Crystal growth prediction method based on GBDT and its application

Abstract

The present invention discloses a method for predicting crystal growth based on GBDT and its application, the method comprising: obtaining historical data of crystal growth test, determining three influencing factors of temperature, crystal height and solution saturation point temperature that affect crystal growth, constructing a data set, wherein the crystal is any one of potassium dideuterium phosphate, potassium dihydrogen phosphate, ammonium dihydrogen phosphate and dideuterium ammonium phosphate crystals; after pretreatment, taking the mean square error of crystal growth prediction as the objective function, constructing a gradient boosting decision tree model, and training the gradient boosting decision tree model with the pretreated training set, using Bayesian optimization model hyperparameters until the model prediction reaches the optimum; obtaining the influencing factor data of the current crystal growth test, and predicting the crystal growth with the trained gradient boosting decision tree model. The present invention uses the gradient boosting decision tree model to predict the crystal growth value, and realizes real-time monitoring and scientific regulation of the crystal growth process.

Description

Technical Field

[0001] The present invention relates to the technical field of crystal growth prediction and growth optimization, and in particular to a GBDT-based crystal growth prediction method and application thereof. Background Art

[0002] Crystals such as DKDP (dideuterium potassium phosphate) and KDP (monopotassium phosphate) are important nonlinear optical crystal materials that can be used to make electro-optical switches and frequency conversion devices for solid-state lasers with high power levels and good beam quality. They are widely used in the field of high-power lasers. However, when such crystals are prepared using traditional methods or single-slot circulating flow methods, the growth cycle of the crystals, especially large-diameter crystals, is relatively long, generally requiring 2-3 years, and the defect rate is high. Typical defects such as inclusions, growth streaks, and cracks are very likely to appear during the growth process, seriously affecting the crystal quality and the final yield. Therefore, quickly and accurately predicting the growth state of the crystal, monitoring and controlling its growth process in real time, and achieving efficient and rapid growth of high-quality crystals have always been key scientific issues and technical difficulties in the current field.

[0003] At present, the prediction and control of the crystal growth process mainly rely on the physical and chemical model based on the summary of expert experience. However, this method needs to consider multiple coupled physical processes such as temperature field, flow field, concentration field, etc. The modeling is complex, the amount of calculation is large, the probability of deviation between the predicted results and the actual results is high, and it does not have the ability of autonomous learning and iterative optimization, which makes it difficult to meet the needs of industrial mass production. In contrast, the data-driven modeling method can directly establish the mapping relationship between data and results "end to end" without in-depth understanding of the physical mechanism. Through continuous learning and continuous optimization, it can better meet the fast and flexible needs of industrial production. Therefore, compared with physical modeling, it is an important direction of current crystal growth research to study new data-driven methods, use experimental data training models to quickly and accurately predict the growth state of crystals, and realize real-time monitoring and scientific control of the growth process.

[0004] However, existing solutions that use machine learning algorithms to build prediction models for crystal growth prediction usually use conventional neural networks to build prediction models. When this method is applied specifically, on the one hand, it is necessary to consider the selection of features that affect crystal growth, and on the other hand, it is also necessary to consider the compatibility of the network model with crystal growth prediction. Existing methods make it difficult to achieve efficient and accurate crystal growth predictions. Summary of the invention

[0005] To address the deficiencies of the above-mentioned prior art, the present invention provides a crystal growth prediction method based on GBDT and its application, constructs a gradient boosting decision tree (GBDT) model based on Bayesian optimization training, and trains the model using historical data of crystal growth experiments to achieve fast and accurate crystal growth prediction, and compares the predicted value of crystal growth with the actual value, analyzes the crystal growth state according to the comparison result, and realizes real-time monitoring and scientific regulation of the crystal growth process, thereby guiding the high-quality growth of crystals, improving the yield rate, and reducing production costs.

[0006] In a first aspect, the present invention provides a method for predicting crystal growth based on GBDT, comprising:

[0007] Obtain historical data of crystal growth tests in the test device, determine the influencing factors affecting crystal growth, and construct a data set with daily influencing factor data and actual crystal growth data; the influencing factors include crystal growth environment temperature, crystal height and solution saturation point temperature; the crystal is any one of dideuterium potassium phosphate DKDP, dipotassium dihydrogen phosphate KDP, diammonium dihydrogen phosphate ADP, and dideuterium ammonium phosphate DADP crystals;

[0008] Preprocessing the data in the data set; the preprocessing includes removing outliers, extracting data features and normalizing the data;

[0009] Taking minimizing the mean square error of crystal growth prediction as the objective function, a gradient boosting decision tree model was constructed, and the gradient boosting decision tree model was trained using the preprocessed training set. The Bayesian optimization model hyperparameters were used until the model prediction reached the optimal value.

[0010] Obtain the influencing factor data of the current crystal growth experiment in the same experimental device, and use the trained gradient boosting decision tree model to predict the crystal growth amount.

[0011] Further technical solutions include:

[0012] The predicted value of crystal growth for the day is compared with the actual value of growth, and the experimental environment for controlling crystal growth is adjusted according to the comparison result.

[0013] In a second aspect, the present invention provides an application of a GBDT-based crystal growth prediction method in the cultivation of DKDP crystals and related crystals, wherein the related crystals include potassium dihydrogen phosphate KDP, ammonium dihydrogen phosphate ADP, and dideuterium ammonium phosphate DADP crystals.

[0014] In a third aspect, the present invention further provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method described in the first aspect are completed.

[0015] In a fourth aspect, the present invention further provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the method described in the first aspect.

[0016] One or more of the above technical solutions have the following beneficial effects:

[0017] 1. The present invention provides a crystal growth prediction method based on GBDT and its application, constructs a gradient boosting decision tree model based on Bayesian optimization training, and trains the model using historical data of crystal growth experiments to achieve fast and accurate crystal growth prediction, and compares the predicted value of crystal growth with the actual value, analyzes the crystal growth state according to the comparison result, and realizes real-time monitoring and scientific regulation of the crystal growth process, thereby guiding the high-quality growth of crystals, improving the yield rate, and reducing production costs.

[0018] 2. The crystal growth prediction method proposed in the present invention can be applied to the growth prediction of potassium dideuterium phosphate DKDP and its related crystals such as potassium dihydrogen phosphate KDP, ammonium dihydrogen phosphate ADP, dideuterium ammonium phosphate DADP crystals, etc. Deuterium and hydrogen are isotopes, and the above crystals have the same chemical properties as DKDP crystals. Therefore, the proposed prediction method can be applied to the growth prediction of DKDP and its related crystals, and has a wider scope of application.

[0019] 3. Compared with the traditional physical model method, the data-driven machine learning model constructed by the present invention does not require understanding of the physical mechanism of crystal growth. Through continuous optimization, it can continuously improve the prediction performance and accuracy, significantly reduce the growth defect rate of DKDP crystals, and realize the rapid and high-quality industrial production of DKDP crystals. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0021] Figure 1 This is a flow chart of the GBDT-based crystal growth prediction method according to an embodiment of the present invention;

[0022] Figure 2 A curve diagram of the DKDP crystal growth prediction error in an embodiment of the present invention;

[0023] Figure 3It is a schematic diagram of a programmed model in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] It should be noted that the following detailed descriptions are exemplary only, are intended to describe specific embodiments, are intended to provide further explanation of the present invention, and are not intended to limit exemplary embodiments according to the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art to which the present invention belongs. In addition, it should also be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0025] Embodiment 1

[0026] In view of the problem that the current DKDP crystal and its related crystals (such as potassium dihydrogen phosphate KDP, ammonium dihydrogen phosphate ADP, dideuterium ammonium phosphate DADP crystals, etc.), especially large-diameter crystals, have a long growth cycle and are prone to defects such as cracks and dense clouds, which lead to scrapping, this embodiment provides a crystal growth prediction method based on GBDT, which can achieve rapid and accurate prediction of crystal growth and provide the possibility for optimizing and controlling the crystal growth process. The method proposed in this embodiment is as follows Figure 1 As shown, the specific steps include:

[0027] Obtain historical data of crystal growth tests in the test device, determine the influencing factors affecting crystal growth, and construct a data set with daily influencing factor data and actual crystal growth data; the influencing factors include crystal growth environment temperature, crystal height and solution saturation point temperature; the crystal is any one of dideuterium potassium phosphate DKDP, dipotassium dihydrogen phosphate KDP, diammonium dihydrogen phosphate ADP, and dideuterium ammonium phosphate DADP crystals;

[0028] Preprocessing the data in the data set; the preprocessing includes removing outliers, extracting data features and normalizing the data;

[0029] Taking minimizing the mean square error of crystal growth prediction as the objective function, a gradient boosting decision tree model was constructed, and the gradient boosting decision tree model was trained using the preprocessed training set. The Bayesian optimization model hyperparameters were used until the model prediction reached the optimal value.

[0030] Obtain the influencing factor data of the current crystal growth experiment in the same experimental device, and use the trained gradient boosting decision tree model to predict the crystal growth amount.

[0031] In this embodiment, the gradient boosting decision tree algorithm is used for crystal growth prediction. The gradient boosting decision tree (GBDT) is an integrated learning method that learns the conditional probability distribution of the target variable by constructing multiple decision trees and gradually reducing the training error of the loss function. It can handle nonlinear relationships and is very suitable for simulating the complex process of crystal growth. In the built gradient boosting decision tree model, each tree focuses on fitting the residual of the current model. This idea of ​​incremental learning enables the model to continuously optimize and improve the prediction accuracy; the model contains a large number of adjustable optimization parameters, such as the number of trees, the maximum depth, the minimum number of leaf node samples, etc., which also provides the possibility for optimizing the model; in addition, the algorithm can realize feature selection, automatically screen the input variables that have the greatest impact on the target variable, and ensure the accuracy of modeling; by controlling the depth of the tree and stopping in advance, the problem of overfitting can be avoided. That is, the GBDT used in this embodiment can handle complex nonlinear relationships, has strong adjustability, avoids overfitting, and is efficient in calculation. It is suitable for application in solution crystal growth process data modeling and prediction tasks.

[0032] Of course, there are certain problems in using the GBDT model for training and prediction, namely: GBDT is a complex model that contains a large number of parameters and takes a long time to train, making it difficult to adjust and optimize parameters; the model is difficult to explain directly and is not as simple and intuitive as a linear model, requiring certain data analysis work to understand the importance of variables; it is not good at processing high-dimensional sparse features, and the effect will be reduced when the feature dimension is extremely high.

[0033] To this end, when this embodiment uses the GBDT model for training and prediction, taking into account the problems of long model training time and many parameters, the Bayesian optimization algorithm is used to optimize the model training process, which can narrow down the results to certain points, and then find the optimal parameters, effectively shortening the time to complete the work; for problems that are difficult to explain, the results are compared with crystals that have completed growth tests to complete the explanation; for the problem that the model is not good at high-dimensional sparse features, only 3 feature data sets are selected. Based on these 3 features, most other factors affecting crystal growth and their relationships can be inferred, and the intervals between these data are small and not sparse, which can effectively guarantee the prediction effect of the GBDT model.

[0034] In this embodiment, taking the growth prediction of DKDP crystal as an example, the DKDP crystal growth amount prediction method based on GBDT proposed in this embodiment is introduced in more detail through the following content.

[0035] Step S1, collecting historical data of the DKDP crystal growth test in the test device, the historical data including daily crystal growth environment temperature, crystal height, solution saturation point temperature and daily actual crystal growth amount data.

[0036] In this embodiment, the crystal growth environment temperature, crystal height and solution saturation point temperature are selected as influencing factors, which are specifically determined by the following analysis:

[0037] Firstly, the factors influencing the growth of DKDP crystals are analyzed. The main influencing factors are the ambient temperature and the solution saturation temperature. The supersaturation temperature of the solution can be calculated by the solution saturation temperature and the ambient temperature, which is the most important factor to be considered in the thermodynamics of crystal growth.

[0038] Specifically, in the solution method crystal growth process, controlling the actual temperature in the growth tank to accurately reduce to the solution saturation point so that the solute precipitates is a key temperature control process, which affects the crystal growth dynamics and crystal quality. First, (1) the concepts of actual temperature and solution saturation point temperature are clarified. The solution saturation point temperature is the temperature corresponding to the solute completely dissolved in the solvent to reach the saturation concentration, and the actual temperature is the current growth environment temperature of the solution in the reactor. After the solution is prepared, it enters the reactor at a temperature higher than the saturation point, and the solute is in a dissolved state. By gradually lowering the growth tank temperature, when the two temperatures tend to be equal, the solution changes from a supersaturated state to a saturated state, and the solute begins to precipitate. Secondly, (2) whether large-sized single crystals can be formed is strongly related to the design of the cooling curve. If the temperature is lowered too quickly, the solution will quickly become strongly supersaturated, and a large number of spontaneously nucleated nuclei may be generated. These nuclei compete fiercely for space and are difficult to continue growing, forming fine polycrystals. At the same time, rapid cooling will cause temperature field disturbances, resulting in thermal stress and cracks. On the contrary, if the temperature is lowered too slowly, the solution will be in a weakly supersaturated state for a long time, the number of directional growth nuclei will be too small, the growth rate will be extremely slow, and large-sized single crystals cannot be obtained efficiently. In addition, overheating may also cause the solute to decompose. Therefore, it is crucial to carefully design the cooling curve to control the rate of change of supersaturation to obtain good products. In the process of regulating crystal growth, the stability of the solute needs to be considered and the appropriate cooling rate should be selected. Usually, a linear cooling mode is used, with a rate of 0.1-1℃ / h. The rate needs to be slowed down when approaching the saturation point to ensure stable growth.

[0039] In addition, (3) equipment parameters also need to be optimized and matched. The stirring method affects mixing and mass transfer. The temperature uniformity of different parts determines the location of solute precipitation. The heat conduction rate of the cooling medium will also affect the cooling process. For example, for DKDP solution, its saturation point is 25°C, the initial cooling rate is 0.6°C / h, and it is reduced to 0.2°C / h at 27°C. The final temperature is controlled at 25.5°C to obtain single crystal products with consistent size. Therefore, to accurately control the temperature change process of the solution from supersaturation to stable saturation, it is necessary to comprehensively consider the solute characteristics and equipment process parameters, and obtain a scientific and reasonable temperature control strategy through repeated experiments and optimization, which is directly related to the crystallization rate and final quality. Therefore, temperature is the main factor affecting the growth of DKDP crystals.

[0040] Secondly, regarding the influence of crystal height, when the height of crystals growing in a direction in a solution increases, it will significantly affect the flow state of the solution and the mass transfer process through various mechanisms, thereby inhibiting the growth rate of the crystals. First, (1) the increase in crystal height will compress the effective volume of the solution in the reactor. For example, in a reactor with a total volume of 5 liters, when the initial crystal height is 1 cm, the solution volume is about 4.9 liters; when the height reaches 2 cm, the solution volume decreases to 4.8 liters. According to the Bernoulli equation, under the same stirring conditions, the same flow rate will increase through a narrower pipe. If the above volume decreases by 10%, the speed will theoretically increase by 11%. Secondly, (2) as the mass of the crystal increases, the moment of inertia of the system also increases. The moment of inertia is related to the shape and mass distribution of the object. It is a physical quantity that represents rotational energy. Under the condition of constant rotation speed, the increase in inertia will reduce the rotation speed. For example, if the mass of the crystal increases from 10 grams to 15 grams, its moment of inertia increases by 50%. Then, under the same input torque, the rotation speed will decrease by about 20%, which will reduce the relative movement speed of the solution near the crystal surface and affect the solute transfer near the interface. In addition, (3) the increase in height will also change the flow field distribution state in the reactor, which may cause abnormal temperature field, resulting in uneven growth rates in different parts and product morphology distortion; it may also produce local low flow rate areas, making it easy for secondary nuclei to form. Usually, the flow rate is the lowest at the top and center, and secondary nuclei are most likely to appear. In addition, (4) as the size of the crystal increases, its resistance to the flow of the solution also increases. According to Darcy's law, the pressure difference on both sides of the pipeline is related to the fluid viscosity and the length and radius of the pipeline. The solution needs to bypass crystals with larger cross-sectional areas, which will result in higher pressure losses and reduce local and overall flow rates. Therefore, by inputting the parameters of crystal height, the problem of model training difficulties caused by the training process of many parameters can be reduced.

[0041] Step S2: preprocess the data in the original data set to check the data quality through preprocessing. The preprocessing includes removing outliers, interpolating missing values, extracting data features and normalizing data to obtain a high-quality model training data set after processing.

[0042] In this embodiment, a historical growth data set of DKDP crystals is collected. Historical data of DKDP crystal growth in the past 10 years are collected from a certain experimental device, including information such as temperature, crystal height, solution saturation point temperature, and actual daily growth. The original data set contains a total of about 15,000 samples with a time span of 1,000 days. The original data set is then preprocessed, including:

[0043] (1) Delete outliers: Calculate the Z-axis height of the daily growth, and remove abnormal data whose absolute value of the Z-axis height exceeds a threshold (3 mm is taken as the threshold in this embodiment). Specifically, the Z-axis height refers to the height difference obtained by subtracting the height measured at a fixed time point in the current morning (such as 8 o'clock) from the height at a fixed time point the previous day (i.e. 8 o'clock). Compare the height difference with the set threshold to determine whether it is abnormal data. Usually, according to laboratory experience, the daily growth should not exceed 1 mm. If it exceeds 1 mm, it may be caused by the error of the worker operating the ruler. In view of the fact that the gradient boosting decision tree has a certain tolerance for individual errors, 3 mm is selected as the threshold. Based on this threshold, the data that exceeds the threshold is selected to be deleted.

[0044] (2) Feature Engineering: The daily crystal growth environment temperature data is split into hourly levels to obtain the temperature change curve of the day, and the average temperature and temperature variance of the day are extracted as feature variables. At the same time, the crystal height and solution saturation point temperature are used as feature variables.

[0045] (3) Data normalization: Use the Min-Max method to normalize all feature variables to the interval [0, 1].

[0046] Step S3, taking minimizing the mean square error of the DKDP crystal growth prediction as the objective function, constructing a gradient boosting decision tree model, and using the preprocessed training set to train the gradient boosting decision tree model, using Bayesian optimization model hyperparameters until the model prediction reaches the optimal.

[0047] Specifically, the Gradient Boosting Decision Tree (GBDT) algorithm is a machine learning algorithm that improves the prediction ability of the model through continuous iteration. The core idea of ​​the algorithm is to combine multiple weak classifiers into a strong classifier to improve the accuracy of the model. Based on the preprocessed training data set, the established gradient boosting decision tree model is trained. The relevant characteristic variables of the three parameters of temperature, crystal height, and solution saturation point temperature are set, and the daily growth change is used as the target variable. The CART algorithm is used to recursively generate multiple regression trees to learn the complex nonlinear mapping relationship between the characteristic variables and the target variables.

[0048] Furthermore, in order to improve the prediction performance of the model, the Bayesian optimization algorithm is used to tune key hyperparameters such as the number of decision trees, the maximum depth of the tree, the minimum number of samples required for internal node re-segmentation, and the minimum number of samples for leaf nodes. The mean square error of the prediction results is defined as the objective function of Bayesian optimization, and the optimization variable range, number of iterations, sampling method, etc. are set. A set of approximately optimal parameter combinations is found through multiple iterative searches.

[0049] In this embodiment, the objective function is to minimize the mean square error of the DKDP crystal growth prediction, set the model hyperparameter search space, and construct a gradient boosting decision tree model. The construction of the gradient boosting decision tree model includes:

[0050] (1) Set the post-processing characteristic variables, with the daily growth change as the target variable;

[0051] (2) The optimization objective function is defined as minimizing the mean square error of the DKDP crystal growth prediction;

[0052] (3) Set the hyperparameter search space as follows: the range of the number of trees is [1000, 5000], the range of the learning rate is [0.001, 0.2], the range of the maximum tree depth is [1, 55], the range of the minimum number of samples for internal node splitting is [2, 50], and the range of the minimum number of samples for leaf nodes is [1, 50];

[0053] (4) Set a random seed and iteratively search the hyperparameter space through the Bayesian optimization algorithm to obtain a set of optimized hyperparameter combinations.

[0054] Furthermore, the preprocessed training set is used to train the gradient boosting decision tree model. During the training process, a random seed is set, and the Bayesian optimization algorithm is used to iteratively search the hyperparameter search space. Through continuous iterative search, the optimal hyperparameter combination that maximizes the model prediction effect is obtained to complete the model training. Among them, the Bayesian optimization model hyperparameters include:

[0055] (1) Define the objective function as the coefficient of determination R^2 of growth prediction;

[0056] (2) Setting the range space of parameters such as the number of trees and the maximum depth of trees;

[0057] (3) Using the Bayesian algorithm, after 500 iterative searches, a set of optimized hyperparameter combinations was obtained to complete the model training.

[0058] Step S4, obtaining the influencing factor data of the current DKDP crystal growth test in the same test device, and using the trained gradient boosting decision tree model to predict the DKDP crystal growth amount. In this embodiment, during the DKDP crystal growth process, the temperature, height and saturation point temperature data monitored on the day are used as model inputs, and the gradient boosting decision tree model optimized with hyperparameters is used to predict the daily growth amount of the DKDP crystal, so that the growth amount prediction value under this condition can be quickly obtained.

[0059] Furthermore, the above method is applied to the cultivation of DKDP crystals, and the predicted value of the DKDP crystal growth on that day is compared with the actual value of the growth. A judgment is made based on the comparison result. If the growth is abnormal, the experimental environment for controlling the growth of the DKDP crystal is adjusted in time to realize the monitoring and scientific control of the DKDP growth process.

[0060] In this embodiment, during the DKDP crystal growth process, the monitored temperature, height and saturation point temperature data are input daily, the growth amount is predicted using the optimized model, and the actual growth amount is compared and analyzed to determine the growth state, so as to adjust and control the growth environment and reduce the defect rate.

[0061] Furthermore, the experimental environment for controlling the growth of the DKDP crystal is adjusted according to the comparison results, including: if the predicted value is greater than the actual value, the X-axis and Y-axis of the DKDP crystal both grow while the Z-axis growth is inhibited, and the Z-axis growth rate of the DKDP crystal is slow; if the predicted value is less than the actual value, the DKDP crystal exhibits an inclusion phenomenon, and the Z-axis growth rate of the DKDP crystal is fast.

[0062] like Figure 2 As shown, the newly grown crystal data is imported into the model for calculation. By inputting T, Ts, and h every day during the growth process of the new crystal into the model, the model predicts the growth amount (mm), and generates an image of the predicted value-real value and the number of days. There are two obvious deviation areas in the image. The actual growth rate of the crystal z axis in the first 60 days is very slow. Through crystal slicing and recording, it can be found that the crystal has a certain expansion in the X and Y axes during the actual growth process, which leads to a slow growth rate on the Z axis; in the area from 110 days to 180 days, it is mainly due to the phenomenon of crystal inclusion, which is the reason for the faster actual growth rate of the Z axis during this period. Therefore, the model can obtain the difference between the actual growth and predicted growth rates from the record of the actual growth parameters of the input crystal, and thus deduce the approximate type of defects in the crystal growth process according to the speed of the Z axis growth rate.

[0063] As another implementation method, the data of the current DKDP crystal growth test process is recorded, the recorded data is added to the data set, the updated data set is used to train and optimize the gradient boosting decision tree model again, and the optimized gradient boosting decision tree model is used to predict and optimize the next DKDP crystal growth test.

[0064] As another implementation, based on the influencing factor data of the current DKDP crystal growth test for multiple days in the same test device, and the difference between the predicted value and the actual value of the crystal growth amount, a regression function is fitted to obtain the relationship between the influencing factor data of this DKDP crystal growth test and the predicted difference in crystal growth amount. This functional relationship can construct a functional relationship between the temperature T, the solution saturation point temperature Ts, the crystal height h and the difference △h, and the fitted functional relationship is as follows:

[0065] △h=A*exp(B*T+C*Ts+D*h);

[0066] Among them, A=0.285042003231673; B=0.1175022564591499; C=-0.22475564624536615; D=0.7072196248592029.

[0067] Afterwards, based on the regression function, by reading the current T, Ts, and h data on the instrument and substituting them into the functional relationship, the crystal z-axis growth △h of the day can be directly obtained, thereby quickly calculating and determining whether the relationship between the current DKDP crystal growth prediction value and the actual value is consistent. If not, adjust the experimental environment that controls the DKDP crystal growth to reduce the possibility of crystal defects.

[0068] As another implementation, the obtained optimal model is made into a python software package, and △h can be predicted directly by sliding the data bars corresponding to T, Ts, and h, such as Figure 3 As shown, the model is encapsulated into a python software package, which can greatly simplify the errors caused by related processes such as the calculation of equations in practical applications and reduce the amount of operations in the actual process.

[0069] Compared with the existing method that relies on expert experience growth model, the above method proposed in this embodiment does not need to understand the physical mechanism of crystal growth, and the constructed data-driven model is lightweight and efficient, and the growth prediction accuracy can reach more than 80%. Through continuous optimization, the prediction performance and accuracy can be continuously improved, and the DKDP growth defect rate can be significantly reduced, realizing the industrialized rapid and high-quality production of DKDP. The method described in this embodiment can be widely used in the quality monitoring and optimization of crystal materials of all sizes, especially large-caliber (usually tens of centimeters) crystal materials, and has important scientific research value and application prospects.

[0070] In addition, it should be pointed out that the crystal growth prediction method proposed in this embodiment can be applied to the growth prediction of potassium dideuterium phosphate DKDP crystals. Considering that deuterium and hydrogen are isotopes, the related crystals of the above-mentioned DKDP, such as potassium dihydrogen phosphate KDP, ammonium dihydrogen phosphate ADP, dideuterium ammonium phosphate DADP crystals, etc., have the same chemical properties as DKDP crystals. Therefore, the proposed prediction method can also be applied to the growth prediction of potassium dihydrogen phosphate KDP, ammonium dihydrogen phosphate ADP, dideuterium ammonium phosphate DADP crystals, etc.

[0071] Embodiment 2

[0072] This embodiment provides an application of a GBDT-based crystal growth prediction method in crystal cultivation to achieve quality monitoring and optimization of large-diameter crystal material production, wherein the crystal is any one of dideuterium potassium phosphate DKDP, dipotassium dihydrogen phosphate KDP, diammonium dihydrogen phosphate ADP, and dideuterium ammonium phosphate DADP crystals.

[0073] Embodiment 3

[0074] This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the GBDT-based crystal growth prediction method as described above are completed.

[0075] Embodiment 4

[0076] This embodiment also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps in the crystal growth amount prediction method based on GBDT as described above are completed.

[0077] The steps involved in the above embodiments 2 to 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0078] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0079] The above description is only a preferred embodiment of the present invention. Although the specific implementation mode of the present invention is described in conjunction with the accompanying drawings, it is not a limitation of the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the protection scope of the present invention.

Claims

1. A method for predicting crystal growth based on GBDT, characterized in that: include: Obtain historical data of crystal growth tests in the test device, determine the influencing factors affecting crystal growth, and construct a data set with daily influencing factor data and actual crystal growth data; the influencing factors include crystal growth environment temperature, crystal height and solution saturation point temperature; the crystal is any one of dideuterium potassium phosphate DKDP, dipotassium dihydrogen phosphate KDP, diammonium dihydrogen phosphate ADP, and dideuterium ammonium phosphate DADP crystals; Preprocessing the data in the data set; the preprocessing includes removing outliers, extracting data features and normalizing the data, wherein removing outliers is used to reduce data noise and improve prediction accuracy; Taking minimizing the mean square error of crystal growth prediction as the objective function, a gradient boosting decision tree model is constructed, and the gradient boosting decision tree model is trained using the preprocessed training set, and the Bayesian optimization model hyperparameters are used until the model prediction reaches the optimal; taking minimizing the mean square error of crystal growth prediction as the objective function, the model hyperparameter search space is set, and the gradient boosting decision tree model is constructed, wherein the hyperparameters include the number of trees, the learning rate, the maximum tree depth, the minimum number of samples for internal node splitting, and the minimum number of samples for leaf nodes; the range of the number of trees is [1000, 5000], the range of the learning rate is [0.001, 0.2], the range of the maximum tree depth is [1, 55], the range of the minimum number of samples for internal node splitting is [2, 50], and the range of the minimum number of samples for leaf nodes is [1, 50]; The gradient boosting decision tree model is trained using the preprocessed training set. During the training process, a random seed is set and the Bayesian optimization algorithm is used to iteratively search the hyperparameter search space. Through continuous iterative search, the optimal hyperparameter combination that maximizes the model prediction effect is obtained to complete the model training. Obtain the influencing factor data of the current crystal growth test in the same test device, and use the trained gradient boosting decision tree model to predict the crystal growth amount; The removing of outliers includes: calculating the Z-axis height of the actual daily growth of the crystal, and removing abnormal data whose absolute value of the Z-axis height exceeds a threshold; The data feature extraction includes: performing hourly splitting on the daily crystal growth environment temperature data, obtaining the temperature change curve of the day, extracting the average temperature and temperature variance of the day as characteristic variables; and taking the crystal height and the solution saturation point temperature as characteristic variables; Compare the predicted value of crystal growth on that day with the actual value of growth, and adjust the experimental environment for controlling crystal growth according to the comparison results, including: If the predicted value is greater than the actual value, the crystal grows in both the Y and Y axes and inhibits the growth of the Z axis, and the growth rate of the crystal in the Z axis is slow; If the predicted value is less than the actual value, the crystal will be included and the crystal will grow fast along the Z axis; According to the influencing factor data of the current crystal growth test for multiple days in the same test device, and the difference between the predicted value and the actual value of the crystal growth amount, a regression function is fitted to obtain the relationship between the influencing factor data of this crystal growth test and the predicted difference of the crystal growth amount; the fitted function relationship is as follows: △h=A*exp(B*T+C*Ts+D*h) Wherein, △h is the difference between the predicted value and the actual value of the growth amount; T is the temperature; Ts is the saturation point temperature of the solution; h is the crystal height; A = 0.285042003231673; B = 0.1175022564591499; C = -0.22475564624536615; D = 0.7072196248592029; Based on the regression function, a quick calculation is performed to determine whether the relationship between the current crystal growth prediction value and the actual value is consistent. If not, the experimental environment for controlling crystal growth is adjusted.

2. The GBDT-based crystal growth prediction method according to claim 1, wherein: include: Record the data of the current crystal growth test process, add the recorded data to the data set, use the updated data set to train and optimize the gradient boosting decision tree model again, and use the optimized gradient boosting decision tree model to predict and optimize the next crystal growth test.

3. Application of a GBDT-based crystal growth prediction method as described in any one of claims 1-2 in the cultivation of DKDP crystals and related crystals, wherein the related crystals include potassium dihydrogen phosphate KDP, ammonium dihydrogen phosphate ADP, and dideuterium ammonium phosphate DADP crystals.

4. An electronic device, characterized in that: The invention comprises a memory and a processor and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of a method for predicting crystal growth based on GBDT as claimed in any one of claims 1 to 2 are completed.

5. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of a crystal growth prediction method based on GBDT as described in any one of claims 1-2.

Images