A Method for Constructing a Mechanism-Guided Few-Shot Prediction Model of Polishing Material Removal Rate

Abstract

The present invention discloses a method for constructing a mechanism-guided few-shot polishing material removal rate prediction model with low cost and high precision. Based on the Preston equation, combined with the mechanism characteristics of the polishing technology and the simulation modeling of the polishing process, this method deeply studies the relationship between the polishing process parameters and the material removal rate, establishes a mathematical model for predicting the material removal rate with a certain accuracy, and then designs the neural network model structure based on the mathematical form of this mathematical model, and trains this neural network model to approximate the established mathematical model, so as to obtain a pre-trained neural network model; finally, based on the transfer learning strategy, continue to train this neural network model on a small number of real samples to obtain the final material removal rate prediction model; this prediction model integrates the interpretability of the mathematical model and the non-linear fitting ability of the neural network model, has high prediction accuracy, and can provide reliable prediction support for the actual polishing process.

Description

Technical Field

[0001] The present invention relates to the fields of ultra-precision machining and artificial intelligence technology, and specifically relates to a method for constructing a few-shot polishing material removal rate prediction model guided by mechanism. Background Art

[0002] Shear thickening polishing technology is a non-contact ultra-precision polishing method that utilizes the non-linear thickening rheological properties of non-Newtonian fluid polishing liquid under the action of shear stress to achieve material removal. Deterministic shear thickening polishing means that the active rotation of the polishing head drives the flow of the polishing liquid, causing the flowing polishing liquid to produce a shear thickening effect, forming particle clusters that enclose abrasive grains, thereby performing deterministic removal machining on a small area of the workpiece material.

[0003] In order to optimize the polishing process, it is usually necessary to predict and model the removal rate of the polishing material. In the prior art, some researchers have conducted relevant research based on mathematical and physical theories and simulation methods. For example, in the prior art, someone designed a constitutive equation for shear thickening polishing based on the Preston equation, then studied the velocity field and pressure field distributions during shear thickening deterministic polishing based on fluid simulation, and finally corrected the simulation coefficients through experiments to achieve the modeling of the prediction of the shear thickening polishing material removal rate. Although starting from principles and physical laws, it has strong theoretical support and good generalization. However, due to the complexity of the shear thickening polishing process, the theoretical-based modeling method is often a modeling under certain ideal conditions, which assumes that the material removal rate is only related to the polishing speed, gap height, and polishing liquid concentration, and cannot take into account various complex dynamic factors during the polishing process, such as changes in polishing liquid temperature and rheological properties, resulting in it being difficult to obtain high-precision prediction results in real scenarios.

[0004] In recent years, with the development of artificial intelligence technology, more and more researchers have started to study modeling methods based on machine learning / deep learning. Deep learning models represented by neural networks have powerful non-linear fitting capabilities and can update model parameters during continuous learning based on backpropagation. This powerful non-linear fitting ability can, from the perspective of real data, transform the potential laws between input and output into the parameters of the neural network through continuous learning, without relying on various restrictive assumptions, thereby achieving high-precision prediction in real scenarios. However, its modeling process relies on a large number of high-quality data samples for training, and real high-quality experimental data samples must rely on professional experimental personnel to conduct polishing experiments, requiring a large amount of manpower and material resources, and are often difficult to obtain. Generally, only a small amount of experimental data is available for use. If directly based on deep learning modeling in the case of few samples, it often leads to problems of overfitting / underfitting, and a high-precision prediction model cannot be obtained. This makes it difficult for artificial intelligence technology to be effectively applied in the prediction modeling of the material removal rate of shear thickening polishing materials. Therefore, researchers have proposed transfer learning technology, that is, first let the model be trained on a large number of easily obtainable data samples in similar scenarios to obtain a pre-trained model, and then fine-tune the model on a small number of difficult-to-obtain real target samples, so as to achieve good modeling results using a small number of samples in the target scenario. However, in the field of prediction modeling of the material removal rate of shear thickening polishing materials, the above method cannot be directly applied, because on the one hand, there is a lack of such a publicly available large-scale dataset for pre-training in the polishing field, and on the other hand, the principles and characteristics of various polishing methods have significant differences, making it difficult to establish a publicly available large-scale dataset with a unified modeling form. To sum up, how to utilize the non-linear fitting ability of the neural network model and construct a high-precision material removal rate prediction model based on a small number of shear thickening polishing experimental sample data is an urgent problem to be solved at present. Summary of the Invention

[0005] The present invention provides a method for constructing a mechanism-guided few-shot polishing material removal rate prediction model with low cost and high precision, which is used to solve the problem of lack of real experimental samples in traditional neural network-based material removal rate modeling methods, and the problem of low modeling accuracy caused by the inability of pure mathematical and physical modeling methods to reflect the real complex processing process. The prediction model construction method of the present invention first constructs a mathematical model for predicting the material removal rate with a certain accuracy based on the Preston equation, combined with the mechanism characteristics of the polishing technology and the simulation modeling of the polishing process; then designs a neural network model based on the mathematical form of the mathematical model and guides the pre-training process of the neural network to obtain a pre-trained neural network model with a certain prediction accuracy; finally, based on the transfer learning strategy, continue to train the neural network model on a small number of real samples to obtain the final polishing material removal rate prediction model; this prediction model integrates the interpretability of the mathematical model and the non-linear fitting ability of the neural network model, has high prediction accuracy, can provide reliable prediction support for the actual polishing process, thus guiding production practice, optimizing the polishing process, and improving production efficiency and product quality.

[0006] The technical solution of the present invention is as follows:

[0007] A method for constructing a mechanism-guided few-shot polishing material removal rate prediction model, the method includes the following steps: S1, determine the polishing object, design a polishing experiment, construct a real experimental data set D, each sample in the real experimental data set D contains a real polishing process parameter combination and the corresponding material removal rate, and then combine the polishing principle and the Preston equation: , establish a mathematical model M for predicting the polishing material removal rate; where MRR is the material removal rate, P is the polishing pressure, V is the polishing speed, and k is a coefficient; S2, set different polishing process parameter combinations and input them into the mathematical model M, and output the predicted results of the material removal rate to obtain a simulated experimental data set Dm, each experimental data contains a set of simulated polishing process parameter combinations and the predicted results of the corresponding material removal rate; S3, according to the mathematical form of the mathematical model M, design an initial state neural network model N, and train it on the simulated experimental data set Dm until convergence, so that the trained model N infinitely approaches the mathematical model M to obtain a pre-trained model N_1; S4, fine-tune the pre-trained model N_1 on the real experimental data set D, optimize the model parameters based on the real experimental data, so as to further improve the prediction accuracy of the pre-trained model N_1 and obtain the final polishing material removal rate prediction model N_2.

[0008] Compared with the prior art, the method for constructing a mechanism-guided few-shot polishing material removal rate prediction model organically combines theoretical modeling and neural network training, realizes high-precision prediction modeling of the polishing material removal rate in the few-shot case, and provides strong technical support for the optimization of the polishing process. It has the following advantages:

[0009] (1) The present invention constructs a mathematical model for predicting the polishing material removal rate based on theoretical analysis and simulation modeling, and on this basis, guides the pre-training process of the neural network model, enabling it to have a certain prediction accuracy before contacting real experimental data, and fine-tuning and training on a small number of real samples, thereby further improving the prediction accuracy. The final prediction model not only has the high-precision prediction ability of the neural network but also inherits the interpretability of the mathematical model. This interpretability is crucial for understanding the decision-making process of the model, verifying the rationality of the model, and establishing users' trust in the model prediction results;

[0010] (2) Using simulated data for pre-training reduces the dependence on real experimental samples, thereby reducing experimental costs and time consumption. Moreover, the model can learn a wider range of polishing conditions, enhancing the model's adaptability to unknown conditions;

[0011] (3) The present invention reduces the number of experiments and material consumption, conforms to the production concept of green environmental protection, is not only applicable to the prediction of the shear thickening polishing material removal rate but also can be extended to the construction of prediction models for other material processing processes, and has high practical value and market prospects.

[0012] As an optimization, in the method for constructing a mechanism-guided few-shot polishing material removal rate prediction model described above, in step S1, the polishing experiment is a shear thickening polishing experiment, and the formula of the mathematical model M extended based on the Preston equation is: ; where Kc is the abrasive concentration correction coefficient, Ks is the simulation correction coefficient, K(w,h) is the correction coefficient related to the polishing speed and polishing gap, P(x,y) is the pressure distribution on the xy plane, and V(x,y) is the velocity distribution on the xy plane. Further, the simulation correction coefficient Ks and the abrasive concentration correction coefficient Kc are obtained by modeling the pressure and velocity fields in the polishing area and fitting them with the polishing experimental data.

[0013] As an optimization, in the method for constructing a mechanism-guided few-shot polishing material removal rate prediction model described above, in step S3, the formula of the neural network model N in the initial state is: ; where X represents various combinations of polishing process parameters; represents a custom activation function designed according to the mathematical form of the model M and related to the variables Xi, Xj, Xz; α represents a non-linear activation function; F represents the forward calculation process of the multi-layer perceptron.

[0014] Further, in step S3, the simulated experimental dataset Dm is divided into a simulated training set Dm_train and a simulated test set Dm_test; first, the neural network model N is trained on the simulated training set Dm_train until the difference between the prediction result of the trained neural network model on the test set Dm_test and the prediction result of the mathematical model M is less than the set threshold T, obtaining a pre-trained model N_1, that is . At this time, the acquisition of the pre-trained model N_1 not only depends on the training effect of the model on the simulated data, but also needs to be verified through the test set, so as to ensure the generalization ability of the model and the prediction accuracy of the model.

[0015] Further, the training process of the neural network model N includes forward propagation, loss calculation, backward propagation, and parameter update, and through iterative training, the model parameters are continuously optimized. Thus, the neural network model N can gradually improve its performance and accuracy.

[0016] As an optimization, in the above-mentioned method for constructing a mechanism-guided few-shot polishing material removal rate prediction model, in step S4, the real experimental dataset D is divided into a real training set D_train and a real test set D_test; first, the pre-trained model N_1 is fine-tuned on the real training set D_train until the prediction accuracy P_n of the pre-trained model N_1 on the real test set D_test is greater than the prediction accuracy P_m of the mathematical model M on the real test set D_test and the accuracy no longer rises, then the final material removal rate prediction model N_2 is obtained.

[0017] The specific judgment process is as follows. Use D T and D N 's MSE as the loss of the pre-trained model N_1 during the fine-tuning training process: ; where D T represents the material removal rate label value on the real test set D_test, D N is the output value of the pre-trained model N_1, D M is the output value of the mathematical model M, and MSE (Mean Squre Error) represents the mean square error, which is a commonly used indicator to measure the prediction accuracy of the model; when it reaches on the real test set D_test, then it is considered that P_n is greater than P_m.

[0018] As an optimization, in the method for constructing the mechanism-guided few-shot polishing material removal rate prediction model described above, when dividing the simulated experimental dataset and the real experimental dataset, the division ratio of the training set and the test set can be adjusted according to the actual situation, but usually remains at a ratio of 65-75% for training and 25-35% for testing. Specifically: the data ratio in the simulated training set Dm_train and the simulated test set Dm_test is 7:3; the data ratio in the real training set D_train and the real test set D_test is 7:3, so as to ensure the effectiveness of model training and the accuracy of testing.

[0019] As an optimization, in the method for constructing the mechanism-guided few-shot polishing material removal rate prediction model described above, in step S1, the shear thickening polishing experiment is realized by using a shear thickening deterministic polishing device; the deterministic polishing device includes a polishing tank, a polishing head arranged above the polishing tank, and a lifting mechanism for driving the polishing head to move up and down; a fixture is arranged inside the polishing tank, and a pressure sensor is arranged at the bottom; the polishing head is connected to a motor, and a horizontal measuring instrument is arranged on the motor; during the experiment, the workpiece is fixed in the polishing tank by the fixture, and a polishing liquid is filled in the polishing tank to immerse the workpiece, and the lower part of the polishing head is located in the polishing liquid; then the motor drives the polishing head to rotate to remove the surface material of the workpiece. Using the polishing device with the above structure for polishing experiments has low implementation difficulty and is easy to obtain various polishing process parameters. Description of the Drawings

[0020] Figure 1 is a schematic diagram of the method for constructing the mechanism-guided few-shot polishing material removal rate prediction model of the present application;

[0021] Figure 2 is a schematic structural diagram of the deterministic polishing device in the embodiment of the present application;

[0022] Figure 3 is a schematic diagram of the neural network model in the initial state designed according to the mathematical model.

[0023] The marks in the drawings are: 1 - polishing tank; 2 - polishing head; 3 - fixture; 4 - pressure sensor; 5 - motor; 6 - horizontal measuring instrument; 7 - workpiece; 8 - polishing liquid. Detailed Embodiments

[0024] The present invention will be further described below with reference to the drawings and embodiments; the content not described in detail in the following embodiments is all common technical knowledge in the art.

[0025] To solve the problems of scarce sample size for predicting and modeling the material removal rate in shear thickening polishing process and low prediction accuracy based on mathematical and physical theories, the present invention proposes a method for constructing a mechanism-guided few-sample polishing material removal rate prediction model. The core lies in combining theoretical modeling with artificial intelligence algorithms. This method can achieve high-precision prediction modeling of the polishing material removal rate with a small number of real samples, providing strong technical support for the optimization of the polishing process.

[0026] This method is implemented through the following key steps:

[0027] First, based on the Preston equation, combined with the mechanism characteristics of the polishing technology and the simulation modeling of the polishing process, the present invention deeply studies the relationship between the polishing process parameters and the material removal rate, and establishes a mathematical model M for predicting the polishing material removal rate with a certain accuracy, thereby providing a theoretical basis for the subsequent construction of the prediction model and enhancing the interpretability of the constructed prediction model.

[0028] Next, the present invention adopts simulation data generation technology to generate a large number of simulated experimental data sets based on the mathematical model M; the key to this step lies in the diversity and representativeness of the simulated data, ensuring that the simulated data sets can cover various possible polishing conditions and providing rich data resources for the training of the neural network model.

[0029] On this basis, the present invention designs and trains a neural network model N. The structure of this model refers to the mathematical model M, aiming to achieve fast convergence in the training on the simulated experimental data, enabling the model to infinitely approach the prediction accuracy of the mathematical model M, thereby obtaining the pre-trained model N_1; this process makes full use of the advantages of the neural network in dealing with complex non-linear relationships. Through the training of the simulated data, the pre-trained model N_1 has the prediction ability equivalent to that of the mathematical model.

[0030] Finally, the present invention fine-tunes and trains the pre-trained model N_1 with a small amount of real experimental data; the purpose of this step is to improve the prediction accuracy of the model on the real experimental data. By fine-tuning the model parameters, the model can better fit the complex real processing process, and finally obtain a higher-precision polishing material removal rate prediction model; this prediction model not only has the high interpretability of the mathematical model M but also realizes the high-precision prediction ability based on the neural network.

[0031] Embodiment:

[0032] A method for constructing a mechanism-guided few-sample polishing material removal rate prediction model, which specifically includes the following steps.

[0033] S1. Determine the polishing object, design a shear thickening polishing experiment, and the experimental data is shown in the following table. Based on this, construct a real experimental dataset D. Each sample in the real experimental dataset D contains a set of real polishing process parameter combinations and the corresponding material removal rate. Then, combine the shear thickening polishing principle and the Preston equation: , establish a mathematical model M for the material removal rate of shear thickening polishing. Among them, MRR is the material removal rate, P is the polishing pressure, V is the polishing speed, and k is a coefficient related to the polishing object and the polishing fluid.

[0034] Experimental conditions Parameters Glass sheet 1.5 (inch) Abrasive <![CDATA[Al2O3]]> Abrasive grain size number #3000 Abrasive grain concentration 3, 5, 7, 9 (wt%) Polishing head rotation speed 4, 8, 12, 16 (r / s) Polishing gap 0.6, 0.8, 1, 1.2 (mm)

[0035] When designing the polishing experiment, it is necessary to consider the characteristics of the polishing object, including but not limited to material hardness, surface roughness, chemical composition, etc. These characteristics will directly affect the parameter settings of the polishing experiment and the final material removal rate.

[0036] In this embodiment, the shear thickening polishing experiment is realized by a shear thickening deterministic polishing device. The deterministic polishing device includes a polishing tank 1, a polishing head 2 arranged above the polishing tank 1, and a lifting mechanism for driving the polishing head 2 to move up and down. A fixture 3 is arranged inside the polishing tank 1, and a pressure sensor 4 is arranged at the bottom. The polishing head 2 is connected to a motor 5, and a horizontal measuring instrument 6 is arranged on the motor 5. During the experiment, the workpiece 7 is fixed in the polishing tank 1 by the fixture 3, and a polishing fluid 8 is filled in the polishing tank 1 to immerse the workpiece 7. The lower part of the polishing head 2 is located in the polishing fluid 8, and there is a polishing gap between the polishing head 2 and the workpiece 7. Then, the motor 5 drives the polishing head 2 to rotate, so that the polishing fluid 8 between the polishing head 2 and the workpiece 7 generates a shear rheological effect to remove the surface material of the workpiece 7.

[0037] The polishing gap height between the workpiece 7 and the polishing head 2 and the rotation speed of the polishing head 2 respectively affect the pressure distribution and flow field distribution of the polishing fluid 8 on the surface of the workpiece 7, which are important influencing factors for the material removal rate. In addition, the concentration of the polishing fluid 8 is also an important factor affecting the material removal rate. Based on the shear thickening polishing principle, the Preston equation is extended to establish a constitutive equation for the material removal rate of shear thickening polishing as: ; where Kc is the abrasive concentration correction coefficient, Ks is the simulation correction coefficient, K(w,h) is the correction coefficient related to the polishing speed and the polishing gap, and P(x,y) is the pressure distribution on the xy plane, and V(x,y) is the velocity distribution on the xy plane.

[0038] Model the pressure and velocity fields in the polishing area, fit the modeling results into a mathematical form and substitute them into the above constitutive equation for the material removal rate. Then, combine the polishing experimental data to fit the simulation correction coefficient Ks and the abrasive concentration correction coefficient Kc. The finally constructed mathematical model M is: Among them, C represents the abrasive concentration; w and h respectively represent the rotational speed of the polishing head and the height of the polishing gap; x and y represent the coordinates of the polishing area.

[0039] S2. Set different combinations of polishing process parameters as inputs to the mathematical model M, and output the predicted results of the material removal rate to obtain the simulated experimental data set Dm. Each piece of experimental data contains a set of simulated polishing process parameter combinations and the corresponding predicted results of the material removal rate. Divide the simulated experimental data set Dm into a simulated training set Dm_train and a simulated test set Dm_test according to a ratio of 7:3. The generation of the simulated experimental data set needs to consider the randomness and diversity of the polishing process parameters to ensure that the generated data set can cover a wide range of experimental conditions, thereby improving the generalization ability of the model.

[0040] S3. According to the mathematical form of the mathematical model M, design the neural network model N in the initial state. When designing, select and design appropriate network architectures, activation functions, loss functions, optimization algorithms, etc. with reference to the form of the mathematical model M to ensure that the model can effectively learn and approximate the mathematical model M. Its formula is: ; where X represents various combinations of polishing process parameters; represents a custom activation function related to variables Xi, Xj, Xz designed according to the mathematical form of model M; α represents a non-linear activation function, such as relu, etc., used to provide non-linear capabilities and fit the real experimental data; F represents the forward calculation process of the multi-layer perceptron.

[0041] In this embodiment, according to the mathematical form of the mathematical model M, custom activation functions θ and λ are defined to simulate the influence of factors such as the rotational speed of the polishing head, the height of the polishing gap, and the abrasive concentration on the material removal rate, and improve the fitting ability of the neural network model N to the mathematical model M. By observing the form of the mathematical model M, it can be found that w, h, and x, y all satisfy a certain Gaussian distribution with the material removal rate. Therefore, the Gaussian surface equations can be set as the activation functions θ and λ respectively and designed in the neural network model N to simulate the mathematical representation of the mathematical model M: ;

[0042]

[0043] Then, train the neural network model N on the simulated training set Dm_train with model parameters W and set the accuracy threshold T. The training process includes forward propagation, loss calculation, backward propagation, and parameter update. Through iterative training, continuously optimize the model parameters until the prediction performance of the model on the simulated experimental data approaches the mathematical model M infinitely, obtaining the pre-trained model N_1, that is, M≈N_1 in terms of prediction accuracy. The so-called infinitely approaching the mathematical model M means that on the same batch of simulated test sets Dm_test, the difference between the prediction results of the neural network pre-trained model N_1 and the mathematical model M is less than the set threshold T, that is , the neural network model parameters are updated to W_1. At this time, the acquisition of the pre-trained model N_1 not only depends on the training effect of the model on the simulated data, but also needs to be verified through the test set to ensure the generalization ability and prediction accuracy of the model.

[0044] S4, divide the real experimental data set D into a real training set D_train and a real test set D_test according to a ratio of 7:3. Then, let the pre-trained model N_1 perform fine-tuning training on the real training set D_train, optimize the model parameters based on the real experimental data, and further improve the prediction accuracy of the pre-trained model N_1 until the prediction accuracy P_n of the pre-trained model on the real test set D_test is greater than the prediction accuracy P_m of the mathematical model M on the real test set D_test and no longer increases, obtaining the final material removal rate prediction model N_2, and the neural network model parameters are updated from W_1 to W_2, and the modeling is completed.

[0045] The process of fine-tuning training needs to be carefully designed, including selecting appropriate learning rates, fine-tuning strategies, and training cycles, etc., to ensure that the model can quickly adapt and improve the prediction accuracy on a small amount of real data. When verifying the prediction model, D T and D N 's MSE can be used as the loss of the prediction model during the fine-tuning training process: ; where, D T represents the material removal rate label value on the real test set D_test, D N is the output value of the prediction model (the predicted value obtained by inputting the data on the real test set into the prediction model), D M is the output value of the mathematical model M (the predicted value obtained by inputting the data on the real test set into the mathematical model M), MSE (Mean Squre Error) represents the mean square error, which is a commonly used indicator to measure the prediction accuracy of the model. The smaller this value is, the higher the prediction accuracy of the model. When reaching on the real test set D_test, it is considered that the prediction accuracy of the finally trained model has been further improved on the basis of the mathematical model constructed based on the polishing process mechanism theory.

[0046] In this embodiment, in the real test set D_test, the true values of the material removal rate are [10.7, 5.5, 10.2, 5.9, 8.45, 8.27, 9.96, 10.57, 6.0] (μm / h); the predicted values of the mathematical model M on the real test set D_test are [10.27656018, 5.54128245, 10.17580959, 5.94428481, 8.46304956, 7.40925152, 8.0060315, 11.38117627, 6.0450354]; the predicted values of the pre-trained model N_1 on the real test set D_test are [10.15728069, 4.79993754, 9.72893041, 5.53493383, 8.34791813, 7.09903125, 8.96807892, 10.51678517, 5.19792818]; the predicted values of the final prediction model N_2 on the real test set D_test are [10.27388505, 5.00993888, 9.64422017, 6.07567224, 8.85241334, 7.85554219, 9.678999, 10.15477847, 5.4139422]; the mean squared error (MSE) is used to measure the difference between the predicted values and the true values of the models. MSE(true values, predicted values of the mathematical model) = 0.6, MSE(true values, predicted values of the pre-trained model) = 0.461, MSE(true values, predicted values of the final prediction model) = 0.188; it can be seen that the prediction accuracy of the final material removal rate prediction model obtained by the construction method of the present application is relatively high.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a mechanism-guided few-shot polishing material removal rate prediction model, characterized in that, The method includes the following steps: S1. Determine the polishing object, design a shear thickening polishing experiment, and construct a real experimental data set D. Each sample in the real experimental data set D contains a real combination of polishing process parameters and the corresponding material removal rate. Then, based on the shear thickening polishing principle, expand the Preston equation: to establish the constitutive equation of the material removal rate for shear thickening polishing as: ; where MRR is the material removal rate, P is the polishing pressure, V is the polishing speed, and k is a coefficient; Kc is the abrasive concentration correction coefficient, Ks is the simulation correction coefficient, K(w, h) is the correction coefficient related to the polishing speed and polishing gap, P(x, y) is the pressure distribution on the xy plane, and V(x, y) is the velocity distribution on the xy plane; c represents the abrasive concentration, w and h represent the rotational speed of the polishing head and the height of the polishing gap respectively, and x, y represent the coordinates of the polishing area. Then, model the pressure and velocity fields in the polishing area, fit the modeling results into a mathematical form and substitute it into the above constitutive equation of the material removal rate, and combine the polishing experimental data to fit the simulation correction coefficient Ks and the abrasive concentration correction coefficient Kc to obtain the mathematical model M for predicting the polishing material removal rate. S2. Set different combinations of polishing process parameters as inputs to the mathematical model M, and output the predicted results of the material removal rate to obtain the simulated experimental data set Dm. Each piece of experimental data includes a set of simulated polishing process parameter combinations and the corresponding predicted results of the material removal rate; S3. According to the mathematical form of the mathematical model M, design the neural network model N in the initial state, and train it on the simulated experimental data set Dm until convergence, so that the trained model N approximates the mathematical model M infinitely to obtain the pre-trained model N_1; S4. Fine-tune the pre-trained model N_1 on the real experimental data set D, and optimize the model parameters based on the real experimental data, so as to further improve the prediction accuracy of the pre-trained model N_1 to obtain the final predicted model N_2 of the polishing material removal rate, and the modeling is completed.

2. The method for constructing a few-shot polishing material removal rate prediction model guided by a mechanism according to claim 1, wherein: In the step S3, the formula of the neural network model N in the initial state is as follows: where X represents various combinations of polishing process parameters; represents a custom activation function related to variables Xi, Xj, Xz designed according to the mathematical form of the model M; α represents a non-linear activation function; F represents the forward calculation process of the multi-layer perceptron.

3. The method for constructing a few-shot polishing material removal rate prediction model guided by a mechanism according to claim 2, wherein: In the step S3, the simulated experimental data set Dm is divided into a simulated training set Dm_train and a simulated test set Dm_test; the neural network model N is trained on the simulated training set Dm_train until the difference between the prediction result of the trained neural network model on the test set Dm_test and the prediction result of the mathematical model M is less than the set threshold T, and a pre-trained model N_1 is obtained, that is .

4. The method for constructing a few-shot polishing material removal rate prediction model guided by a mechanism according to claim 3, wherein: The training process of the neural network model N includes forward propagation, loss calculation, backpropagation, and parameter update, and continuously optimizes the model parameters through iterative training.

5. The method for constructing a few-shot polishing material removal rate prediction model guided by a mechanism according to claim 3, characterized in that: In step S4, the real experimental data set D is divided into a real training set D_train and a real test set D_test; first, let the pre-trained model N_1 perform fine-tuning training on the real training set D_train until the prediction accuracy P_n of the pre-trained model N_1 on the real test set D_test is greater than the prediction accuracy P_m of the mathematical model M on the real test set D_test and no longer rises, then the final predicted model N_2 of the material removal rate is obtained.

6. The method for constructing a mechanism-guided few-shot polishing material removal rate prediction model according to claim 5, characterized in that: Use D T and D N The MSE of is used as the loss of the pre-trained model N_1 during the fine-tuning training process: ; D T represents the label value of the material removal rate on the real test set D_test, D N is the output value of the pre-trained model N_1, D M is the output value of the mathematical model M, and MSE represents the mean square error; when it reaches on the real test set D_test, it is considered that P_n is greater than P_m.

7. The method for constructing a few-shot polishing material removal rate prediction model guided by a mechanism according to claim 5, characterized in that: The data ratio in the simulated training set Dm_train and the simulated test set Dm_test is 7:3; the data ratio in the real training set D_train and the real test set D_test is 7:

3.

8. The method for constructing a few-shot polishing material removal rate prediction model guided by a mechanism according to claim 1, characterized in that: In step S1, the shear thickening polishing experiment is realized by using a shear thickening deterministic polishing device; the deterministic polishing device includes a polishing tank (1), a polishing head (2) arranged above the polishing tank (1), and a lifting mechanism for driving the polishing head (2) to move up and down; a fixture (3) is arranged inside the polishing tank (1), and a pressure sensor (4) is arranged at the bottom; the polishing head (2) is connected to a motor (5), and a horizontal measuring instrument (6) is arranged on the motor (5); during the experiment, the workpiece (7) is fixed in the polishing tank (1) by the fixture (3), and a polishing liquid (8) is filled in the polishing tank (1) to immerse the workpiece (7), and the lower part of the polishing head (2) is located in the polishing liquid (8); then the motor (5) drives the polishing head (2) to rotate to remove the surface material of the workpiece (7).

Images