Intelligent design and preparation method of AI-driven inorganic hydrated salt phase change material

Abstract

The invention relates to an AI-driven intelligent design and preparation method of an inorganic hydrated salt phase change material, and solves the problem that the traditional technology is mainly based on experience trial and error and single performance optimization and cannot give consideration to multi-performance balance and multi-scene efficient adaptation development requirements of the inorganic hydrated salt phase change material. The method comprises the following steps: acquiring multi-dimensional performance requirements (including phase change temperature, latent heat value and the like) of a material, generating a candidate formula and a prediction result by using a trained Gaussian process regression model, and performing multi-objective optimization to screen out a Pareto optimal formula; and carrying out experimental verification and calculating deviation, retraining the model by complementary data exceeding a threshold value, and determining a final formula after reaching the standard so as to be matched with continuous process large-scale preparation. The method has the advantages that the AI replaces experience trial and error, multi-performance cooperation of materials is achieved, the research and development period is greatly shortened, the cost is reduced, and the method is suitable for multiple energy storage scenes.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of energy saving, and in particular to an AI-driven intelligent design and preparation method of inorganic hydrated salt phase change materials. BACKGROUND

[0002] As a kind of functional material that can efficiently absorb, store and release latent heat during phase transition, phase change materials (PCMs) are the key supporting materials for realizing energy cascade utilization, promoting building energy saving, industrial waste heat recovery, cold chain transportation and other fields of energy saving and emission reduction. In engineering applications, ideal PCMs need to meet multiple performance requirements: the phase change temperature needs to be accurately matched with the specific scene heat source characteristics, the phase change enthalpy needs to ensure high heat storage density, the supercooling degree needs to be controlled below 2 DEG C to ensure the accuracy of phase change triggering, the cycle life needs to exceed 1000 times and the performance decay needs to be less than 10%, and the rheological performance needs to adapt to different construction processes. Among them, inorganic hydrated salt PCMs (such as sodium sulfate decahydrate and calcium chloride hexahydrate) have high volumetric heat storage density (usually > 300 MJ / m³), low cost (only 1 / 3-1 / 5 of organic PCMs), good thermal conductivity and non-flammability, and become the most potential PCM system for large-scale engineering applications.

[0003] Currently, the development of inorganic hydrated salt PCMs mainly relies on the traditional technical mode of "experience trial and error + single performance optimization". To meet different performance requirements, the research and development process often adopts the idea of "breaking through one by one": if the phase change temperature needs to be adjusted, different salts are simply mixed for ratio adjustment; if the liquid flowability after phase change needs to be improved, porous carriers such as expanded perlite are added for adsorption and packaging, or polyacrylic acid sodium and other polymer thickeners are added; if the supercooling degree needs to be reduced, borax and other nucleating agents are selected based on experience; if the cyclic phase separation needs to be relieved, further supplement of gelling agents is needed. This mode takes "solving single performance problem" as the core, lacks systematic planning and collaborative design of comprehensive performance of materials, and all adjustments are based on the experience accumulation of research and development personnel without clear theoretical model or data support.

[0004] However, this traditional technology mode has two core limitations, which are difficult to adapt to the comprehensive performance requirements of PCMs in engineering: on the one hand, the multi-performance collaborative optimization capability is missing. Since existing technologies only design and control means for a single performance, optimizing a certain indicator often leads to degradation of other key performances. For example, when a large amount of salt is mixed to reduce the phase change temperature to adapt to the building heating scene, the phase change enthalpy value of the material will be significantly sacrificed. When a high-molecular thickening agent is added to improve the rheological property, the heat storage performance of the material will be weakened, and ultimately the balance of temperature, latent heat, supercooling degree, and durability cannot be achieved. On the other hand, the research and development efficiency and scene adaptability are insufficient. Due to the reliance on empirical trial and error, there is a lack of systematic design methods, and the development of materials for each scene requires repeated debugging for months or even years. Not only is the research and development cost high, but it is also difficult to quickly respond to differentiated performance requirements of different scenes (such as air conditioning cold storage 5-10℃, heat pump heat storage 45-55℃), resulting in a gap between the developed materials and the actual needs of engineering. SUMMARY

[0005] In order to replace empirical trial and error with AI, achieve multi-performance collaboration of materials, significantly shorten the research and development cycle, reduce costs, and adapt to multiple energy storage scenes, the present application provides an AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials.

[0006] The present application provides an AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials, which adopts the following technical solution:

[0007] An AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials, comprising:

[0008] Obtaining multi-dimensional performance requirement indicators of inorganic hydrated salt phase change materials for target application scenarios, wherein the multi-dimensional performance requirement indicators include phase change temperature range, minimum phase change latent heat value, maximum allowable supercooling degree, viscosity range suitable for construction process, target cost upper limit, and minimum cycle life;

[0009] Using the trained Gaussian process regression model to analyze the performance requirement indicators, generating candidate formulations and outputting the performance prediction results of each candidate formulation;

[0010] Based on the candidate formulations and performance prediction results, and with the predicted values of each performance indicator meeting the corresponding preset requirement threshold and the predicted value of the formulation cost being ≤ the target cost upper limit as constraints, a multi-objective optimization algorithm is used to select a Pareto optimal formulation;

[0011] Preparing samples and measuring the performance of the Pareto optimal formulation, calculating the deviation values of the measured data and the performance prediction results on each performance indicator, and comparing the deviation values with the preset threshold;

[0012] If the deviation values of all performance indicators do not exceed the preset threshold value, the formula is taken as the target formula;

[0013] If the deviation value of any performance indicator exceeds the preset threshold value, the measured data is supplemented to the training data set, the Gaussian process regression model is retrained, the Pareto optimal formula is selected again, and the sample preparation, performance measurement and deviation comparison steps are repeated until one or more target formulas are obtained, all of which meet the preset threshold value requirements;

[0014] Based on the preset comprehensive performance evaluation function, the formula with the highest evaluation score is selected from all obtained target formulas as the final formula, and the continuous process matched with the final formula is used to complete the large-scale preparation.

[0015] By using the above technical scheme, the Gaussian process regression modeling and the AI scheme of multi-objective optimization are used to replace the traditional experience trial and error, and the coordination of multiple performances (temperature, latent heat, etc.) of materials is accurately realized; the closed-loop feedback improves the prediction accuracy, greatly shortens the research and development cycle, reduces the cost, adapts to multiple scenes, realizes large-scale preparation combined with the continuous process, and breaks through the traditional bottleneck. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of an AI-driven inorganic hydrated salt phase change material intelligent design and preparation method according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be further described in detail below with reference to the accompanying drawings.

[0018] Reference Figure 1 An AI-driven inorganic hydrated salt phase change material intelligent design and preparation method disclosed in the present application comprises:

[0019] In step S100, the multi-dimensional performance requirement indicators of the inorganic hydrated salt phase change material for the target application scene are obtained, wherein the multi-dimensional performance requirement indicators include the phase change temperature range, the minimum phase change latent heat value, the maximum allowable supercooling degree, the viscosity range suitable for the construction process, the target cost upper limit and the minimum cycle life.

[0020] The phase change temperature range refers to the temperature interval at which the material undergoes solid-liquid or liquid-solid transition during the phase change process, which is a key parameter for the application of phase change materials and determines whether the material can accurately match the heat source characteristics of a specific application scenario. The minimum phase change latent heat value refers to the minimum heat absorbed or released by the material during the phase change process, usually measured in J / g. A high latent heat value means that the material has a higher energy storage density. The maximum allowable supercooling degree refers to the maximum temperature difference between the actual crystallization temperature and the theoretical phase change temperature of the material during the cooling process. The smaller the supercooling degree, the more accurate the phase change triggering temperature of the material. The viscosity range suitable for construction process refers to the viscosity interval of the material in the molten state suitable for a specific construction process (such as pouring, spraying, injection molding, etc.), usually measured in mPa·s. Moderate viscosity can ensure good flowability and stability of the material during the construction process. The upper limit of the target cost refers to the maximum cost allowed in the material preparation process, ensuring that the material is economically feasible. The minimum cycle life refers to the minimum number of cycles after which the performance degradation of the material does not exceed a certain percentage, usually requiring a performance degradation of less than 10% after more than 1000 cycles.

[0021] The multi-dimensional performance requirement indicators of inorganic hydrated salt phase change materials for target application scenarios can be directly obtained from users. Users fill in the required phase change temperature range, minimum phase change latent heat value, maximum allowable supercooling degree, viscosity range suitable for construction process, target cost upper limit, and minimum cycle life, etc.

[0022] The necessary process is described as follows: Taking building wall insulation as an example, the user requires the phase change temperature of the phase change material to be between 22-26℃, the latent heat value to be greater than 180J / g, the supercooling degree to be no more than 2℃, the viscosity to be within the range of 5000-8000mPa·s to adapt to spraying construction, and the cycle life to be more than 1000 times, and the target cost upper limit to be no more than 10 yuan per kilogram of material. These requirement indicators serve as target parameters for design optimization and are input into the system to provide a basis for subsequent intelligent design.

[0023] In step S200, the trained Gaussian process regression model is used to analyze the performance requirement indicators, generate candidate formulations, and output the performance prediction results of each candidate formulation.

[0024] Among them, Gaussian Process Regression (GPR): a non-parametric Bayesian regression method based on Gaussian process, used to model the nonlinear relationship between input variables (such as material component ratio) and output variables (such as material performance indicators). GPR model can provide prediction value and uncertainty estimation, suitable for small sample data modeling. Candidate formula: according to the performance requirement index, the possible material formula combination that meets the target performance requirement is generated by GPR model prediction. Performance prediction result: the predicted value of the GPR model for the performance indicators such as phase transition temperature, latent heat, supercooling degree and viscosity of the candidate formula.

[0025] The training process of a specific Gaussian Process Regression model can refer to steps S210 to S250, which will not be repeated here.

[0026] The necessary process is described as follows:

[0027] Taking the development of phase change materials for air conditioning cold energy storage as an example, the performance requirement indicators input by the user include phase transition temperature around 10℃, latent heat value ≥180J / g, supercooling degree <2℃, etc. Based on these requirement indicators, the trained GPR model is used for analysis.

[0028] First, the performance requirement indicators input by the user are used as the input conditions of the model. The GPR model generates a series of candidate formulas that may meet the requirements based on the relationship between the learned material formula and performance. For example, the model may generate the following candidate formulas and their performance prediction results: Formula A: sodium sulfate accounts for 70%, ammonium chloride accounts for 8%, and water accounts for 22%; predicted phase transition temperature 10.0℃, latent heat 200J / g, supercooling degree 1.8℃. Formula B: sodium sulfate accounts for 65%, ammonium chloride accounts for 12%, and water accounts for 23℃; predicted phase transition temperature 9.5℃, latent heat 190J / g, supercooling degree 2.0℃. Formula C: sodium sulfate accounts for 75%, ammonium chloride accounts for 5%, and water accounts for 20℃; predicted phase transition temperature 10.5℃, latent heat 205J / g, supercooling degree 1.5℃.

[0029] Step S300, based on the candidate formula and performance prediction result, with the predicted value of each performance indicator meeting the corresponding preset requirement threshold, and the predicted value of the formula cost ≤ the target cost upper limit as the constraint, a multi-objective optimization algorithm is used to screen out the Pareto optimal formula.

[0030] Among them, the preset demand threshold: the minimum or maximum acceptable value of the performance index set by the user according to the application scenario, such as the phase change temperature range, the minimum phase change latent heat value, the maximum allowed subcooling degree, etc. The formula cost prediction value: according to the current market price and the proportion of each component in the formula, the cost of unit mass material is calculated. Pareto optimal formula: under the condition of meeting all performance index preset demand threshold and cost constraint, the formula combination that cannot improve all performance indexes or reduce cost by changing formula proportion. Multi-objective optimization algorithm: an algorithm that can optimize multiple objective functions at the same time, such as non-dominated sorting genetic algorithm (NSGA-II) or particle swarm optimization algorithm (PSO), which balances the conflict between multiple objectives by generating a set of Pareto optimal solutions.

[0031] Determine optimization objectives and constraints: according to the performance requirement index and the upper limit of cost input by the user, set the objective function and constraint condition of multi-objective optimization problem. The objective function usually includes phase change temperature, phase change latent heat, subcooling degree and other performance indexes, and the constraint condition includes the preset demand threshold of each performance index and the predicted cost of formula, which does not exceed the target cost upper limit.

[0032] Select multi-objective optimization algorithm: select a suitable multi-objective optimization algorithm, such as non-dominated sorting genetic algorithm (NSGA-II) or particle swarm optimization algorithm (PSO), to optimize and screen the candidate formula.

[0033] Algorithm execution and result screening: run the multi-objective optimization algorithm to evaluate and screen the candidate formula, and get a set of Pareto optimal solutions that meet all the constraints. Select one or more formulas from the Pareto optimal solution set as the object of subsequent experimental verification.

[0034] Assuming that the multi-objective optimization algorithm uses the non-dominated sorting genetic algorithm, the main components and operation process of the non-dominated sorting genetic algorithm (NSGA-II) are as follows: 1. Initialize the population: randomly generate a set of initial formula combinations, each formula combination contains the proportion of each component. Calculate the performance prediction value (such as phase change temperature, latent heat, supercooling degree) and cost prediction value of each formula. 2. Non-dominated sorting: perform non-dominated sorting on each formula in the population, and divide the population into multiple non-dominated layers. The formulas in the non-dominated layer are not dominated by other formulas in all objective functions. For example, formula A and formula B, if A is better than B in latent heat, and B is better than A in cost, then A and B are not dominated by each other, and belong to the same non-dominated layer. 3. Calculate the crowding distance: in each non-dominated layer, calculate the crowding distance of each formula. The crowding distance represents the proximity of the formula in the target space, and the greater the distance, the more isolated the formula is in the target space. Through the crowding distance, the algorithm can maintain the diversity of the population and avoid falling into local optimum. 4. Selection operation: select formulas from the non-dominated layer to enter the next generation population. Preferentially select formulas in the non-dominated layer, if the number of formulas in the non-dominated layer exceeds the population size, select according to the crowding distance. The selection operation ensures that the algorithm maintains the diversity and convergence of the population during optimization. 5. Cross and mutation operation: cross the selected formulas to generate new formula combinations. Cross operation generates new offspring by combining the characteristics of two parent formulas. Mutate the generated offspring to introduce new genetic variation and increase the diversity of the population. 6. Evaluation and screening: calculate the performance prediction value and cost prediction value of the newly generated offspring. Combine the offspring with the parent formulas, re-sort and calculate the crowding distance, and select the next generation population. 7. Iterative optimization: repeat the above selection, cross, mutation, evaluation and screening process until the preset number of iterations is reached or the convergence condition is met. The convergence condition can be that the non-dominated layer of the population remains unchanged in multiple iterations, or the improvement of the objective function is less than a certain threshold. 8. Output the Pareto optimal solution set: finally output the Pareto optimal solution set that meets all the constraints, each solution is a formula combination and its performance prediction value. Select one or more formulas from the Pareto optimal solution set as the object of subsequent experimental verification.

[0035] The necessary process is described as follows:

[0036] Taking the development of phase change materials for air conditioning cold energy storage as an example, the user input performance requirement indicators include phase change temperature around 10℃, latent heat value ≥180J / g, supercooling degree <2℃, and target cost upper limit of 10 yuan / kg. Based on these demand indicators, a number of candidate formulas and their performance prediction results are generated using the Gaussian process regression model.

[0037] Next, the non-dominated sorting genetic algorithm (NSGA-II) is used to optimize and screen the candidate formulations. First, the candidate formulations and their performance prediction results are input into the NSGA-II algorithm, and the objective functions are set as the phase change temperature, latent heat, supercooling degree, and cost, with the constraint conditions being that the phase change temperature is between 9-11℃, the latent heat value is ≥180J / g, the supercooling degree is <2℃, and the cost is ≤10 yuan / kg.

[0038] After running the NSGA-II algorithm, a set of Pareto optimal solutions is obtained. From these solutions, a formulation is selected that satisfies all the constraint conditions and has the best overall performance. For example, the selected Pareto optimal formulation is: sodium sulfate accounts for 70%, ammonium chloride accounts for 8%, and water accounts for 22%. The predicted phase change temperature of this formulation is 10.0℃, the latent heat is 200J / g, the supercooling degree is 1.8℃, and the cost prediction value is 9.2 yuan / kg, fully meeting the user's performance requirements and cost constraints.

[0039] In step S400, the sample preparation and performance measurement of the Pareto optimal formulation are performed, the deviation values between the measured data and the performance prediction results in each performance indicator are calculated, and the deviation values are compared with the preset threshold.

[0040] Among them, sample preparation: according to the component proportion of the Pareto optimal formulation, accurately weigh each raw material, uniformly mix through mechanical stirring, ultrasonic dispersion, etc., and if necessary, control the temperature and time to ensure complete reaction / crystallization, to prepare the target phase change material sample. Performance measurement: measure the actual performance of the sample through experimental methods, including phase change temperature, latent heat, supercooling degree, viscosity, etc. Use a differential scanning calorimeter (DSC) to measure the phase change temperature and latent heat value, a cooling curve to measure the supercooling degree, and a rheometer to measure the viscosity, etc., to obtain the actual performance data of the sample. Deviation value: the difference between the measured performance value and the predicted performance value, usually expressed in absolute error or relative error. Deviation calculation: calculate the deviation value between the measured performance value and the predicted performance value, including phase change temperature deviation, latent heat deviation, supercooling degree deviation, and viscosity deviation, etc. Preset threshold: the allowed deviation range set by the user, used to judge whether the measured performance meets the prediction results.

[0041] The necessary procedures are as follows:

[0042] Taking the development of phase change materials for air conditioning cold energy storage as an example, the Pareto optimal formulation screened out in step S300 is: sodium sulfate accounts for 70%, ammonium chloride accounts for 8%, and water accounts for 22%. The predicted phase change temperature of this formulation is 10.0℃, the latent heat is 200J / g, the supercooling degree is 1.8℃, and the cost prediction value is 9.2 yuan / kg. The specific process is as follows:

[0043] 1. Sample preparation: Weigh 70 g of anhydrous sodium sulfate, 8 g of ammonium chloride, and 22 g of deionized water. Mix the components in a beaker, heat to 50-60°C to dissolve the sodium sulfate completely, then slowly cool to form a solid-liquid mixture, obtaining a uniform solid phase change material sample.

[0044] 2. Performance testing: Use a differential scanning calorimeter (DSC) to measure the phase change temperature and latent heat value of the sample, record the main melting point temperature and melting enthalpy value. Observe the difference between the crystallization point and melting point through the cooling curve, and record the supercooling degree. Use a rheometer to measure the viscosity of the sample in the molten state, and record the apparent viscosity value.

[0045] 3. Deviation calculation: Calculate the phase change temperature deviation: the measured phase change temperature is 10.3°C, the predicted phase change temperature is 10.0°C, and the deviation is 0.3°C. Calculate the latent heat deviation: the measured latent heat is 195 J / g, the predicted latent heat is 200 J / g, and the deviation is 5 J / g. Calculate the supercooling degree deviation: the measured supercooling degree is 1.8°C, the predicted supercooling degree is 1.8°C, and the deviation is 0°C. Calculate the viscosity deviation: the measured viscosity is 6000 mPa·s, the predicted viscosity is 5000 mPa·s, and the deviation is 1000 mPa·s.

[0046] 4. Threshold comparison: The user sets the preset threshold as follows: phase change temperature deviation ≤ 0.5°C, latent heat deviation ≤ 10 J / g, supercooling degree deviation ≤ 0.5°C, and viscosity deviation ≤ 1000 mPa·s. Compare the calculated deviation values with the preset threshold to determine whether the measured performance meets the predicted results. In this example, the phase change temperature deviation of 0.3°C, the latent heat deviation of 5 J / g, and the supercooling degree deviation of 0°C are all within the preset threshold range, but the viscosity deviation of 1000 mPa·s just reaches the upper limit of the threshold.

[0047] Step S500, if the deviation values of all performance indicators do not exceed the preset threshold, the formula is taken as the target formula.

[0048] Among them, the target formula: after measured verification, the performance deviation is within the preset threshold range, and meets the application requirements, which can be used as the final formula candidate.

[0049] Step S600, if the deviation value of any performance indicator exceeds the preset threshold, the measured data is supplemented to the training data set, the Gaussian process regression model is retrained, and the Pareto optimal formula is selected again, and the sample preparation, performance measurement and deviation comparison steps are repeated until one or more target formulas are obtained, all of which meet the preset threshold requirements.

[0050] The training data set is used to train the Gaussian process regression (GPR) model and contains the formula combination and its corresponding performance indicators. Feedback learning: new experimental data (measured data) is supplemented to the training data set, and the model is retrained to improve the prediction accuracy of the model.

[0051] The general process is as follows: 1. Bias value evaluation: According to the bias value calculated in step S400, evaluate whether the difference between the measured value and the predicted value of each performance indicator is within the user-set preset threshold range. 2. Data feedback: If the bias value exceeds the preset threshold, supplement the measured performance data of the formula to the training data set. 3. Model retraining: Use the updated training data set to retrain the Gaussian process regression model to improve the prediction accuracy of the model. 4. Re-screening: Use the retrained model to run the multi-objective optimization algorithm (such as NSGA-II) again to select new Pareto optimal formulas. 5. Iterative optimization: Repeat steps S400 to S600 until the bias values of all performance indicators are within the preset threshold range.

[0052] For example, develop a phase change material for air conditioning cold energy storage. Assuming that in step S400, the measured performance of a certain Pareto optimal formula deviates from the predicted performance as follows: phase change temperature deviation: 0.3℃ (preset threshold ≤0.5℃); latent heat deviation: 5J / g (preset threshold ≤10J / g); supercooling degree deviation: 0℃ (preset threshold ≤0.5℃); viscosity deviation: 1200mPa·s (preset threshold ≤1000mPa·s).

[0053] Since the viscosity deviation exceeds the preset threshold, step S600 needs to be performed, the process is as follows: 1. Data feedback: Supplement the measured performance data of the formula (phase change temperature 10.3℃, latent heat 195J / g, supercooling degree 1.8℃, viscosity 6000mPa·s) to the training data set. 2. Model retraining: Use the updated training data set to retrain the Gaussian process regression model. Adjust the model parameters through cross-validation and other methods to ensure that the model has good prediction accuracy. 3. Re-screening: Use the retrained model to run the non-dominated sorting genetic algorithm (NSGA-II) again to select new Pareto optimal formulas. 4. Iterative optimization: Prepare samples and measure the performance of the new Pareto optimal formulas, calculate the bias value of the measured data and the performance prediction result, and compare it with the preset threshold. Repeat the above process until the bias values of all performance indicators are within the preset threshold range.

[0054] Step S700, based on the preset comprehensive performance evaluation function, select the formula with the highest evaluation score from all obtained target formulas as the final formula, and use the continuous process matching the final formula to complete large-scale preparation.

[0055] The pre-set comprehensive performance evaluation function is used to evaluate the comprehensive performance of each candidate formula. This function combines the normalized values of multiple key performance indicators and takes into account the cost factor. The formula is as follows:

[0056] where S i is the comprehensive score of the i-th candidate formula. w j is the weight of the j-th performance indicator, and the sum of the weights is 1. p i,j is the normalized value of the j-th performance indicator of the i-th candidate formula, ranging from 0 to 1. λ is the cost penalty coefficient, usually set to 0.2. C i is the cost of the i-th candidate formula. C target is the upper limit of the target cost.

[0057] Then calculate the comprehensive score S i for all candidate formulas and sort them in descending order of score. Finally, select the formula with the highest score that meets all the hard thresholds as the final formula from the sorted candidate formulas.

[0058] The specific process of large-scale preparation using the continuous process matched to the final formula can refer to steps S710 to S760.

[0059] The training process of the Gaussian process regression model is as follows:

[0060] Step S210, determine the composition of the multi-component system containing the base hydrated salt and additives, including temperature adjuster, nucleating agent and rheological modifier, based on the phase change function requirements and pre-experiment data to determine the mass fraction range of each component.

[0061] where the base hydrated salt is the main component of the phase change material, with a specific phase change temperature and latent heat value, and is the core substance for realizing phase change energy storage. Additives: used to adjust the performance of the base hydrated salt, including temperature adjuster, nucleating agent and rheological modifier, etc., to meet the needs of specific application scenarios. Temperature adjuster: used to adjust the phase change temperature to make it closer to the temperature range required by the target application scenario. Nucleating agent: used to reduce the supercooling degree and improve the crystallization performance of the phase change material, to ensure the stability and reliability of the phase change process. Rheological modifier: used to adjust the viscosity and flowability of the material to adapt to different construction process requirements. Mass fraction range: the mass proportion range of each component in the formula, used to define the feasible space of the formula.

[0062] The general process is as follows: 1. Determine the base hydrated salt: According to the required phase change temperature range of the target application scenario, select the appropriate base hydrated salt. For example, for air conditioning cold energy storage, sodium sulfate decahydrate (Na2SO4·10H2O) can be selected as the base material, with a melting point of about 32.4°C. 2. Select additives: Temperature adjuster: select compounds that can form eutectic mixtures with the base hydrated salt, such as ammonium chloride (NH4Cl), etc., for adjusting the phase change temperature. Nucleating agent: select substances that can reduce supercooling, such as nano-silicon dioxide (SiO2), graphene oxide, etc. Rheological modifier: select substances that can adjust the viscosity and flowability of the material, such as diatomaceous earth, bentonite, etc. Define the mass fraction range: according to literature and pre-experiment data, determine the mass fraction range of each component. For example: 1. Base hydrated salt: 50%-80%. 2. Temperature adjuster: 0%-15%. 3. Nucleating agent: 0%-5%; 4. Rheological modifier: 0%-10%. Finally, ensure that the sum of the mass fractions of each component is 100%.

[0063] Step S220, the initial formula is obtained by small sample sampling of the ratio of the multi-component system using the preset design experiment method, the invalid formula is removed after reasonable verification, the effective sampling formula is reserved, the sample is prepared according to the effective sampling formula, and the phase change temperature, phase change latent heat value, supercooling degree, viscosity and cycle life data are measured, and the initial data set is constructed combined with the component mass fraction and corresponding performance data of the same type system in the historical literature. Among them, the design experiment method: a systematic experimental design method for efficiently selecting representative formula combinations in multi-component systems, common methods include central composite design (CCD), Latin hypercube design, etc. Small sample sampling: in the formula space of the multi-component system, a small number of representative formula combinations are selected for experiment to reduce the experimental workload and cost. Reasonable check: preliminary evaluation of the generated formula, eliminate those physically or chemically impossible formulas, such as exceeding the solubility range or leading to unstable formulas. Invalid formula: those that are physically or chemically infeasible, or obviously do not meet the target performance requirements. Effective sampling formula: after reasonable verification, the formula with actual feasibility and representativeness is retained. Initial data set: a data set containing effective sampling formula and its corresponding measured performance data (such as phase change temperature, latent heat value, supercooling degree, viscosity and cycle life), used for model training.

[0064] For example, develop a phase change material for air conditioning cold energy storage, the target phase change temperature is about 10℃, latent heat value ≥180J / g, supercooling degree <2℃, moderate viscosity, and low cost. The specific process is as follows: 1. Select the experimental method: choose central composite design (CCD) as the experimental method, because it can efficiently select representative formula combinations in multi-component systems. 2. Small sample sampling: set the mass fraction range of the base hydrated salt (sodium sulfate decahydrate) to 50%-80%, the temperature regulator (ammonium chloride) to 0%-15%, the nucleating agent (nano silicon dioxide) to 0%-5%, and the rheological modifier (diatomite) to 0%-10%. Use the CCD method to select 13 representative formula combinations within the above range. 3. Reasonableness check: preliminarily evaluate the 13 generated formulas, and eliminate those that are physically or chemically impossible. For example, some formulas may exceed the solubility range or cause unstable phase change processes. Assuming that after checking, 10 effective sampling formulas are retained. 4. Sample preparation and measurement: prepare samples according to the 10 retained effective sampling formulas, the specific steps are as follows: weigh the corresponding mass of each component, mix the base hydrated salt and additives, heat to 50-60℃ to fully dissolve. Slowly cool to room temperature to form a solid-liquid mixture. Use a differential scanning calorimeter (DSC) to measure the phase change temperature and latent heat value, a cooling curve to measure the supercooling degree, a rheometer to measure the viscosity, and a cycle test device to measure the cycle life. 5. Data integration: combine the above measured performance data with the component mass fraction and corresponding performance data of the same type of system in historical literature to construct an initial data set. Assuming that 5 data of the same type of system are obtained from historical literature, combine these data with the measured data to form an initial data set containing 15 data points.

[0065] Step S230, with the mass fraction of each component in the initial data set as the input feature, and the corresponding phase change temperature, latent heat value, supercooling degree, viscosity and cycle life as the output label, the data is standardized to construct the training data set.

[0066] Among them, the input feature: the mass fraction of each component in the initial data set, the input variable for model training. The output label: the performance indicators corresponding to the input feature, such as phase change temperature, latent heat value, supercooling degree, viscosity and cycle life, as the target variable for model training. Data standardization: scale the data proportionally to fall within a certain small interval (such as 0 to 1 or -1 to 1) to eliminate the dimensional influence between different features and improve the efficiency and accuracy of model training. Training data set: the combination of standardized input features and output labels, used to train the Gaussian process regression model.

[0067] For example, develop phase change materials for air conditioning cold energy storage, target phase change temperature is about 10℃, latent heat value ≥180J / g, supercooling degree <2℃, moderate viscosity, lower cost. In step S220, an initial data set containing 15 data points has been constructed. The specific process is as follows: 1. Extract input features and output labels: input features: mass fraction of each component, including base water-salt, temperature regulator, nucleating agent, rheological modifier and water. Output label: phase change temperature, phase change latent heat value, supercooling degree, viscosity and cycle life. 2. Data standardization: standardize the input features, normalize the mass fraction of each component to 0 to 1. Standardize the output label, normalize the phase change temperature, phase change latent heat value, supercooling degree, viscosity and cycle life to 0 to 1. For example, use Z-score standardization method. 3. Build training data set: combine the standardized input features and output labels to form the training data set.

[0068] Step S240, based on the training data set, the model parameters are iteratively optimized by Bayesian inference to fit the mapping relationship between components and performance, forming a formula-performance mapping model.

[0069] Wherein, Bayesian inference: a statistical method based on Bayes theorem, used to update the probability distribution of model parameters according to observed data, so as to optimize the model parameters. Formula-performance mapping model: the model obtained by training, which can predict the performance of phase change materials under different formula combinations.

[0070] The necessary procedures are as follows: 1. Initialize model parameters: Select Gaussian Process Regression (GPR) model, initialize kernel function parameters (such as length scale parameter of radial basis function RBF) and noise level, etc. For example, select RBF kernel function, initialize length scale parameter as 1.0, and noise level as 0.1. 2. Bayesian inference: Use the training data set to update the probability distribution of model parameters through Bayesian inference. The specific steps are as follows: Calculate the mean vector and covariance matrix of the training data. Update the posterior distribution of model parameters using Bayesian theorem. For example, for RBF kernel function, update the posterior distribution of length scale parameter and noise level. 3. Iterative optimization: Adjust model parameters through multiple iterations to gradually improve the fitting accuracy of the model to the training data. In each iteration, the optimization goal is to maximize the log-likelihood function. For example, use gradient descent or conjugate gradient method to optimize model parameters, gradually adjust length scale parameter and noise level, until the fitting error of the model to the training data no longer significantly decreases. 4. Model validation: After each iteration, use the validation set to evaluate the prediction accuracy of the model to ensure that the model does not overfit. The validation set can be a part of the training data set or generated by cross-validation method. For example, use cross-validation method to divide the training data set into training set and validation set, and evaluate the prediction error of the model on the validation set. 5. Form a mapping model: After multiple iterations of optimization, a model that can accurately predict the mapping relationship between formulations and performance is formed. The final model parameters include the optimized kernel function parameters and noise level. For example, after 10 iterations of optimization, the final RBF kernel function length scale parameter is 1.2, the noise level is 0.05, and the fitting error of the model to the training data is 0.01.

[0071] Step S250, the preset cross-validation method is used to verify the prediction error of the formulation-performance mapping model. If the prediction error of each performance indicator does not exceed the preset threshold, the Gaussian process regression model is directly determined as trained; if the prediction error exceeds the preset threshold, the error exceeding item is supplemented by directional sampling according to the aforementioned preset experimental design method, and the data is measured and included in the training data set, and the model parameter optimization step is repeated until the model prediction error meets the preset threshold, and the Gaussian process regression model is determined as trained.

[0072] Among them, the cross-validation method is a statistical analysis method used to evaluate and validate the prediction performance of the model. Common methods include k-fold cross-validation, leave-one-out cross-validation, etc. Prediction error: the difference between the model prediction value and the actual value, usually measured by mean square error (MSE), mean absolute error (MAE), etc. Preset threshold: user-set allowed prediction error range, used to judge whether the prediction accuracy of the model meets the requirements. Directional supplementary sampling: according to the model prediction error exceeding item, new formulation combinations are selected for experiments to supplement the training data set.

[0073] The general process is as follows: 1. Select the cross-validation method: according to the characteristics of the data set and the complexity of the model, select the appropriate cross-validation method, such as k-fold cross-validation. 2. Verify the prediction accuracy of the model: use the cross-validation method to evaluate the prediction accuracy of the model, and calculate the prediction error of each performance indicator. 3. Evaluate the prediction error: compare the prediction error with the preset threshold value to determine whether the prediction accuracy of the model meets the requirements. 4. Directional supplementary sampling: if the prediction error exceeds the preset threshold, according to the error exceeding item, directional supplementary sampling is carried out according to the preset design experiment method and the data is measured. 5. Update the training data set: incorporate the new experimental data into the training data set and re-optimize the model parameters. 6. Repeat the optimization step: repeat the model parameter optimization step until the model prediction error meets the preset threshold.

[0074] The specific examples are as follows: 1. Select the cross-validation method: select the 5-fold cross-validation method, divide the training data set into 5 subsets, use 4 subsets for training each time, and the remaining 1 subset for validation, repeat 5 times. 2. Verify the prediction accuracy of the model: use the 5-fold cross-validation method to evaluate the prediction accuracy of the model, and calculate the prediction error of each performance indicator, such as the mean square error (MSE) of the phase change temperature, the mean absolute error (MAE) of the latent heat value, etc. 3. Evaluate the prediction error: the preset threshold: phase change temperature MSE≤0.05, latent heat value MAE≤5J / g, supercooling degree MAE≤0.5℃, viscosity MAE≤500mPa·s, cycle life MAE≤30 times. Evaluation results: assuming that after the first cross-validation, the phase change temperature MSE=0.06, the latent heat value MAE=4J / g, the supercooling degree MAE=0.4℃, the viscosity MAE=600mPa·s, and the cycle life MAE=20 times. 4. Directional supplementary sampling: since the phase change temperature MSE and the viscosity MAE exceed the preset threshold, directional supplementary sampling is needed. According to the error exceeding item, select new formula combinations for experiments. For example, for the phase change temperature error, select the formula close to the phase change temperature boundary value; for the viscosity error, select the formula with high viscosity. 5. Update the training data set: prepare samples according to the new formula combination, and measure their phase change temperature, latent heat value, supercooling degree, viscosity and cycle life. Incorporate the new experimental data into the training data set and update the data set. 6. Repeat the optimization step: use the updated training data set to re-optimize the model parameters. Repeat the cross-validation step to evaluate the prediction accuracy of the model until the prediction error of all performance indicators meets the preset threshold.

[0075] The initial formula of the multi-component system is obtained by small sample sampling of the ratio of the preset design experiment method, including:

[0076] Step S221, determine the mass proportion variables of the base hydrated salt, temperature regulator, nucleating agent, and rheology modifier in the multi-component system and the preset value range, and the sum of the mass proportions of each component is constrained to be about 100%.

[0077] For example, develop a phase change material for air conditioning cold energy storage, the target phase change temperature is about 10℃, latent heat value ≥180J / g, supercooling degree <2℃, moderate viscosity, and low cost. The specific process is as follows: 1. Determine the base hydrated salt: choose sodium sulfate decahydrate (Na2SO4·10H2O) as the base material, its melting point is about 32.4℃, the latent heat value is high, but it needs to be adjusted by additives to adjust the phase change temperature. Mass proportion range: 50%-80% 2. Select additives: temperature regulator: choose ammonium chloride (NH4Cl) to reduce the phase change temperature from 32.4℃ to about 10℃. Mass proportion range: 0%-15%. Nucleating agent: choose nano silicon dioxide (SiO2) to reduce the supercooling degree and ensure the stability of the phase change process. Mass proportion range: 0%-5%. Rheology modifier: choose diatomite to adjust the viscosity and flowability of the material, making it suitable for pouring and packaging. Mass proportion range: 0%-10%.

[0078] 2. Define the mass proportion range: 1. Base hydrated salt (sodium sulfate decahydrate): 50%-80%; 2. Temperature regulator (ammonium chloride): 0%-15%; 3. Nucleating agent (nano silicon dioxide): 0%-5%; 4. Rheology modifier (diatomite): 0%-10%; water: the rest, to ensure that the sum of the mass proportions of each component is 100%.

[0079] Step S222, based on the determined mass proportion variables and ranges of each component, use central composite design as the preset design experiment method, generate an experimental point set containing vertex, axis point and center point in the formula space defined by the above variables according to the preset small sample size, the structure of the experimental point set constitutes a second-order experimental design for analyzing linear and quadratic interaction effects, and the corresponding ratio value combination of each experimental point is the initial formula obtained by small sample sampling.

[0080] The detailed process is described as follows:

[0081] 1. Select central composite design (CCD): Central composite design (CCD) is an experimental design method that can efficiently explore the combination of factor levels in multi-factor experiments, covering linear effects and quadratic interaction effects of the formula space. CCD design includes three parts: vertex, axis point and center point.

[0082] 2. Generate vertices: Vertices are combinations of the minimum or maximum values of the mass fractions of each component. For 4 factors, there are 24 = 16 vertices. For example, one vertex could be 50% base hydrate, 0% temperature modifier, 0% nucleating agent, 0% rheology modifier, and 50% water. Another vertex could be 80% base hydrate, 15% temperature modifier, 5% nucleating agent, 10% rheology modifier, and 0% water.

[0083] 3. Generate axis points: Axis points are used to explore the interaction effects between factor levels. Axis points are usually located at the middle values of the factor levels. For example, an axis point could be 65% base hydrate (middle value), 7.5% temperature modifier (middle value), 2.5% nucleating agent (middle value), 5% rheology modifier (middle value), and 17.5% water.

[0084] 4. Generate center points: Center points are points where all factor levels take the middle values, used to evaluate experimental error and curvature of the model. For example, a center point could be 65% base hydrate, 7.5% temperature modifier, 2.5% nucleating agent, 5% rheology modifier, and 17.5% water.

[0085] 5. Combine experimental points: Combine vertices, axis points, and center points to form a complete set of experimental points. Each experimental point corresponds to a specific formulation combination, which covers the linear effects and quadratic interaction effects of the formulation space.

[0086] The initial formulation obtained by the preset design experiment method for the ratio of the multi-component system to the small sample also includes the following methods, which are as follows:

[0087] Step 1, using the preset multi-component system high nonlinearity response judgment condition to judge the response characteristics of the multi-component system composition.

[0088] 1. Define the judgment condition of high nonlinearity response:

[0089] High nonlinearity response: refers to the interaction between components in a multi-component system, which leads to a high nonlinearity characteristic of the performance of the system to the component ratio. This nonlinearity characteristic means that a simple linear model cannot accurately describe the relationship between the component ratio and the performance.

[0090] Judgment condition: a series of conditions are preset to judge whether the multi-component system has high nonlinearity response. Common judgment conditions include:

[0091] Historical data or literature: refer to existing experimental data or literature to judge whether similar systems exhibit high nonlinearity response.

[0092] Preliminary experimental data: Through a small number of preliminary experiments, observe the response curve of performance indicators to component allocation ratios, and determine whether there are obvious nonlinear characteristics.

[0093] Correlation analysis: Calculate the correlation coefficient between each component allocation ratio and performance indicators. If the correlation coefficient is low, it indicates that there is a nonlinear relationship.

[0094] Model fitting goodness: Use a linear model to fit preliminary data and calculate the goodness of fit (such as R² value). If the goodness of fit is low, it indicates a highly nonlinear response.

[0095] 2, Collect and analyze data:

[0096] Collect data: Collect relevant data for multi-component systems, including historical experimental data, literature data, and preliminary experimental data.

[0097] Preliminary experiments: Design and conduct a small number of preliminary experiments to measure performance indicators at different component allocation ratios, such as phase change temperature, latent heat value, supercooling degree, viscosity, and cycle life.

[0098] Correlation analysis: Calculate the correlation coefficient between each component allocation ratio and performance indicators to assess the strength of the linear relationship.

[0099] Model fitting: Use a linear model to fit preliminary data and calculate the goodness of fit (such as R² value) to assess the applicability of the linear model.

[0100] 3, Determine the response characteristics:

[0101] Evaluate nonlinear characteristics: Based on the analysis results above, evaluate the nonlinear characteristics of multi-component systems. If the correlation coefficient is low and the goodness of fit of the linear model is poor, it is determined that the system has a highly nonlinear response. Record the determination results: Record the response characteristic determination results of the multi-component system to provide a basis for subsequent steps.

[0102] Step 2, if it is determined to be a highly nonlinear response system, then collect the basic ratio values of the base hydrate salt and additives and the corresponding performance data according to the pre-set sample size to construct an initial sample set.

[0103] The detailed process is as follows:

[0104] 1, Determine the pre-set sample size: Determine the initial sample size according to experimental design and resource constraints. The sample size should be small enough to reduce experimental costs, but large enough to provide sufficient information for model construction. For example, the pre-set sample size can be 10 to 20 samples.

[0105] 2. Select base ratio values: Based on the composition of the multi-component system, select the base ratio values of the base hydrate salt and additives. These ratio values should cover the key areas of the formulation space, including the minimum, maximum, and intermediate values of each component. For example, select the base ratio values of the base hydrate salt (50%, 65%, 80%), temperature modifier (0%, 7.5%, 15%), nucleating agent (0%, 2.5%, 5%), and rheology modifier (0%, 5%, 10%).

[0106] 3. Collect base ratio values and corresponding performance data: According to the preset sample size, select the base ratio value combination, and prepare the corresponding sample. Perform performance testing on each sample to measure performance indicators such as phase change temperature, latent heat value, supercooling degree, viscosity, and cycle life. For example, prepare samples with the following base ratio values and test their performance: Sample 1: base hydrate salt 50%, temperature modifier 0%, nucleating agent 0%, rheology modifier 0%, water content 50%. Sample 2: base hydrate salt 80%, temperature modifier 15%, nucleating agent 5%, rheology modifier 10%, water content 0%. Sample 3: base hydrate salt 65%, temperature modifier 7.5%, nucleating agent 2.5%, rheology modifier 5%, water content 17.5%. And so on, until the preset sample size is reached.

[0107] 4. Construct the initial sample set: Combine the collected base ratio values and corresponding performance data to form the initial sample set. The initial sample set should include detailed ratio information and performance test results for each sample, which can be used for subsequent model construction and optimization.

[0108] Step 3, based on the initial sample set, construct a nonlinear surrogate model through Gaussian process regression, taking the ratio values as input and the performance data as output, outputting the performance prediction value and prediction variance corresponding to the ratio values.

[0109] Assuming we have constructed the initial sample set through step 2, we now need to construct a nonlinear surrogate model through Gaussian process regression. The specific process is as follows:

[0110] 1. Prepare the initial sample set: The initial sample set contains 15 samples, each including the mass percentage of each component and the corresponding performance indicators. For example: Sample 1: base hydrate salt 50%, temperature modifier 0%, nucleating agent 0%, rheology modifier 0%, water content 50%, phase change temperature 10.2℃, latent heat 190J / g, supercooling degree 1.8℃, viscosity 5500mPa·s, cycle life 1020 times. Sample 2: base hydrate salt 80%, temperature modifier 15%, nucleating agent 5%, rheology modifier 10%, water content 0%, phase change temperature 10.5℃, latent heat 205J / g, supercooling degree 1.5℃, viscosity 6500mPa·s, cycle life 1080 times. And so on, until 15 samples.

[0111] 2. Select Gaussian Process Regression (GPR) model: Select a Gaussian Process Regression (GPR) model suitable for modeling and predicting the nonlinear relationship between input variables and output variables.

[0112] 3. Construct training dataset: Use the ratio values in the initial sample set as input features and performance data as output labels to construct the training dataset. Standardize the input features and output labels. For example, use the Min-Max normalization method to normalize the input features to 0 to 1, and use the Z-score standardization method to normalize the output labels to a distribution with a mean of 0 and a standard deviation of 1.

[0113] 4. Select kernel function: Select the Radial Basis Function (RBF) as the kernel function, initialize the length scale parameter to 1.0, and the signal variance to 1.0.

[0114] 5. Train GPR model: Train the GPR model using the training dataset, optimize the kernel function parameters by maximizing the log-likelihood function. Update the probability distribution of model parameters using Bayesian inference to ensure the model can accurately fit the training data.

[0115] 6. Model validation: Use 5-fold cross-validation method to evaluate the prediction accuracy of the model, calculate the Mean Squared Error (MSE) and Mean Absolute Error (MAE). Ensure that the fitting error of the model on the training data is within a reasonable range, while avoiding overfitting. For example, the MSE should be less than 0.1, and the MAE should be less than 5J / g.

[0116] 7. Output prediction value and prediction variance: For a new ratio value, use the trained GPR model to make predictions, output the performance prediction value and its prediction variance. For example, for a new ratio value (base hydrate salt 60%, temperature regulator 10%, nucleating agent 3%, rheological modifier 7%, water 20%), the model predicts the phase change temperature to be 10.4℃, the latent heat value to be 198J / g, the supercooling degree to be 1.6℃, the viscosity to be 6200mPa·s, and the cycle life to be 1040 times, with prediction variances of 0.01, 0.02, 0.01, 0.03, and 0.02, respectively.

[0117] Step 4, call the preset expected improvement function, and calculate the optimization value by substituting the performance prediction value and prediction variance. Define the area where the optimization value is greater than the preset value threshold and the prediction variance is greater than the preset variance threshold as the high information sampling area.

[0118] 1. Define the expected improvement function: The expected improvement function is a sampling criterion for Bayesian optimization, used to measure the potential improvement value of a new sampling point. It combines the prediction value and prediction variance, aiming to balance exploration and utilization.

[0119] The definition of the expected improvement function is: ; wherein, is the current known best performance value, is the predicted performance value of the new sampling point x.

[0120] 2. Calculate the optimization value: For each candidate sampling point, use the predicted value and the predicted variance of the Gaussian process regression model to calculate its optimization value by substituting into the expected improvement function. The optimization value reflects the potential improvement of the new sampling point relative to the current best point.

[0121] 3. Set the preset threshold:

[0122] Value threshold: Set a preset value threshold to filter out areas with high optimization value. For example, set the value threshold to 0.1.

[0123] Variance threshold: Set a preset variance threshold to filter out areas with high uncertainty. For example, set the variance threshold to 0.05.

[0124] 4. Define high information sampling areas: For each candidate sampling point, check whether its optimization value is greater than the preset value threshold, and at the same time check whether its predicted variance is greater than the preset variance threshold. If both conditions are met, the area is divided into high information sampling areas, which have high improvement potential and high uncertainty, and are the key areas for subsequent sampling.

[0125] Step 5, in the high information sampling area, select the ratio value according to the preset encryption ratio for experimental sampling, measure the corresponding performance data, and supplement the newly collected ratio value and performance data to the initial sample set to update the sample set.

[0126] The detailed process is as follows: 1. Determine the high information sampling area: According to the results of step 4, determine the high information sampling area. These areas are the areas with optimization value greater than the preset value threshold and predicted variance greater than the preset variance threshold, with high improvement potential and high uncertainty. 2. Set the encryption ratio: Pre-set an encryption ratio to determine the number of samples that need to be added in the high information sampling area. For example, the encryption ratio can be 20% to 50% of the initial sample amount. 3. Select the ratio value: In the high information sampling area, select new ratio values uniformly or randomly according to the preset encryption ratio. These new ratio values should cover the key areas of the high information sampling area to ensure the representativeness of the sampling. 4. Experimental sampling: Prepare samples according to the selected ratio values and conduct experimental tests to measure the corresponding performance data, such as phase change temperature, phase change latent heat value, supercooling degree, viscosity and cycle life, etc. 5. Update the sample set: Supplement the newly collected ratio value and its corresponding performance data to the initial sample set to form the updated sample set.

[0127] Step 6: Iteratively optimize the nonlinear surrogate model based on the updated sample set, calculate the error reduction rate of the nonlinear surrogate model before and after iteration on the preset validation set, repeat steps 3 to 5 until the preset iteration number is reached or the error reduction rate is less than the preset reduction threshold.

[0128] The detailed process is as follows: 1. Update the sample set: supplement the new data points (ratio and corresponding performance data) collected in step 5 to the initial sample set to form an updated sample set. This step ensures that the sample set contains more data points in the high information area, which helps improve the prediction accuracy of the model. 2. Re-train the nonlinear surrogate model: re-train the Gaussian process regression (GPR) model using the updated sample set. This step aims to optimize the model parameters using new data points to improve the model's fitting ability for the relationship between the formula and performance. 3. Calculate the prediction error: evaluate the performance of the updated model on the preset validation set. The validation set is a data set independent of the training set, used to evaluate the generalization ability of the model. Calculate the prediction error of the model on the validation set, such as mean square error (MSE) or mean absolute error (MAE). 4. Calculate the error reduction rate: compare the prediction error of the current iteration model with the last iteration model on the validation set, and calculate the error reduction rate. The error reduction rate reflects the degree of improvement in model performance. 5. Determine whether the termination condition is met: set a preset error reduction threshold, for example, 0.01. If the error reduction rate is less than the threshold, or the preset iteration number (such as 10) is reached, stop the iteration process. If the termination condition is not met, return to step 3 and continue to optimize the model.

[0129] Step 7: Combine the final generated ratio set as the initial formula obtained by small sample sampling.

[0130] The detailed process is as follows: 1. Integrate all sampling data: integrate all ratio values and their corresponding performance data generated during the iterative optimization process in the previous steps (steps 3 to 6). These data include the initial sample set, the encrypted sample in the high information sampling area, and the newly added sample in the subsequent iteration process. 2. Form the initial formula set: use the integrated ratio combination and its performance data as the initial formula set obtained by small sample sampling. This formula set contains formulas that have been optimized and verified multiple times, and has high representativeness and reliability.

[0131] Using the trained Gaussian process regression model to analyze the performance requirement index, generating candidate formulas and outputting the performance prediction results of each candidate formula include:

[0132] Step S2A0, input the preset multi-dimensional performance requirement index as a constraint condition into the trained Gaussian process regression model, generate a candidate formula set that meets the constraint condition through the model search, and output the component proportion, performance prediction value and prediction variance of each formula.

[0133] The detailed process is described as follows: 1. Input performance requirement indicators: input the multi-dimensional performance requirement indicators preset by the user as constraint conditions into the trained Gaussian Process Regression (GPR) model. These performance requirement indicators usually include phase change temperature range, minimum phase change latent heat value, maximum allowable supercooling degree, viscosity range suitable for construction process, target cost upper limit, and minimum cycle life, etc. 2. Model search to generate a set of candidate formulations: use the prediction ability of the GPR model to search for candidate formulations that meet all the constraint conditions within the formulation space. The GPR model can provide performance prediction values and their prediction variances for each candidate formulation, which helps to evaluate the uncertainty and improvement potential of the formulation. 3. Output formulation information: for each candidate formulation that meets the constraint conditions, output its component proportion, performance prediction value, and prediction variance. These information will provide important basis for subsequent optimization and decision-making.

[0134] Step S2B0, call the preset expected improvement function, substitute the performance prediction value and the prediction variance to calculate the initial optimization value of each candidate formulation, and form a preliminary ranking.

[0135] For each candidate formulation, use its performance prediction value and prediction variance to calculate the initial optimization value. The formula of the expected improvement function is:

[0136] wherein, is the best known performance value. is the performance prediction value of the candidate formulation x. is the prediction variance of the candidate formulation x. Φ is the cumulative distribution function of the standard normal distribution. is the probability density function of the standard normal distribution.

[0137] According to the calculated initial optimization value, all candidate formulations are ranked. The higher the initial optimization value, the greater the potential improvement value of the formulation, and the higher the priority.

[0138] Step S2C0, call the SHAP explainability tool, for each candidate formulation, calculate the Shapley value of each component to the performance indicator and take the absolute value to obtain the contribution degree, and define the components with contribution degree > preset contribution threshold as the key components of the formulation.

[0139] ​The detailed process is as follows: 1. Prepare the data: prepare the data set for SHAP analysis, including the proportion of each candidate formula component as input features, and the corresponding performance prediction value as output label. These data should come from the prediction results of Gaussian Process Regression (GPR) model. 2. Call SHAP tool: select appropriate SHAP tool, such as shap library, to calculate the Shapley value of each input feature. Shapley value is a method based on game theory to measure the contribution of each feature to model prediction. 3. Calculate Shapley value: use SHAP tool to calculate the Shapley value of each component. Shapley value represents the average contribution of each component to performance prediction. For each candidate formula, calculate the Shapley value of each component. 4. Calculate contribution: take the absolute value of the Shapley value of each component as the contribution. The contribution reflects the importance of each component to performance prediction. 5. Set contribution threshold: set a contribution threshold to filter out components that have significant contribution to performance prediction. For example, set the contribution threshold to 0.1. 6. Define key components: filter out components with contribution greater than the preset contribution threshold, and define these components as key components. Key components have significant impact on performance prediction and are the focus of subsequent optimization.

[0140] Step S2D0, based on the key components of each candidate formula and their contribution, the initial optimization value is modified by a preset adjustment coefficient.

[0141] The general process is as follows: 1. Extract the contribution of key components: from step S2C0, we have calculated the contribution of each key component, and selected the key components with contribution greater than the preset contribution threshold. 2. Set adjustment coefficient: preset an adjustment coefficient to control the influence of key component contribution on initial optimization value. The adjustment coefficient is usually a value between 0 and 1, for example 0.5. 3. Calculate adjustment factor: take the contribution of key components as adjustment factor. The adjustment factor reflects the importance of each key component in the optimization process. Specifically, calculate the sum of the standardized contribution of all key components . Assuming the contribution of key components is:

[0142] Shapley A =0.2; Shapley B =0.3; Shapley C =0.1.

[0143] Then the adjustment factor is: =0.2+0.3+0.1=0.6.

[0144] 4. Correct the initial optimization value: Embed the adjustment factor into the initial optimization value by a preset adjustment coefficient, and calculate the corrected optimization value. The correction formula is: .

[0145] wherein EI(x) is the initial optimization value, a is the preset adjustment coefficient, is the sum of the normalized contribution degrees of all key components.

[0146] Step S2E0, reorder the candidate formula set based on the corrected optimization value, and mark the formula whose ranking promotion amplitude > preset promotion threshold as an anti-intuitive formula.

[0147] The detailed process is described as follows: 1. Reorder the candidate formula set: According to the corrected optimization value calculated in step S2D0, reorder all candidate formulas. The higher the corrected optimization value, the greater the potential improvement value of the formula, and the higher the priority. 2. Set the ranking promotion threshold: A ranking promotion threshold is preset to identify formulas with significant ranking promotion. For example, set the ranking promotion threshold to 2 positions. 3. Calculate the ranking promotion amplitude: Compare the ranking of each candidate formula in the preliminary sorting (step S2B0) and the reordering (step S2D0), and calculate the ranking promotion amplitude. 4. Mark the anti-intuitive formula: Mark the formula whose ranking promotion amplitude is greater than the preset promotion threshold as an anti-intuitive formula.

[0148] Step S2F0, generate an explanation report for the anti-intuitive formula using a preset generation method.

[0149] Among them, the common generation methods include: text report: detailed description of the component proportion, performance prediction value, initial optimization value, corrected optimization value and ranking promotion amplitude of the anti-intuitive formula. Visual chart: use charts (such as bar chart, line chart) to intuitively display the changes of performance prediction value and optimization value of the anti-intuitive formula. SHAP value visualization: through the visualization of SHAP value, the contribution of key components to performance prediction is displayed.

[0150] The specific process of step S2F0 can refer to steps S2F1 to S2F6, which will not be repeated here.

[0151] Step S2G0, output the final candidate formula list and supporting explanation report.

[0152] Based on the key components and their contribution degrees of each candidate formula, the initial optimization value is corrected by a preset adjustment coefficient, including:

[0153] Step S2D1, based on the key components of each candidate formula and their contribution degrees, the initial optimization value is corrected by a preset adjustment coefficient, and the performance prediction value, prediction variance and initial optimization value of the candidate formula are extracted synchronously to form a parameter dataset to be corrected.

[0154] The detailed process is described as follows: 1. Obtain the key components and their contribution degrees data: From step S2C0, extract the key components and their contribution degrees data of each candidate formula. These data include the name of each key component and the corresponding contribution degree normalized value. 2. Extract the performance prediction value, prediction variance and initial optimization value of the candidate formula: From steps S2A0 and S2B0, extract the performance prediction value, prediction variance and initial optimization value of each candidate formula. These data will be used for subsequent correction calculation.

[0155] 3. Integrate to form a parameter dataset to be corrected: integrate the contribution degree data of the key components, performance prediction value, prediction variance and initial optimization value into a dataset to form a parameter dataset to be corrected.

[0156] Step S2D2, based on the phase change material field knowledge base, set the preset adjustment coefficient, and assign mechanism priority weight according to the priority of the key components to the core performance demand of the formula.

[0157] The general process is described as follows: 1. Build a phase change material field knowledge base: collect and organize relevant knowledge in the field of phase change materials, including the regulation mechanism of each component on performance, historical experimental data, literature data, etc. These knowledge will be used as the basis for setting the preset adjustment coefficient and assigning mechanism priority weight. 2. Set the preset adjustment coefficient: according to the field knowledge base, set a preset adjustment coefficient to control the influence degree of key component contribution degree on initial optimization value. The adjustment coefficient is usually a value between 0 and 1, for example 0.5. 3. Determine the regulation priority of key components to performance: analyze the regulation mechanism of key components to performance, and determine the regulation priority of each key component to performance. Analyze the regulation mechanism: for example, some components may have a significant impact on phase change temperature, while others may have a greater contribution to latent heat value. Through data and literature analysis in the field knowledge base, determine the regulation ability of each key component to different performance indicators. Determine the priority: according to the importance of performance indicators in the target application scenario, determine the regulation priority of key components. For example, if the target application scenario has the highest requirement on phase change temperature, the components that have a significant impact on phase change temperature will get higher priority. 4. Assign mechanism priority weight: assign mechanism priority weight according to the regulation priority of key components. Assign weight: assign a weight to each key component to reflect its importance in the optimization process. The weight can be assigned according to the regulation priority, and the components with higher priority get higher weight. The sum of the weights should be 1 to ensure the relative nature of the weights.

[0158] Step S2D3, based on the performance prediction value, prediction variance and key component contribution degree in the parameter data set to be corrected, a composite adjustment factor integrating the key component contribution degree and the mechanism priority weight is constructed, which is introduced into the optimization value correction calculation to derive a correction formula.

[0159] The process of constructing the composite adjustment factor is as follows: 1. Normalization of key component contribution degree: the contribution degree of each key component is normalized to make its value between 0 and 1. The normalization formula is: ; wherein, is the contribution degree of the i-th key component, and n is the total number of key components. 2. Mechanism priority weight: according to the mechanism priority weight ω i assigned in step S2D2, the weight of each key component reflects its importance in the formula. 3. Composite adjustment factor: the normalized contribution degree and the mechanism priority weight are combined to construct the composite adjustment factor: .

[0160] Derivation of correction formula: the composite adjustment factor is introduced into the optimization value correction calculation to derive a correction formula. The correction formula is as follows:

[0161] . wherein, is the initial optimization value. is a preset adjustment coefficient, usually between 0 and 1. is the composite adjustment factor of the i-th key component. is the sum of the composite adjustment factors of all key components.

[0162] Step S2D4, according to the correction formula and the initial optimization value in the parameter data set to be corrected, hierarchical correction calculation is implemented according to the contribution degree of key components to update the optimization value of each candidate formula.

[0163] Hierarchical correction calculation is performed according to the following logic:

[0164] 1. Level division: for each candidate formula, the system first sorts all the identified "key components" according to their contribution degree values from high to low. This sorting naturally forms a correction priority hierarchy, with the component with the highest contribution degree at the first level, the component with the second highest contribution degree at the second level, and so on.

[0165] 2. Sequential correction: then, the system applies the correction formula to each level in turn. The correction process starts from the level with the highest contribution degree, substitutes the data of the key components in this level into the correction formula to correct the current optimization value; then, uses the result of the last correction as the new current value, and repeats the process in the next level until all the key components in all levels are corrected.

[0166] 3. Value Update: The final result obtained after the above sequence of calculations is the revised optimization value of the candidate formulation. The system uses this new value to update the optimization parameters of the formulation.

[0167] Step S2D5, using the mechanism consistency verification method based on historical experimental data comparison, verifies the matching degree of the revised optimization value and the actual role of the key components.

[0168] Among them, "mechanism consistency verification method based on historical experimental data comparison" is a method to verify the consistency of model prediction results and actual physical and chemical mechanisms. By comparing the model prediction value with historical experimental data, the accuracy of model prediction is evaluated. The core of this method is to use existing experimental data as a benchmark to ensure that the correction results of the model conform to the actual physical and chemical laws.

[0169] The specific steps of the method are as follows: First, collect all historical experimental data related to the current research, including performance test results of different formulations such as phase change temperature, latent heat value, supercooling degree, viscosity and cycle life, etc. Then organize the collected data into a structured form to ensure the accuracy and completeness of the data. Then, according to the importance of the target application scenario, assign weights to each performance indicator. The allocation of weights should be based on the knowledge of domain experts and the analysis of historical data, and ensure that the sum of all weights is 1 to maintain the relative nature of the weights. Finally, calculate the weighted average value of each performance indicator in the historical experimental data as the benchmark value.

[0170] The calculation formula of the weighted average value is: .

[0171] Among them, is the weighted average value of the i-th performance indicator. is the actual value of the i-th performance indicator in the j-th sample. is the weight of the j-th sample. m is the total number of samples of historical experimental data.

[0172] 4. Compare model prediction value with historical data benchmark value:

[0173] Calculate the deviation: for each candidate formulation, calculate the deviation between the predicted value of each performance indicator and the historical data benchmark value. The calculation formula of the deviation is:

[0174] .

[0175] Among them, is the value of the i-th performance indicator predicted by the model.

[0176] is the weighted average value of the i-th performance indicator in the historical data.

[0177] 5. Calculate the matching degree: Based on the weight and deviation of the performance indicators, calculate the matching degree. The calculation formula of the matching degree is:

[0178] ;

[0179] Wherein, is the matching degree, between 0 and 1. is the deviation of the i-th performance indicator. is the weight of the i-th performance indicator. n is the total number of performance indicators.

[0180] 6. Set the matching degree threshold: According to the actual demand and domain knowledge, preset a matching degree threshold. For example, the matching degree threshold can be set to 0.8.

[0181] 7. Check the matching degree: Calculate the matching degree of each candidate formula and compare it with the preset matching degree threshold.

[0182] Step S2D6, if the matching degree is greater than or equal to the preset threshold, directly output the revised optimization value, the adjustment weight proportion and the verification result, and feed back to the candidate formula reordering step.

[0183] Wherein, the adjustment weight proportion: In the revision formula, the product of the key component contribution degree and the mechanism priority weight accounts for the proportion of the total revision amount. The verification result: The verification result of the mechanism consistency method, including the matching degree value and the judgment of whether it meets the expectation.

[0184] Step S2D7, if the matching degree is less than the preset threshold, dynamically adjust the preset adjustment coefficient and the mechanism priority weight according to the deviation ratio, and execute the composite adjustment factor construction, hierarchical revision calculation and matching degree verification steps based on the to-be-revised parameter data set again until the matching degree is greater than or equal to the preset threshold or the preset maximum iteration number is reached, and then feed back to the candidate formula reordering step.

[0185] Deviation ratio (DeviationRatio): Measure the gap between the matching degree and the preset threshold, the calculation formula is: ; wherein, is the preset matching degree threshold, is the matching degree of the current candidate formula.

[0186] Adjust the adjustment coefficient and the weight: 1. The preset adjustment coefficient (α): used to control the influence degree of the key component contribution degree on the initial optimization value. The mechanism priority weight (mechanism_weighti): reflects the importance of each key component in performance control.

[0187] The adjustment formula is: ;

[0188] .

[0189] The preset generation means for generating an explanation report for an anti-intuitive formula includes the following steps:

[0190] In step S2F1, the component proportion, performance prediction value, key component contribution degree, and ranking change data of the anti-intuitive formula are obtained, and the historical data of the same type of traditional formula are synchronously extracted as a reference benchmark.

[0191] The necessary process is described as follows:

[0192] 1. Extracting anti-intuitive formula data: Obtain detailed data of the anti-intuitive formula from the optimization model, including component proportion, performance prediction value, key component contribution degree, and ranking change. For example, the component proportion of the anti-intuitive formula is 60% for the base hydrated salt, 10% for the temperature regulator, 3% for the nucleating agent, 7% for the rheological modifier, and 20% for water; the performance prediction value is 10.2°C for the phase change temperature, 190 J / g for the latent heat, 1.8°C for the supercooling degree, 6000 mPa·s for the viscosity, and 1050 times for the cycle life; the key component contribution degree is 0.8 for the base hydrated salt and 0.6 for the temperature regulator; and the ranking is improved from the 3rd to the 1st.

[0193] 2. Extracting traditional formula data: Extract data of the same type of traditional formula from the historical database or literature as a reference benchmark. For example, the component proportion of the traditional formula is 70% for the base hydrated salt, 5% for the temperature regulator, 2% for the nucleating agent, 3% for the rheological modifier, and 20% for water; the performance prediction value is 10.5°C for the phase change temperature, 180 J / g for the latent heat, 2.0°C for the supercooling degree, 5500 mPa·s for the viscosity, and 1000 times for the cycle life.

[0194] In step S2F2, the differences between the anti-intuitive formula and the traditional formula design rules are quantified by preset comparison parameters to form an intuitive comparison table, which constitutes the core content of the phenomenon description.

[0195] Among them, the preset comparison parameters are pre-set performance indicators for comparing the anti-intuitive formula and the traditional formula, such as phase change temperature, latent heat value, supercooling degree, viscosity, and cycle life. The difference characteristics are the differences between the anti-intuitive formula and the traditional formula in the preset comparison parameters, which are obtained by quantitative analysis. The intuitive comparison table: The differences between the anti-intuitive formula and the traditional formula in each preset comparison parameter are displayed in table form, which is convenient for intuitive understanding.

[0196] In step S2F3, the SHAP visualization tool is called to generate SHAP force explanation graphs and component influence relationship graphs to analyze the contribution direction, synergistic effect, and nonlinear action path of the key components, identify the characteristic combination that significantly conflicts with the traditional design rules, and establish the evidence base for mechanism analysis.

[0197] Force Plot: A visualization tool that shows the direction (positive or negative) and magnitude of the impact of each feature on the model's prediction. It displays the contribution of each feature through intuitive bar plots, helping to identify key components and their specific influence on the model's prediction. Component Influence Diagram: A visualization tool that reveals the interactions between features and their combined impact on the model's prediction. It uncovers the mechanisms of feature combinations, including synergistic effects and nonlinear interaction pathways, aiding in understanding the comprehensive influence of components in complex systems.

[0198] 1. Call SHAP Tool: Use the SHAP (SHapley Additive exPlanations) tool, a game-theoretic explanation method that assigns a SHAP value to each feature, representing its contribution to the model's prediction. SHAP provides various visualization methods suitable for explaining complex machine learning models.

[0199] Assuming we have a trained machine learning model for predicting the performance of inorganic hydrate salt phase change materials, we use the SHAP tool to explain the model's predictions. SHAP values are calculated based on the model's output and input features, quantifying the specific contribution of each feature to the model's prediction.

[0200] 2. Generate Force Plot: Use the SHAP tool to generate a force plot, which displays the direction and magnitude of the impact of each feature on the model's prediction. The force plot visually shows the contribution of each feature through bar plots, with positive bars indicating positive contributions and negative bars indicating negative contributions. This step helps identify key components and their specific influence direction on the model's prediction.

[0201] Assuming we have a feature A with a SHAP value of 0.5, indicating a positive contribution to the model's prediction with a magnitude of 0.5. In the force plot, feature A will be displayed with a positive bar of length 0.5. In this way, we can visually see the impact of each feature on the model's prediction.

[0202] 3. Generate Component Influence Diagram: Use the SHAP tool to generate a component influence diagram, which reveals the interactions between features and their combined impact on the model's prediction. This diagram uncovers the mechanisms of feature combinations, including synergistic effects and nonlinear interaction pathways. This step helps understand the comprehensive influence of components in complex systems and identify which component combinations have significant joint effects on the model's prediction.

[0203] Suppose we have two features A and B with SHAP values of 0.5 and 0.3, respectively. By examining the component impact relationship graph, we can see the interaction between features A and B. If the graph shows that the combination of A and B has a significant positive contribution to the model's prediction, it indicates a synergistic effect between them. Additionally, if the graph shows that this contribution changes with the values of A and B, it indicates a nonlinear interaction path between them.

[0204] 4. Analyze the mechanism of key components:

[0205] By combining the force interpretation graph and the component impact relationship graph, we can analyze the contribution direction, synergistic effect, and nonlinear interaction path of key components in detail. By analyzing the force interpretation graph, we can determine the positive or negative contribution of each key component. By analyzing the component impact relationship graph, we can identify which components have synergistic effects and their nonlinear interaction paths. This step is crucial for understanding the model's prediction and can reveal complex relationships hidden in the data.

[0206] Suppose we find that the combination of features A and B has a significant positive contribution to the model's prediction, but this contribution is smaller when the values of A and B are low and significantly increases when the values of A and B are high. This indicates a nonlinear synergistic effect between A and B. By further analysis, we can determine the specific form of this nonlinear relationship, such as whether it is a multiplicative relationship or a more complex functional relationship.

[0207] 5. Mark core features that conflict with traditional cognition:

[0208] By comparing the results of the force interpretation graph and the component impact relationship graph with traditional design rules, we can mark features that do not conform to traditional experience and theory. These features may need further verification, as they may reveal new physical and chemical mechanisms or potential biases in the model. This step provides important clues for subsequent experimental verification and model optimization, helping to establish a solid evidence base for mechanism analysis.

[0209] Suppose traditional design rules indicate that feature A has a positive contribution to the model's prediction, but the force interpretation graph shows that feature A has a significant negative contribution. This conflict indicates that traditional cognition may need to be updated. We mark feature A and conduct further experimental verification to confirm its actual role. If the experimental verification results are consistent with the model's prediction, it will provide important evidence for updating traditional design rules.

[0210] Step S2F4, combined with the theory of phase change materials, interprets the identified conflict feature combination, clarifies the correlation logic between component synergistic effect and macroscopic performance breakthrough, and completes the theoretical construction of mechanism analysis.

[0211] wherein the phase change material field theory: involves the physical and chemical principles of phase change materials, including the theoretical basis of performance indicators such as phase change temperature, latent heat, supercooling degree, and viscosity. Conflict feature combination: in the optimization process, those feature combinations that do not conform to traditional experience and theory, such as the synergistic effect of certain components or the impact on performance indicators. Component synergy: the interaction between multiple components affects performance indicators, and this interaction may not be simply linear superposition. Macro performance breakthrough: performance improvement achieved through component synergy, which may exceed the expectations of traditional formulations.

[0212] The necessary process is described as follows: 1. Combine the phase change material field theory: refer to the basic theory of the phase change material field to analyze the physical and chemical basis of counterintuitive features. For example, consider the phase change temperature, latent heat value of the base hydrated salt, and the action mechanism of the temperature regulator, nucleating agent, and rheological modifier. 2. Mechanistic interpretation of conflict feature combination: detailed explanation of the physical and chemical principles of conflict feature combination. For example, if the synergistic effect of a certain component is significantly higher than that of traditional formulations, analyze its action mechanism in the phase change process, such as nucleation, phase change temperature regulation, etc. Hypothesis model analysis shows that the synergistic effect of the base hydrated salt and a certain temperature regulator significantly reduces the supercooling degree at a certain ratio. Through mechanistic interpretation, analyze the specific mechanism of this synergistic effect, such as how the temperature regulator effectively regulates the supercooling degree by changing the crystalline structure or intermolecular interactions of the hydrated salt. This explanation is based not only on experimental data but also on the physical and chemical theory of phase change materials. 3. Clear the correlation logic between component synergy and macro performance breakthrough: explain how component synergy leads to a breakthrough in macro performance. For example, the synergistic effect of the base hydrated salt and the temperature regulator may significantly reduce the supercooling degree at a certain ratio, thereby improving the practical application performance of the material. 4. Complete the theoretical construction of mechanism analysis: based on the above analysis, construct a complete theoretical framework of mechanism analysis. Integrate the correlation logic of conflict feature combination, component synergy, and macro performance breakthrough into a unified theoretical system, providing theoretical support for subsequent experimental verification and model optimization.

[0213] Step S2F5, based on the historical verification database and the preset verification rules, provides empirical support for mechanism interpretation, marks the consistency degree of the verification results and theoretical interpretation, and forms the verification evidence part.

[0214] wherein the historical verification database: contains past experimental data and verification results, used for comparison and verification of current mechanism interpretation. Preset verification rules: pre-set experimental or computational methods for verifying the correctness and reliability of mechanism interpretation. Empirical support: data obtained through experiments or calculations to support the correctness of mechanism interpretation. Consistency degree: the matching degree between the verification results and the theoretical interpretation, usually represented by a value between 0 and 1, the higher the value, the better the consistency.

[0215] The general process is described as follows: 1. Selecting preset verification means based on historical verification database: Referring to the historical verification database, select the preset verification means related to the current machine interpretation. These means can include experimental tests, computational simulations, etc., to ensure the reliability and effectiveness of the verification results. Suppose the effect of component synergy on supercooling needs to be verified, refer to similar verification means in the historical database, select differential scanning calorimetry (DSC) for experimental testing, or use molecular dynamics simulation for computational verification. 2. Implement verification experiments or calculations: According to the preset verification means, implement specific experiments or calculations. Ensure that the conditions of the experiment or calculation are consistent with the assumptions of the theoretical interpretation to ensure the comparability of the verification results. If DSC experiments are selected, follow the standard operating procedures for sample preparation and testing, and record the measured value of supercooling. 3. Calculate the relative error: Use the relative error to quantitatively judge the difference between the measured value and the theoretical expected value, the formula is as follows (only this necessary formula, meaning is popular): Relative error = (|measured value-theoretical expected value| ÷ theoretical expected value) x 100%. Formula meaning: First calculate the "absolute difference" between the measured and expected values, then see what percentage this difference accounts for of the theoretical expected value, the smaller the percentage, the more consistent; Example: the measured value of supercooling is 1.87℃, the theoretical expectation is 1.8℃, then the relative error = (|1.87-1.8| ÷ 1.8) x 100% ≈ 3.9%. 4. Mark the consistency level: mark the consistency level according to the industry commonly used standard, and give the conclusion combined with the test details: the level is divided: relative error ≤5% is "highly consistent", 5%-10% is "moderately consistent", 10%-20% is "basically consistent", more than 20% is "needs to be optimized". 5. Form the verification evidence part: combine the verification results and the consistency level to form the verification evidence part. Record the details of the verification experiment or calculation in detail, including experimental conditions, testing methods, calculation models, etc., to provide comprehensive empirical support for machine interpretation.

[0216] Step S2F6, based on historical application scenario data and material performance-application mapping rules, automatically generates application scenario suggestions and use restriction conditions for the counter-intuitive formula, outputs application guidance and risk warning content.

[0217] The specific process is as follows: 1. Select mapping rules based on historical application scenario data: Referring to historical application scenario data, select material performance-application mapping rules related to counterintuitive formulations. These rules help match material performance with actual application scenarios. Technical points: Use data mining techniques to extract mapping rules related to counterintuitive formulation performance indicators from historical application scenario data. For example, through clustering analysis or decision tree algorithms, identify which performance indicators are highly correlated with which application scenarios. 2. Generate application scenario suggestions: Automatically generate application scenario suggestions for this counterintuitive formulation based on mapping rules. Application scenario suggestions include recommended usage areas and specific application methods. Technical points: Use machine learning models (such as support vector machines, random forests, etc.) to predict the most suitable application scenarios for materials based on their performance. For example, if a formulation has a high latent heat value and low supercooling degree, the model may recommend it for solar thermal storage systems. 3. Determine usage restrictions: Based on material performance and application scenarios, determine usage restrictions. These conditions ensure the safety and effectiveness of materials in specific applications. Technical points: Determine usage restrictions by combining material performance-application mapping rules with the physical and chemical properties of materials. For example, if a material has high viscosity, it may be necessary to ensure that the system has good stirring or flow mechanisms during use. 4. Output application guidance and risk warning: Combine application scenario suggestions and usage restrictions to output application guidance and risk warning content. Application guidance provides detailed usage suggestions, and risk warning points out potential risks and precautions. Technical points: Generate detailed user manuals and risk warning reports, including operation steps, maintenance suggestions, and measures to address potential risks. For example, for high-viscosity materials, it is recommended to regularly check the system flow state to prevent blockage.

[0218] Step S2F7, integrate the core content of the phenomenon description, the theoretical construction of mechanism analysis, the verification evidence part, the application suggestion and the risk warning content, and generate a complete counterintuitive formulation explanation report according to the preset structured report template. Generate a complete explanation report: Integrate the above parts according to the preset structured report template to generate a complete counterintuitive formulation explanation report. Technical points: Use document generation tools (such as LaTeX or Word templates) to ensure uniform report format and complete content.

[0219] Large-scale preparation using continuous processes matching the final formulation includes:

[0220] Step S710, analyze the component characteristics of the final formulation, including the melting point parameters of the base hydrated salt, the dispersion characteristics of the additives, the viscosity control range of the rheological modifier, and the thermal stability interval of each component.

[0221] Take the final formula of "air conditioner cold energy storage phase change material" (sodium sulfate 60% + ammonium chloride 10% + nano silicon dioxide 3% + diatomite 7% + water 20%) as an example, analyze it according to the following logic:

[0222] Extract basic parameters: retrieve the inherent properties of each substance from the formula component database (such as the melting point of sodium sulfate decahydrate 32.4℃, the thermal decomposition temperature of ammonium chloride 337.8℃).

[0223] Combine formula proportion analysis: for the actual proportion, determine the viscosity control range of diatomite 7% (5500-6500 mPa・s) and the dispersion requirement of nano silicon dioxide 3% (ultrasonic assisted dispersion is required).

[0224] Draw the process adaptation boundary: integrate the characteristics of each component, and determine that the heating temperature of the mixing module needs to be higher than the melting point of sodium sulfate (≥35℃) and lower than the decomposition temperature of ammonium chloride (≤300℃), which provides parameter basis for subsequent process module matching.

[0225] Step S720, based on the component characteristics, match the feeding module, mixing module and forming module of the continuous process to adapt to the processing requirements of different components.

[0226] Take the formula of "air conditioner cold energy storage phase change material" (sodium sulfate 60% + ammonium chloride 10% + nano silicon dioxide 3% + diatomite 7% + water 20%) as an example: 1. Feeding module matching: according to the easy agglomeration characteristics of nano silicon dioxide, anti-agglomeration quantitative feeder is matched; according to the easy dissolution characteristics of ammonium chloride, liquid / solid double feeding port is matched, respectively feeding solid components (sodium sulfate, diatomite) and liquid components (ammonium chloride aqueous solution); 2. Mixing module matching: combined with the melting point of sodium sulfate (≥35℃), temperature controlled stirring tank is matched; according to the dispersion requirement of nano silicon dioxide, ultrasonic dispersion assembly is integrated in the tank; 3. Forming module matching: according to the viscosity (5500-6500 mPa・s) regulated by diatomite, continuous forming nozzle suitable for spray construction is matched.

[0227] Step S730, set the preset process parameters for each matched process module;

[0228] Preset process parameters: for the feeding, mixing and forming modules of the continuous process, the quantifiable operation indexes such as feeding rate, heating temperature and stirring speed need to be matched with the formula component characteristics and module functions. Module adaptation parameters: parameters matched with the functions of specific process modules (such as anti-agglomeration feeder, temperature controlled mixing tank), such as the vibration frequency of feeding module and the ultrasonic power of mixing module. Parameter constraint boundary: based on the component characteristics (such as thermal stability, viscosity), set the upper and lower limits of parameters, such as the mixing temperature does not exceed the decomposition temperature of ammonium chloride (337.8℃) and does not lower than the melting point of sodium sulfate (32.4℃).

[0229] The necessary process is described as follows: 1. Derivation based on component characteristics: referring to the component characteristics (such as the melting point of sodium sulfate, the dispersion requirement of nanosilica) analyzed in S710, the parameter range is preliminarily determined, for example, the mixing temperature needs to be higher than the melting point of sodium sulfate to ensure melting. 2. Combining with the rated function of the module (such as the anti-agglomeration feeder, the high-pressure spraying unit): according to the matched module in S720, the feasible value of the parameter is determined, for example, the ultrasonic dispersion power does not exceed the rated power of the module (500W). 3. Verification through pre-experiment: small-batch test of the influence of different parameter combinations on the performance of the sample, and screening of the parameters (such as stirring speed 80 rpm) that can stably meet the performance requirements (such as the mixing uniformity of the system reaching more than 95%).

[0230] Step S740, starting the continuous production line according to the preset process parameters to carry out trial production, the trial production yield is a preset trial production proportion of the total scale production, and the trial production sample is collected to detect the key performance.

[0231] The preset process parameters refer to the quantifiable operation indexes set for the feeding, mixing and molding modules in S730, such as the solid feeding rate (sodium sulfate 50 kg / h), the mixing temperature (40°C), the spraying pressure (1.0 MPa) and the like, which are the operation basis for trial production. The trial production yield is the total amount of products in the trial production stage, which is calculated by “total scale production x preset trial production proportion”, and is used to verify the process stability. The preset trial production proportion is the percentage of the trial production yield in the total scale production (industry convention is 5%-10%), which balances the trial production cost and the effectiveness of process verification. The key performance detection items are detection indexes for the core application requirements of phase change materials, such as phase change temperature, phase change latent heat, supercooling degree and melting viscosity (corresponding to the scene demand of “air conditioner cold energy storage” in the present application).

[0232] Step S750, calculating the deviation value of the key performance of the trial production sample and the laboratory sample, if the deviation value ≤ preset deviation threshold, carrying out scale production according to the trial production parameters; if the deviation value > preset deviation threshold, adjusting the preset process parameters of the corresponding module based on the deviation analysis result, and repeating the trial production step until the deviation value ≤ preset deviation threshold or the preset maximum iteration number is reached.

[0233] The approximate process is described as follows:

[0234] 1. Deviation value calculation: absolute error (suitable for temperature, supercooling degree): deviation value = |trial production sample performance value - laboratory sample performance value|; relative error (suitable for latent heat, viscosity): deviation value = (|trial production sample performance value - laboratory sample performance value| ÷ laboratory sample performance value) x 100%.

[0235] 2. Pre-set deviation threshold determination: Refer to the inorganic hydrated salt phase change material industry standard (such as GB / T35985-2018) and the "air conditioner cold energy storage" scene requirements of the present application, set the key indicator threshold (such as phase change temperature ≤0.5℃, latent heat ≤5%, supercooling degree ≤0.3℃, viscosity ≤10%).

[0236] 3. Deviation cause positioning: Positioning through the "performance-module" correspondence (such as viscosity deviation → mixing module stirring speed / temperature; phase change temperature deviation → feeding module component proportion / mixing temperature).

[0237] 4. Adjusting process parameters based on deviation analysis results: If the deviation value exceeds the pre-set deviation threshold, adjust the pre-set process parameters of the corresponding module according to the deviation cause positioning results. For example: if the viscosity deviation exceeds the threshold, adjust the stirring speed or temperature setting of the mixing module. If the phase change temperature deviation exceeds the threshold, adjust the component proportion or mixing temperature setting of the feeding module.

[0238] 5. Repeat the trial production step: After adjusting the process parameters, re-produce the trial production, calculate the new deviation value. If the new deviation value still exceeds the pre-set deviation threshold, repeat the above adjustment and trial production steps until the deviation value ≤ pre-set deviation threshold or reach the pre-set maximum iteration number (such as 3 times).

[0239] Step S760, in the process of large-scale production, real-time monitoring of pre-set key process parameters, each batch according to the pre-set sampling frequency to detect product performance stability, output the large-scale product that meets the multi-dimensional performance demand index.

[0240] Pre-set key process parameters: directly use the parameters verified by trial production in S730, no need to adjust, ensure the consistency of large-scale production and trial production process. Real-time monitoring means: special sensors are installed for each module (weighing sensor for feeding module, temperature sensor for mixing tank, pressure sensor for forming module), data is transmitted to the central control system in real time, and automatic alarm is given when the threshold is exceeded. Pre-set sampling frequency: refer to the production characteristics of inorganic hydrated salt phase change material (easy to cause performance fluctuation due to uneven mixing), set "sample once every 2 hours" or "sample once every 10 tons of production", covering the whole batch production cycle. Quality standard: integrate S100 user requirements (such as "phase change temperature 10℃ or so, latent heat ≥180J / g") and industry specifications, improve 5%-10% precision as the qualified line (such as latent heat ≥190J / g), ensure the application reliability.

[0241] The embodiments of the specific implementation are the preferred embodiments of the present application, and do not limit the protection scope of the present application, so: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. An AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials, characterized in that, include: To obtain the multi-dimensional performance requirements of inorganic hydrated salt phase change materials for the target application scenarios, the multi-dimensional performance requirements include the phase change temperature range, the minimum latent heat of phase change, the maximum allowable subcooling, the viscosity range suitable for construction processes, the upper limit of the target cost, and the minimum cycle life. The trained Gaussian process regression model is used to analyze the performance requirement index, generate candidate formulations, and output the performance prediction results of each candidate formulation. Based on the candidate formulations and performance prediction results, with the constraints that the predicted values ​​of each performance index meet the corresponding preset demand thresholds and the predicted formulation cost is less than or equal to the target cost upper limit, a multi-objective optimization algorithm is used to select the Pareto optimal formulation. Samples were prepared and performance was measured for the Pareto optimal formulation. The deviation values ​​of the measured data and the performance prediction results on each performance index were calculated and compared with the preset threshold. If the deviation values ​​of all performance indicators do not exceed the preset threshold, then the formula is taken as the target formula. If the deviation of any performance index exceeds the preset threshold, the measured data will be added to the training dataset, the Gaussian process regression model will be retrained, the Pareto optimal formulation will be screened again, and the sample preparation, performance measurement and deviation comparison steps will be repeated until a target formulation in which the deviation of one or more performance indices meets the preset threshold requirements is obtained. Based on a preset comprehensive performance evaluation function, the formulation with the highest evaluation score is selected from all the obtained target formulations as the final formulation, and large-scale preparation is completed using a continuous process that matches the final formulation.

2. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 1, characterized in that, The training process of the Gaussian process regression model is as follows: The composition of a multi-component system containing matrix hydrated salts and additives was determined. The additives included temperature regulators, nucleating agents and rheology modifiers. The mass ratio range of each component was determined based on phase change functional requirements and preliminary experimental data. The initial formulation was obtained by sampling the proportions of the multi-component system using a pre-designed experimental method. After the rationality was verified and invalid formulations were removed, the valid sampled formulations were retained. Samples were prepared according to the valid sampled formulations and the phase transition temperature, latent heat of phase transition, supercooling, viscosity and cycle life data were measured. The initial dataset was constructed by combining the component mass ratios and corresponding performance data of similar systems in historical literature. Using the mass percentage of each component in the initial dataset as input features, and the corresponding phase transition temperature, latent heat of phase transition, supercooling, viscosity, and cycle life as output labels, the data is standardized to construct a training dataset. Based on the training dataset, Bayesian inference is used to iteratively optimize the model parameters to fit the mapping relationship between components and performance, thus forming a formulation-performance mapping model. The prediction error of the recipe-performance mapping model is verified using a pre-defined cross-validation method. If the prediction error of each performance index does not exceed the pre-defined threshold, it is directly identified as a trained Gaussian process regression model. If the prediction error exceeds the pre-defined threshold, for the error exceeding the threshold, the model parameter optimization steps are repeated by adding the sampled and measured data to the training dataset according to the aforementioned pre-defined experimental design method until the model prediction error meets the pre-defined threshold, and then it is identified as a trained Gaussian process regression model.

3. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 2, characterized in that, The initial formulation was obtained by sampling a small number of components in a multi-component system using a pre-designed experimental method. This included: Determine the mass percentage variables and preset value ranges of matrix hydrated salts, temperature regulators, nucleating agents, and rheology modifiers in a multi-component system, and constrain the total mass percentage of each component to 100%; Based on the determined mass proportions and ranges of each component, a central composite design is adopted as the pre-set experimental design method. An experimental point set containing vertices, pivots, and center points is generated within the formulation space defined by the above variables with a pre-set small sample size. The structure of the experimental point set constitutes a second-order experimental design for analyzing linear and quadratic interaction effects. The combination of proportion values ​​corresponding to each experimental point is the initial formulation obtained by small sample sampling.

4. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 2, characterized in that, The initial formulation was obtained by sampling a small number of components in a multi-component system using a pre-designed experimental method. This included: Step 1: Use the preset criteria for determining the highly nonlinear response of a multi-component system to determine the response characteristics of the multi-component system. Step 2: If the system is determined to be a highly nonlinear response system, the basic ratio values ​​and corresponding performance data of the matrix hydrated salt and additives are collected according to the preset sample size to construct an initial sample set. Step 3: Based on the initial sample set, construct a nonlinear surrogate model through Gaussian process regression, with the ratio value as input and performance data as output, outputting the performance prediction value and prediction variance corresponding to the ratio value; Step 4: Call the preset expected improvement function, substitute the performance prediction value and prediction variance into the calculation of the optimization value, and define the region where the optimization value is greater than the preset value threshold and the prediction variance is greater than the preset variance threshold as the high information sampling region. Step 5: Select the ratio value in the high information sampling area according to the preset encryption ratio for experimental sampling, measure the corresponding performance data, and supplement the newly collected ratio value and performance data to the initial sample set to update the sample set. Step 6: Iteratively optimize the nonlinear surrogate model based on the updated sample set, calculate the decrease in prediction error of the nonlinear surrogate model on the preset validation set before and after iteration, and repeat steps 3 to 5 until the preset number of iterations or the decrease in error is less than the preset decrease threshold is reached. Step 7: Use the final set of ratio values ​​as the initial formula obtained from small sample sampling.

5. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 4, characterized in that, The trained Gaussian process regression model is used to analyze performance requirement indicators, generate candidate formulations, and output the performance prediction results for each candidate formulation, including: The Gaussian process regression model is trained by inputting preset multi-dimensional performance requirement indicators as constraints. The model search generates a set of candidate formulations that meet the constraints and outputs the component proportions, performance prediction values ​​and prediction variances of each formulation. Call the preset expected improvement function, substitute the performance prediction value and prediction variance to calculate the initial optimization value of each candidate formulation, and form a preliminary ranking; The SHAP interpretability tool is called to calculate the Shapley value of each component to the performance index for each candidate formulation and take the absolute value to obtain the contribution. Components with a contribution greater than the preset contribution threshold are defined as the key components of the formulation. Based on the key components and their contribution of each candidate formulation, the initial optimization value is corrected by a preset adjustment coefficient; Based on the corrected optimization value, the candidate recipe set is reordered, and recipes whose ranking improvement is greater than the preset improvement threshold are marked as counterintuitive recipes. For counterintuitive formulas, a pre-defined generation method is used to generate an explanatory report; Output the final candidate formulation list and accompanying explanatory report.

6. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 5, characterized in that, Based on the key components and their contributions of each candidate formulation, the initial optimization value is corrected using preset adjustment coefficients, including: Acquire the key components and their contribution data for each candidate formulation, simultaneously extract the performance prediction values, prediction variances and initial optimization values ​​of the candidate formulations, and integrate them to form a dataset of parameters to be corrected. Based on a knowledge base in the field of phase change materials, preset adjustment coefficients are set, and priority weights of mechanisms are assigned according to the control priority of key components for the core performance requirements of the formulation. Based on the performance prediction values, prediction variances, and key component contributions in the dataset of parameters to be corrected, a composite adjustment factor that integrates the key component contribution and mechanism priority weights is constructed. This factor is then introduced into the optimization value correction calculation, and the correction formula is derived. Based on the correction formula and the initial optimization value in the dataset of parameters to be corrected, a hierarchical correction calculation is performed according to the contribution of key components to update the optimization value of each candidate formulation. A mechanism consistency verification method based on historical experimental data comparison was adopted to verify the matching degree between the optimized value after correction and the actual role of key components. If the matching degree is greater than or equal to the preset threshold, the corrected optimization value, the adjusted weight ratio and the verification result will be directly output and fed back to the candidate formula re-ranking step. If the matching degree is less than the preset threshold, the preset adjustment coefficient and mechanism priority weight are dynamically adjusted according to the deviation ratio. Based on the dataset of parameters to be corrected, the composite adjustment factor construction, hierarchical correction calculation and matching degree verification steps are re-executed until the matching degree is greater than or equal to the preset threshold or the preset maximum number of iterations is reached. Then, the results are fed back to the candidate formulation reordering step.

7. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 5, characterized in that, Generating an explanatory report for counterintuitive formulas using pre-defined generation methods includes the following steps: Obtain data on the component proportions, performance predictions, key component contributions, and ranking changes of counterintuitive formulations, and simultaneously extract historical data of similar traditional formulations as a benchmark. By pre-setting comparison parameters, the differences between counterintuitive formulas and traditional formula design rules are quantified to form an intuitive comparison table, which constitutes the core content of the phenomenon description. The SHAP visualization tool was used to generate SHAP force interpretation diagrams and component influence relationship diagrams, analyze the contribution direction, synergistic effect and nonlinear action path of key components, identify characteristic combinations that significantly conflict with traditional design rules, and establish the evidence basis for mechanism analysis. Combining theory in the field of phase change materials, we provide a mechanistic explanation for the identified combinations of conflicting features, clarify the correlation between the synergistic effect of components and the breakthrough in macroscopic performance, and complete the theoretical construction of the mechanism analysis. Based on historical verification databases and preset verification rules, empirical support is provided for the explanation of the mechanism, and the degree of consistency between the verification results and the theoretical explanation is marked to form the verification evidence section. Based on historical application scenario data and material property-application mapping rules, the system automatically generates application scenario suggestions and usage restrictions for this counterintuitive formulation, producing application guidance and risk warning content. The report integrates the core content of the phenomenon description, the theoretical construction of the mechanism analysis, the verification evidence, the application suggestions and risk warnings, and generates a complete counterintuitive formula explanation report according to the preset structured report template.

8. The AI-driven intelligent design and preparation method for inorganic hydrated salt phase change materials according to claim 1, characterized in that, Large-scale preparation using a continuous process matched to the final formulation includes: The component characteristics of the final formulation were analyzed, including the melting point parameters of the matrix hydrated salt, the dispersion characteristics of the additives, the viscosity control range of the rheology modifier, and the thermal stability range of each component. The feeding module, mixing module, and molding module of the continuous process are matched based on the characteristics of the component to adapt to the processing requirements of different components. Set preset process parameters for each matched process module; The continuous production line was started according to the preset process parameters for trial production. The trial production output was a preset trial production ratio of the total scale production output. Trial production samples were collected to test key performance. Calculate the deviation between the key performance of the trial production sample and the laboratory sample. If the deviation is less than or equal to the preset deviation threshold, then proceed with large-scale production according to the trial production parameters. If the deviation is greater than the preset deviation threshold, adjust the preset process parameters of the corresponding module based on the deviation analysis results and repeat the trial production steps until the deviation is less than or equal to the preset deviation threshold or the preset maximum number of iterations is reached. During large-scale production, preset key process parameters are monitored in real time, and the performance stability of each batch is tested at a preset sampling frequency to output large-scale products that meet multi-dimensional performance requirements.

Images