results. Subsequently, based on the difference between the predicted optical performance and the preset target optical performance, a fitness evaluation function is constructed to quantify the merits of the candidate formulations. The fitness evaluation function can adopt squared error, absolute error, or distance metric forms to adapt to the performance optimization needs of different application scenarios. It should be noted that the performance evaluation model is only used for candidate formulation performance prediction and does not directly participate in formulation generation or evolutionary decision-making, thus achieving functional decoupling between the prediction model and the optimization algorithm.

[0085] The genetic algorithm module is responsible for generating and optimizing candidate ceramic formulations within a continuous formulation space that satisfies physical constraints; the performance evaluation module is used to predict the optical performance of candidate formulations and feed the prediction results back to the fitness evaluation module; the fitness evaluation results serve as feedback information for the evolutionary process of the genetic algorithm, guiding the selection and evolution of candidate formulations, thereby achieving intelligent screening of ceramic formulations based on target optical performance.

[0086] Example 3 In this embodiment, the percentage content of oxide components and the corresponding solar reflectance and infrared emissivity of multiple known ceramic samples are first collected. After the data is normalized, it is used to train an optical performance evaluation model based on the Transformer structure offline.

[0087] Subsequently, in the formulation screening stage, a genetic algorithm is used to generate an initial population of ceramic formulations that meet the component constraints. Each candidate formulation is then input into a trained Transformer model to obtain the corresponding optical performance prediction results. The fitness value is calculated based on the error between the prediction results and the target optical performance indicators, and the selection, crossover, and mutation operations of the genetic algorithm are performed accordingly.

[0088] After multiple generations of evolution and iteration, several sets of ceramic formulations with optimal adaptability are output as screening results. The screening results are given in the form of oxide component percentage content, the sum of the percentage content of all components is 100%, and they meet the preset optical performance target requirements.

[0089] This invention enables automated screening of radiation-cooled ceramic formulations under given target optical performance conditions; significantly reduces reliance on human experience and trial-and-error in materials research and development; improves the efficiency and success rate of multi-component ceramic material formulation design; and supports multi-objective formulation optimization under complex constraints.

[0090] It should be noted that in this embodiment, the Transformer model is only used to evaluate the optical performance of candidate formulations. Its training method, network structure improvement and prediction accuracy enhancement are not the focus of this invention. The core of this invention is to use a genetic algorithm combined with a performance evaluation model to achieve intelligent screening of radiation-cooled ceramic formulations.

[0091] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A method for intelligent screening of radiation-cooled ceramic formulations based on the Transformer model and genetic algorithm, characterized in that, Includes the following steps: (1) The radiation-cooled ceramic formulation is parameterized and a physical and chemical constraint model of the oxide composition ratio is established to construct a feasible formulation solution space that satisfies the constraint conditions. (2) Construction of optical performance evaluation model; (3) Within the feasible solution space constructed in step (1), an initial candidate ceramic formula population that satisfies the constraints is generated using a random initialization method, which serves as the initial individuals for the genetic algorithm; (4) Definition and evaluation of fitness function; (5) Genetic evolution and formula screening output.

2. The method according to claim 1, characterized in that, The parameterization described in step (1) is as follows: the chemical composition of the radiation-cooled ceramic is represented as a vector of percentage content of various oxides; The percentage content vector of the components composed of oxides is a continuous variable; The constraint modeling in step (1) is as follows: construct a feasible solution space for the percentage content vector of oxide components, such that it at least satisfies the constraint that the percentage content of each oxide component is not less than zero, and the sum of the percentage contents of all oxide components is 100%.

3. The method according to claim 1, characterized in that, The method described in step (2) is as follows: a neural network model based on the Transformer structure is used as input, which is the percentage content vector of the oxide components, and the output is the corresponding optical performance parameters.

4. The method according to claim 3, characterized in that, The optical performance parameters include at least one of solar reflectivity and infrared emissivity; The neural network model with the Transformer structure is trained offline using existing experimental sample data to learn the nonlinear mapping relationship between the chemical composition and optical properties of ceramics.

5. The method according to claim 1, characterized in that, The random initialization method in step (3) is as follows: under the constraints, a genetic algorithm is used to model the initial candidate formula, and a continuous numerical encoding method is used to represent the ceramic formula, so as to generate an initial candidate formula population that meets the constraints.

6. The method according to claim 5, characterized in that, The genetic algorithm uses a continuous numerical encoding method for gene coding, with each gene code corresponding to the percentage content of an initial candidate ceramic oxide component.

7. The method according to claim 1, characterized in that, The fitness function in step (4) is defined as follows: the candidate ceramic formula obtained in step (3) is input into the optical performance evaluation model constructed in step (2) to obtain the corresponding predicted optical performance parameters. The fitness function is constructed based on the difference between the predicted optical performance parameters and the preset target optical performance index.

8. The method according to claim 7, characterized in that, The preset target optical performance indicators are target values ​​pre-set according to the specific application scenario requirements or design indicators, including target solar reflectivity and / or target infrared emissivity; The optical performance indicators are obtained through experimental measurement or statistical analysis based on existing experimental data; The method for constructing the fitness function is as follows: it is constructed based on the error between the predicted optical performance parameters and the preset target optical performance index. The error includes at least one of the following: squared error, absolute error, and distance metric form; The evaluation in step (4) is as follows: calculate the fitness value of each candidate ceramic formulation, sort the candidate formulations according to the fitness function value, and select the top N candidate formulations with the smallest fitness function value to enter the next generation of genetic evolution process, where N is the preset retention quantity; the fitness function value is the weighted sum of the errors between the predicted optical performance parameters and the preset target optical performance index.

9. The method according to claim 1, characterized in that, The genetic evolution in step (5) is as follows: based on the fitness function, the initial candidate formula population obtained in step (3) is subjected to multiple generations of evolutionary iteration through genetic operations; The genetic operations include at least one of selection, crossover, and mutation; The formula screening output in step (5) is as follows: when the genetic algorithm meets the preset termination condition, it outputs one or more sets of radiation-cooled ceramic formula schemes that meet the target optical performance requirements. The formula schemes are given in the form of oxide component percentage content and can be further converted into corresponding raw material ratio schemes. The preset termination condition is: reaching a preset number of iterations or the fitness function converging to a set threshold. The number of iterations is any integer within a preset range of 50-1000. The set threshold is any preset convergence criterion in which the rate of change of the fitness function is less than 0.1%-5% over several consecutive generations.

10. The application of the method according to any one of claims 1-9 in the preparation of radiation-cooled ceramic materials.