Recommended methods, apparatus, equipment and storage media for wafer surface treatment processes.

Abstract

This application proposes a method, apparatus, device, and storage medium for recommending wafer surface treatment processes, relating to the field of semiconductor manufacturing technology. The method includes: acquiring the surface physicochemical characteristic parameters of a target wafer and preprocessing these parameters to obtain target physicochemical characteristic parameters; inputting the target physicochemical characteristic parameters into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for each process parameter of the target wafer; concatenating the target physicochemical characteristic parameters and the preliminary recommended parameters and inputting them together into a Transformer encoder based on an attention mechanism, utilizing a multi-head self-attention mechanism to model the complex coupling relationship between input features, and extracting deep representations to output the target recommended parameters. This method enables a nonlinear mapping between the surface physicochemical characteristic parameters of the target wafer and process parameters, thereby achieving efficient, accurate, and automated recommendation of wafer plasma surface treatment process parameters.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and in particular to a recommended method, apparatus, device and storage medium for wafer surface treatment process. Background Technology

[0002] Heterocrystalline wafer bonding is an advanced manufacturing technology that combines semiconductor wafers made of different materials with different surface treatments. This technology has wide applications in microelectronics manufacturing, microelectromechanical systems (MEMS) packaging, multi-functional chip integration, and other emerging fields. Direct heterocrystalline wafer bonding technology places extremely high demands on surface activation processes, requiring plasma treatment to adjust surface chemical properties to achieve high-quality bonding.

[0003] In wafer plasma surface treatment, there are complex coupling relationships among various process parameters. For example, changes in gas type and power directly affect the activation effect, while there is a delicate balance between activation time and vacuum level. Existing physical models and statistical methods often fail to fully capture the complexity of this nonlinear, multivariate coupling, making it difficult to achieve ideal surface treatment results in actual production. Summary of the Invention

[0004] In view of this, this application proposes a recommended method, apparatus, equipment and storage medium for wafer surface treatment process.

[0005] In a first aspect, this application provides a recommended method for wafer surface treatment, including:

[0006] The surface physicochemical properties parameters of the target wafer are obtained, and the surface physicochemical properties parameters are preprocessed to obtain the target physicochemical properties parameters;

[0007] The target physical and chemical properties parameters are input into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for the target wafer in terms of various process parameters;

[0008] The target materialized characteristic parameters and the preliminary recommendation parameters are concatenated and input into a Transformer encoder based on an attention mechanism. The complex coupling relationship between the input features is modeled using a multi-head self-attention mechanism, and a deep representation is extracted to output the target recommendation parameters.

[0009] In one embodiment, the recommended method for wafer surface treatment further includes:

[0010] The target physical property parameters and the target recommended parameters are input into the error rate prediction model based on the long short-term memory neural network to obtain the prediction error rate of each process parameter of the target wafer;

[0011] The target recommended parameters are calibrated based on the prediction error rate of each process parameter of the target wafer to generate calibrated recommended parameters.

[0012] In one embodiment, the recommended method for wafer surface treatment further includes:

[0013] Based on the target physical property parameters, select the reference wafer with the most similar physical properties from the wafer database;

[0014] Based on the physical and chemical properties parameters of the reference wafer, the single hidden layer neural network model, the Transformer encoder, and the error rate prediction model, the target prediction error rate of the reference wafer in terms of each process parameter is obtained.

[0015] Obtain the first error rate difference between the target predicted error rate and the actual error rate of each process parameter of the reference wafer;

[0016] The calibration recommendation parameters are compensated based on the first error rate difference to generate compensated recommendation parameters.

[0017] In one embodiment, the surface physicochemical property parameters include material chemical property parameters, roughness, and mechanical property parameters; the preprocessing of the surface physicochemical property parameters to obtain the target physicochemical property parameters includes:

[0018] The chemical property parameters of the material are linearly normalized to obtain the first characteristic parameter;

[0019] The roughness and the mechanical property parameters are respectively normalized by quantiles to obtain the second characteristic parameter and the third characteristic parameter;

[0020] The first feature parameter, the second feature parameter, and the third feature parameter are concatenated into an input feature matrix to obtain the target materialization property parameters.

[0021] In one embodiment, the single-hidden-layer neural network model includes: an input layer, a hidden layer, and an output layer;

[0022] The number of neurons in the input layer is the same as the number of parameter types included in the surface physical properties parameters;

[0023] The hidden layer comprises multiple neurons and uses a nonlinear function as the activation function;

[0024] The number of neurons in the output layer is the same as the number of parameter types included in the process parameters, and the activation function of the output layer is either a sigmoid activation function or a linear activation function.

[0025] The single-hidden-layer neural network model is trained using a sample set based on historical wafer processing experimental data to adjust the connection weights and bias parameters between the hidden layer and the output layer.

[0026] In one embodiment, the step of concatenating the target materialized feature parameters and the preliminary recommendation parameters and inputting them together into an attention-based Transformer encoder, utilizing a multi-head self-attention mechanism to model the complex coupling relationship between input features, and extracting deep representations to output target recommendation parameters includes:

[0027] The target materialized characteristic parameters and the preliminary recommended parameters are concatenated to form a multi-dimensional joint input vector;

[0028] The multidimensional joint input vector is mapped to the high-dimensional feature space required by the Transformer encoder through a linear embedding layer;

[0029] The Transformer encoder processes the embedded features in the high-dimensional feature space to generate a highly expressive hidden representation that integrates multi-source features, and then maps it to the target recommendation parameters through a regression head.

[0030] In one embodiment, the step of inputting the target physical property parameters and the target recommended parameters into the error rate prediction model based on a long short-term memory neural network to obtain the prediction error rate of each process parameter of the target wafer includes:

[0031] Error prediction of mature wafer samples is modeled as a time series process and defined as a continuous Ti sequence step error prediction task;

[0032] The target recommended parameters and physicochemical property parameters of the mature wafer sample are input into the error rate prediction model to obtain the prediction error rate of each sequence step output by the error rate prediction model, and the second error rate difference between the prediction error rate and the actual error rate of each sequence step is obtained.

[0033] A cumulative correction trend is formed based on the difference in the second error rate corresponding to each sequence step;

[0034] The target physical property parameters and the target recommended parameters are input into the error rate prediction model based on the long short-term memory neural network, and the preliminary prediction error rate of each process parameter is output.

[0035] The cumulative correction trend is used to correct the initial prediction error rate of each process parameter of the target wafer, so as to obtain the target prediction error rate of each process parameter.

[0036] Secondly, this application provides a recommended apparatus for wafer surface treatment processes, comprising:

[0037] The acquisition module is used to acquire the surface physicochemical property parameters of the target wafer and preprocess the surface physicochemical property parameters to obtain the target physicochemical property parameters.

[0038] The first recommendation module is used to input the target physical property parameters into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for the process parameters.

[0039] The second recommendation module is used to concatenate the target materialized characteristic parameters and the preliminary recommendation parameters and input them into the Transformer encoder based on the attention mechanism. It uses a multi-head self-attention mechanism to model the complex coupling relationship between the input features and extracts deep representations to output the target recommendation parameters.

[0040] Thirdly, this application provides an electronic device including a processor and a memory; the memory has a computer program stored thereon, wherein the computer program, when executed by the processor, implements the recommended method for wafer surface treatment process as described in the first aspect.

[0041] Fourthly, this application also provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the recommended method for wafer surface treatment process as described in the first aspect.

[0042] The proposed wafer surface treatment process method of this application has the following advantages over related technologies:

[0043] 1. This application rapidly generates preliminary recommended parameters for various process parameters of the target wafer by inputting the target physical property parameters into a pre-trained single-hidden-layer neural network model. Subsequently, the preliminary recommended parameters and the target physical property parameters are input into a Transformer encoder based on an attention mechanism. The multi-head self-attention mechanism of the Transformer encoder is used to model the complex coupling relationship between input features, extract deep representations to complete the final recommendation, and generate target recommended parameters. Thus, the neural network model captures the complex coupling relationship between various process parameters, realizes the nonlinear mapping between the surface physical property parameters of the target wafer and the process parameters, and achieves efficient, accurate, and automated recommendation of wafer plasma surface treatment process parameters.

[0044] 2. This application employs a single-hidden-layer neural network model for initial recommendation, and then uses a Transformer encoder to optimize the initial recommendation results to obtain the target recommendation parameters. This combines the fast response capability of a single-hidden-layer neural network with the deep modeling capability of the Transformer structure based on an attention mechanism for complex dependencies in input features. While improving prediction speed, it significantly enhances the model's adaptability to heterogeneous materials and multi-parameter coupling scenarios, exhibiting stronger adaptability and prediction accuracy when facing diverse wafer materials and complex process environments. Through the wafer surface treatment process recommendation method of this application, the wafer surface treatment process recommendation process shifts from traditional experience-driven to data-driven, not only reducing trial-and-error costs and debugging time, but also improving process consistency, yield, and stability, demonstrating good versatility, scalability, and industrialization prospects. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic flowchart of a recommended method for wafer surface treatment in one embodiment of this application;

[0047] Figure 2 This is a flowchart illustrating a recommended method for wafer surface treatment in another embodiment of this application;

[0048] Figure 3 This is a flowchart illustrating a similar wafer difference correction strategy in one embodiment of this application;

[0049] Figure 4 This is a schematic diagram of the network structure of a single hidden layer neural network model in one embodiment of this application;

[0050] Figure 5 This is a schematic diagram of the recommended apparatus for wafer surface treatment process in one embodiment of this application. Detailed Implementation

[0051] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0052] Currently, the process parameters for wafer surface plasma processing (such as vacuum level, gas type, power, processing time, and power frequency) often require extensive experimentation and adjustment based on engineers' experience. The optimal parameters vary significantly depending on the wafer material and surface condition. Each wafer material requires multiple experimental verifications, which is time-consuming and costly, resulting in long debugging cycles and poor process stability. Existing empirical models for new materials (such as GaAs) suffer from a lack of rapid adaptability, insufficient accuracy, and difficulty in accurately modeling the mapping relationship between complex material physicochemical properties (such as roughness and crystal orientation deviation) and process parameters.

[0053] In related technologies, rule-based parameter recommendation systems cannot effectively handle multidimensional nonlinear relationships, leading to recommendations that deviate from actual needs. Furthermore, traditional algorithms have limited responsiveness to dynamic environmental changes, making it difficult to meet the high-precision requirements of advanced manufacturing processes for process parameters.

[0054] Based on this, in some embodiments, such as Figure 1 As shown in the figure, this application provides a method for recommending wafer surface treatment processes, which includes the following steps S101 to S103.

[0055] S101: Obtain the surface physicochemical property parameters of the target wafer, and preprocess the surface physicochemical property parameters to obtain the target physicochemical property parameters.

[0056] The surface physicochemical properties of the target wafer can include material chemical properties, roughness, and mechanical properties. After obtaining these properties, they can be normalized for subsequent input into a neural network model or Transformer encoder.

[0057] S102: Input the target physical and chemical property parameters into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for the target wafer in terms of various process parameters.

[0058] It is understandable that a single-hidden-layer neural network model is used to construct a nonlinear mapping relationship between the physical and chemical properties of the wafer and the plasma surface treatment process parameters. Thus, after the target physical and chemical properties are input into the pre-trained single-hidden-layer neural network model, the model can output preliminary recommended parameters for the target wafer in terms of various process parameters.

[0059] It should be noted that the recommended parameters in this application embodiment are feature parameters that have been uniformly normalized or standard encoded. The recommended parameters need to be reverse-mapping processed to restore them to physical quantities that can be executed in actual engineering.

[0060] In applications, process parameters may include vacuum level parameters, type of medium gas, power parameters, processing time parameters, and frequency parameters.

[0061] Vacuum degree parameter (Op1): The actual vacuum degree value (5Pa to 100Pa) is mapped to the [0,1] interval using a linear normalization method to ensure stable neural network output.

[0062] Medium gas type (Op2): Based on the gas type (oxygen, nitrogen, argon), corresponding intervals are assigned, namely [0,0.3], (0.3,0.6] and (0.6,1], respectively, to achieve standardized coding of gas types.

[0063] Power parameter (Op3): The actual power (50W to 350W) is mapped to the [0,1] interval through linear normalization to eliminate the scale inconsistency problem caused by different power levels.

[0064] Processing time parameter (Op4): The actual processing time (15 seconds to 150 seconds) is mapped to the [0,1] interval through linear normalization to ensure the stability of the time feature value.

[0065] Frequency parameters (Op5): Based on the operating frequency of the processing power supply (13.56kHz or 40kHz), the parameters are assigned values ​​in the intervals [0,0.5] and (0.5,1] respectively using an interval mapping method to complete the frequency feature discretization processing.

[0066] For example, the reference range for the output value of the medium gas can be found in Table 1, and the reference range for the output value of the frequency parameter can be found in Table 2.

[0067] Table 1 shows the reference range for the output value of the medium gas.

[0068]

[0069] Table 2 shows the reference range for frequency parameter output values.

[0070]

[0071] S103: The target materialized characteristic parameters and preliminary recommendation parameters are concatenated and input into the Transformer encoder based on the attention mechanism. The multi-head self-attention mechanism is used to model the complex coupling relationship between the input features and extract the deep representation to output the target recommendation parameters.

[0072] It is understandable that the Transformer encoder has the ability to model the complex coupling relationships between input features, and exhibits stronger adaptability and prediction accuracy when facing diverse wafer materials and complex process environments. Therefore, by concatenating the target materialized property parameters and the preliminary recommended parameters and inputting them together into the attention-based Transformer encoder, the original input information and the coarse recommendation results are effectively fused and processed in the Transformer encoder, which improves the contextual understanding ability of the recommended parameters and the accuracy of the final output, thereby further improving the prediction accuracy.

[0073] The aforementioned wafer surface treatment process recommendation method rapidly generates preliminary recommended parameters for various process parameters of the target wafer by inputting the target physical property parameters into a pre-trained single-hidden-layer neural network model. Subsequently, the preliminary recommended parameters and the target physical property parameters are input into an attention-based Transformer encoder. The Transformer encoder's multi-head self-attention mechanism models the complex coupling relationships between input features, extracting deep representations to complete the final recommendation and generate the target recommended parameters. This method uses a neural network model to capture the complex coupling relationships between various process parameters, achieving a nonlinear mapping between the target wafer's surface physical property parameters and process parameters. This enables efficient, accurate, and automated recommendation of wafer plasma surface treatment process parameters. Furthermore, by combining the fast response capability of a single-hidden-layer neural network with the deep modeling capability of the attention-based Transformer structure for complex dependencies in input features, the prediction speed is improved while significantly enhancing the model's adaptability to heterogeneous materials and multi-parameter coupling scenarios. It exhibits stronger adaptability and prediction accuracy when facing diverse wafer materials and complex process environments.

[0074] In some embodiments, based on the above embodiments, such as Figure 2 As shown, the recommended method for wafer surface treatment also includes the following steps S201 and S202.

[0075] S104: Input the target physical and chemical property parameters and the target recommended parameters into the error rate prediction model based on the long short-term memory neural network to obtain the prediction error rate of each process parameter of the target wafer.

[0076] It is understandable that the target physical property parameters and target recommended parameters are input into a pre-built and trained error rate prediction model based on a long short-term memory neural network (LSTM). This error rate prediction model learns and analyzes the time series features and complex correlations in the input data, and outputs the prediction error rate corresponding to each process parameter of the target wafer, thereby achieving the prediction of wafer manufacturing process parameter errors.

[0077] S105: Based on the prediction error rate of each process parameter of the target wafer, calibrate the target recommended parameters and generate calibration recommended parameters.

[0078] It is understandable that after obtaining the prediction error rate of each process parameter of the target wafer, the target recommended parameters can be modified accordingly. That is, the recommended values ​​of process parameters with high prediction error rates can be adjusted significantly, while parameters with low error rates can be fine-tuned or maintained. This generates calibration recommended parameters that take into account the feedback of prediction errors, thereby optimizing the accuracy of process parameter settings.

[0079] In this embodiment, by introducing a dynamic error correction mechanism, efficient, accurate, and automated recommendation of wafer plasma surface treatment process parameters is achieved. Based on the fast response capability of a single hidden layer neural network and the deep modeling capability of the Transformer structure based on the attention mechanism for complex dependencies of input features, the LSTM error rate prediction model, combined with time series error trend analysis, compensates for errors in the target recommendation results, effectively improving the accuracy of process parameter settings.

[0080] In some embodiments, based on the previous embodiment, such as Figure 3 As shown, the recommended method for wafer surface treatment also includes the following steps S301 to S304.

[0081] S301: Select the reference wafer with the most similar physical properties from the wafer database based on the target physical property parameters.

[0082] Specifically, for the target wafer, its surface physicochemical properties are used as the search benchmark. The most similar sample in physicochemical properties is retrieved from a mature wafer database. This most similar wafer is then used as the reference wafer, and its surface physicochemical properties and process parameters can be obtained. It should be noted that similarity assessment can employ a nearest neighbor search method based on Euclidean distance to ensure that the reference wafer and the target wafer have highly similar physicochemical properties.

[0083] S302: Based on the physical and chemical properties parameters of the reference wafer, a single hidden layer neural network model, a Transformer encoder, and an error rate prediction model, obtain the target predicted error rate of the reference wafer in terms of various process parameters.

[0084] The process can be understood as follows: The physical and chemical properties of a reference wafer are input into a single-hidden-layer neural network model, which then outputs preliminary recommended parameters for the reference wafer. These preliminary recommended parameters and the physical and chemical properties are then input into a Transformer encoder, which outputs the target recommended parameters for the reference wafer. After obtaining the target recommended parameters, these parameters are input into an error rate prediction model. Based on the parameters output by the error rate prediction model, the target predicted error rate for the reference wafer across various process parameters is obtained.

[0085] S303: Obtain the first error rate difference between the target predicted error rate and the actual error rate of each process parameter of the reference wafer.

[0086] It is understandable that since the actual process parameters of the reference wafer are known, the target recommended parameters of the reference wafer are compared with the actual process parameters to determine the actual error rate of each process parameter of the reference wafer. The difference between the target predicted error rate and the actual error rate can be used to obtain the first error rate difference value.

[0087] S304: Compensate the calibration recommended parameters based on the first error rate difference to generate compensated recommended parameters.

[0088] For example, the target recommended parameter for the reference wafer is G. Op1_ref -G Op5_ref Using the target recommended parameters and physicochemical properties of the reference wafer as input, the error rate of predicting each process parameter of the reference wafer using an LSTM model is E. pred_ref1 ~E pred_ref5 Then calculate the actual error rate (E) of the reference wafer. real_ref After that, the difference between the actual error and the predicted error of each process parameter (ΔE) can be further calculated. i The corresponding formula is as follows:

[0089] ;

[0090] ;

[0091] in, This represents the actual process parameter Opi.

[0092] For the target wafer, input its target recommended parameters (G) Op1_tar -G Op5_tar The material properties are fed into the LSTM model to predict the target prediction error rate (E). pred_tar1 - E pred_tar5 The recommended calibration parameters are: .

[0093] Based on the difference ΔEi obtained from the reference wafer, the calibration recommended parameters are compensated according to the first error rate difference to obtain the compensation recommended parameters. The corresponding formula is as follows:

[0094] .

[0095] It should be noted that in application, the target prediction error rate of the target wafer can be compensated first based on the first error rate difference. Then, the target recommended parameters of the target wafer can be corrected based on the compensated error rate, finally obtaining the compensated and corrected wafer surface plasma treatment process parameters, which can be used as a recommendation for practical application. The corresponding formula is as follows:

[0096] ;

[0097] .

[0098] In this embodiment, by introducing a similar wafer difference correction strategy, the first error rate difference between the target predicted error rate and the actual error rate of each process parameter of the reference wafer is obtained. The calibration recommended parameters are compensated based on the first error rate difference to generate compensated recommended parameters, thereby performing secondary optimization on the target recommended parameters, thereby effectively controlling the error rate of the recommended parameters and improving the reliability of the final recommended parameters.

[0099] In some embodiments, the surface physicochemical property parameters include material chemical property parameters, roughness, and mechanical property parameters. Step S101 involves preprocessing the surface physicochemical property parameters to obtain target physicochemical property parameters, including: linearly normalizing the material chemical property parameters to obtain a first feature parameter; performing quantile normalization on the roughness and mechanical property parameters respectively to obtain a second feature parameter and a third feature parameter; and concatenating the first feature parameter, the second feature parameter, and the third feature parameter into an input feature matrix to obtain the target physicochemical property parameters.

[0100] X1 represents the chemical properties of the material, X2 represents the roughness (the roughness of the wafer surface to be treated), and X3 represents the mechanical properties (considering local flatness, curvature, and warpage). The chemical properties (X1) are relatively stable and can be processed using simple normalization. Roughness (X2) and mechanical properties (X3) are mapped to a uniform distribution using quantile normalization.

[0101] First, regarding the material chemical property parameter X1, this application uses the frequency of wafer elements in the semiconductor industry as the contribution and employs a linear normalization (Min-Max Normalization) method to linearly map it to the [0,1] interval. For example, the contribution of common elements can be shown in Table 3.

[0102] Table 3 shows the contribution of common elements.

[0103]

[0104] That is, based on the minimum and maximum contribution values ​​of elements in the training data, normalization is performed according to the following formula:

[0105]

[0106] The contribution of common elements can be recorded in Table 1. Based on this, the average contribution of the elements contained in the chemical formula is calculated after normalization according to the specific chemical formula of the wafer. The calculated value is the X1 value. The X1 values ​​of common wafer materials can be found in Table 4.

[0107] Table 4 shows the X1 values ​​for common wafer materials.

[0108]

[0109] The above processing preserves the relative chemical activity differences between materials and unifies the numerical scale, which helps improve the network convergence speed and training stability. For extreme anomalous material samples, if the X1 value exceeds the statistical range of the training set, interval truncation is also used during the inference stage to ensure that the input value after normalization is stable within the range of [0,1].

[0110] Subsequently, for X2 (wafer surface roughness index) and X3 (mechanical characteristic parameters weighted by local flatness, curvature, and warpage), this application introduces a preprocessing strategy based on quantile normalization. Specifically, cumulative distribution functions are fitted to the X2 and X3 features, the cumulative probability of each sample under the corresponding feature value ranking is calculated, and its cumulative probability is mapped to a uniform distribution in the [0,1] interval. This processing method can effectively suppress skewed distributions, extreme outliers, and scale inconsistencies in the original features, making the processed X2 and X3 more uniform in numerical distribution, further improving the adaptability of the neural network to the input features. For example, the parameters after roughness normalization can be found in Table 5, and the corresponding values ​​of each X3 index can be found in Table 6.

[0111] Table 5 shows the parameters after roughness normalization.

[0112]

[0113] Table 6 shows the corresponding values ​​for each indicator of X3.

[0114]

[0115] After normalization, the linearly normalized X1 is concatenated with the quantile-normalized X2 and X3 to form a unified and standardized new input feature matrix, which serves as the input layer feature of the neural network model.

[0116] During the model inference phase, to maintain consistency between the training and application environments, all new input data can be processed using the normalized parameters and quantile mapping rules fitted during the training phase. This avoids fluctuations in model performance due to differences in data distribution and ensures the reliability and stability of the process parameters output by the recommendation system.

[0117] In this embodiment, by introducing the above-mentioned complete feature preprocessing process, this application can systematically optimize the distribution pattern and scale relationship of each input feature, significantly accelerate the training convergence speed of the neural network model, improve the accuracy and generalization ability of the wafer surface processing process parameter recommendation, and further enhance the application adaptability of the method of this application under heterogeneous materials and complex process conditions.

[0118] In some embodiments, a single-hidden-layer neural network model includes an input layer, a hidden layer, and an output layer.

[0119] Among them, such as Figure 4 As shown, the number of neurons in the input layer is the same as the number of parameter types included in the surface physical properties parameters. The hidden layer consists of multiple neurons and uses a nonlinear function as the activation function. The number of neurons in the output layer is the same as the number of parameter types included in the process parameters, and the output layer activation function is either a sigmoid activation function or a linear activation function. The single-hidden-layer neural network model is trained using a sample set based on historical wafer processing experimental data to adjust the connection weights and bias parameters between the hidden and output layers.

[0120] In this embodiment, the specific network design of the single hidden layer neural network structure can be as follows:

[0121] Input layer: Three neurons are set up, corresponding to the normalized material chemical property parameters (X1), surface roughness parameters (X2), and mechanical property parameters (X3), respectively.

[0122] Hidden layer: A single hidden layer is set, containing 4 neurons (Y11, Y12, Y13, Y14). The activation function is a non-linear function (such as ReLU or Tanh) to enhance the model's expressive power.

[0123] Hidden layer calculation:

[0124] ;

[0125] in, , , This represents the weight parameters between the input layer and the hidden layer. This represents the bias parameter between the input layer and the hidden layer.

[0126] Output layer: Set 5 neurons, corresponding to five normalized output process parameters (Op1 to Op5). The output layer activation function is set to Sigmoid or linear activation as needed to ensure that the output falls within the specified range.

[0127] Output layer computation:

[0128] ;

[0129] in, This represents the weight parameters between the output layer and the hidden layer. This represents the bias parameter between the input layer and the hidden layer.

[0130] The connection weights and bias parameters between the hidden layer and the output layer are dynamically adjusted during training. Specifically, training utilizes a sample set based on historical wafer processing experimental data. Optimization methods include standard deep learning algorithms such as gradient descent (SGD) and the Adam optimizer, aiming to minimize the mean squared error (MSE) between the output prediction and the target process parameters. This process yields a pre-trained single-hidden-layer neural network model. The target physical and chemical properties parameters of the target wafer can then be input into this model to obtain preliminary recommended parameters for various process parameters of the target wafer.

[0131] In some embodiments, step S103 involves concatenating the target materialized feature parameters and the preliminary recommendation parameters and inputting them into a Transformer encoder based on an attention mechanism. A multi-head self-attention mechanism is used to model the complex coupling relationship between the input features, extracting a deep representation to output the target recommendation parameters. This includes: concatenating the target materialized feature parameters and the preliminary recommendation parameters to form a multi-dimensional joint input vector; mapping the multi-dimensional joint input vector to the high-dimensional feature space required by the Transformer encoder through a linear embedding layer; processing the embedded features in the high-dimensional feature space through the Transformer encoder to generate a highly expressive hidden representation that integrates multi-source features, and mapping it to the target recommendation parameters through a regression head.

[0132] Specifically, the normalized original input feature vector X=[X1,X2,X3] is first compared with the preliminary recommended parameter vector G output by the first-stage single-hidden-layer neural network. Op =[G Op1 ′,G Op2 ′,G Op3 ′,G Op4 ′,G Op5The feature-level concatenation of '] forms an 8-dimensional joint input vector:

[0133] Xcombined=[X1,X2,X3,G Op1 ′,G Op2 ′,G Op3 ′,G Op4 ′,G Op5 ′).

[0134] Subsequently, this invention introduces a linear embedding layer to map the 8-dimensional joint vector to the high-dimensional feature space required by the Transformer encoder, thereby enhancing feature representation and unifying the input scale.

[0135] h0 = We⋅Xcombined +be;

[0136] Where h0 represents the embedded feature, We∈Rd×8 is the embedding weight matrix, be∈Rd is the bias term, and d is the embedding dimension (128-dimensional) used internally by the Transformer module. This embedding layer can be regarded as a mapping mechanism from low-dimensional raw data to high-dimensional context space, ensuring that the subsequent self-attention module can make full use of the correlation between all input features.

[0137] After linear embedding is completed, h0 is input as a single token into the Transformer encoder. It is processed by sub-modules such as multi-head self-attention mechanism, residual connection, layer normalization and feedforward network in sequence, and finally generates a hidden representation with high expressive power that integrates multi-source features. It is then mapped to the optimized plasma processing process parameters [Op1, Op2, Op3, Op4, Op5] through the regression head.

[0138] Through the above splicing, fusion, and embedding strategies, this application achieves effective fusion processing of the original input information and coarse recommendation results in the Transformer module without adding extra complex structures, thereby improving the contextual understanding ability of the recommendation parameters and the accuracy of the final output.

[0139] In this embodiment, a simple single-hidden-layer neural network is used to quickly map the standardized wafer physicochemical properties of the input to generate preliminary recommended parameters. This stage has advantages such as lightweight model, fast response, and wide adaptability to material variations. Then, a multi-layer stacked Transformer encoder structure is used to concatenate the preliminary recommended parameters with the original input features and input them into the encoder. A multi-head self-attention mechanism is used to model the high-order dependencies between features, dynamically allocating the importance of each input dimension in the global context, and extracting more representative composite feature representations. The Transformer encoder contains a self-attention sub-layer, a feedforward neural network sub-layer, residual connections, and a layer normalization structure, which can perform deep correction and optimization on the coarse recommendation results, thereby outputting the final optimized five plasma processing parameters (Op1 to Op5), i.e., the target recommended parameters, including recommended vacuum level, gas type, processing power, processing time, and frequency.

[0140] It should be noted that this fusion structure can employ a two-stage training approach: first, the initial recommendation network (i.e., a single-hidden-layer neural network model) is trained independently, and then its output is used as a fixed input to the Transformer encoder for optimization training. Alternatively, an end-to-end joint training approach can be used, simultaneously optimizing the weight parameters of both stages to achieve optimal overall performance. While retaining the rapid response advantage of traditional networks, this model further leverages the Transformer encoder's ability to model complex coupling relationships between input features, exhibiting stronger adaptability and prediction accuracy when facing diverse wafer materials and complex process environments.

[0141] In some embodiments, step S104, inputting the target physical and chemical characteristic parameters and the target recommended parameters into an error rate prediction model based on a long short-term memory neural network to obtain the predicted error rate of each process parameter of the target wafer, includes: modeling the error prediction of the mature wafer sample as a time series process, defined as a continuous Ti sequence step error prediction task; inputting the target recommended parameters and physical and chemical characteristic parameters of the mature wafer sample into the error rate prediction model, obtaining the predicted error rate of each sequence step output by the error rate prediction model, and obtaining the second error rate difference between the predicted error rate of each sequence step and the actual error rate; forming a cumulative correction trend based on the second error rate difference corresponding to each sequence step; inputting the target physical and chemical characteristic parameters and the target recommended parameters into the error rate prediction model based on a long short-term memory neural network to output the preliminary predicted error rate of each process parameter; and using the cumulative correction trend to correct the preliminary predicted error rate of each process parameter of the target wafer to obtain the target predicted error rate of each process parameter.

[0142] In applications, a large amount of historical data on mature wafer types can be collected. This historical data can include wafer physical and chemical property parameters (X1, X2, X3), verified mature process parameters (Op1 to Op5), and target recommended parameters (G) generated based on the model in the aforementioned embodiments. Op1_t To G Op5_t ).

[0143] After collecting historical data on a large number of mature wafer types, an error rate prediction model can be built and trained based on the historical data. First, an LSTM network structure can be designed, with the input features being the features at each time step t, and the initially generated wafer process parameters (G) as input. Op1_t ~G Op5_t The network architecture employs two stacked LSTM units, with 64 hidden nodes per layer, to fully capture the short-term and long-term dependencies in the error evolution process. The output feature at each time step t is the prediction error rate (E) for five process parameters. pred1_t ~ E pred5_t The internal structure of LSTM introduces forget gates, input gates, and output gates. The forget gate determines which historical error information to retain or discard based on the historical memory state and the current input; the input gate determines which new error information to record based on the current input error information; and the output gate controls the error rate prediction output at the current time step.

[0144] Error prediction for each mature wafer sample is modeled as a time series process, defined as a continuous Ti-step (T1, T2, ..., Ti) error prediction task. For each step t, preliminary recommended parameters and wafer physicochemical properties are input, and the LSTM outputs the prediction error rate E for that step. pred_t Record the actual error rate E real_t (Calculated from recommended and mature parameters, the Ti step uses mature wafers as the target. In the Ti+1 step, when targeting the target wafer, the cumulative correction trend ΔEsum is used for compensation and correction. ΔE is updated after the target wafer has been actually processed and verified to obtain new real process parameters.) avg ), calculate the difference ΔE between the predicted error rate and the actual error rate. t = E real_t -E pred_t After each step, the internal state is updated, enabling the LSTM to dynamically adjust the memory content based on the historical prediction-actual difference in the next time step (t+1).

[0145] After the historical error prediction in step Ti is completed, based on the accumulated ΔE i-10 , ΔE i-9 , ..., ΔE iThis forms a cumulative correction trend ΔEsum. Upon entering step Ti+1 (i.e., when predicting the target wafer), the initial recommended parameters and physical properties of the target wafer are input, and the LSTM predicts the initial error rate E for step Ti+1. pred _ Ti+1 Using the cumulative corrected trend ΔEsum to adjust E pred_Ti+1 Compensation is performed to obtain the target prediction error rate, as shown in the following formula:

[0146] .

[0147] Based on the compensated target prediction error rate E corrected_Ti+1 It can correct the target recommended parameters for the target wafer and calculate the compensation recommended parameters:

[0148] .

[0149] In this embodiment, the target prediction error rate is calculated in the above manner. The error rate prediction model not only considers the current parameter state, but also combines historical error patterns, thereby achieving more accurate error prediction and facilitating the acquisition of more accurate compensation recommendation parameters.

[0150] Based on the same inventive concept, in some embodiments, such as Figure 5 As shown, this application also provides a wafer surface treatment process recommendation apparatus 50, including: an acquisition module 51, a first recommendation module 52, and a second recommendation module 53; wherein,

[0151] The acquisition module 51 is used to acquire the surface physical and chemical properties parameters of the target wafer and preprocess the surface physical and chemical properties parameters to obtain the target physical and chemical properties parameters;

[0152] The first recommendation module 52 is used to input the target physical property parameters into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for the process parameters;

[0153] The second recommendation module 53 is used to concatenate the target materialized characteristic parameters and the preliminary recommendation parameters and input them into the Transformer encoder based on the attention mechanism. It uses a multi-head self-attention mechanism to model the complex coupling relationship between the input features and extracts deep representations to output the target recommendation parameters.

[0154] In some embodiments, the wafer surface treatment process recommendation apparatus 50 further includes a first error prediction module and a calibration module. The first error prediction module is used to input the target physicochemical characteristic parameters and the target recommended parameters into an error rate prediction model based on a long short-term memory neural network to obtain the predicted error rate of each process parameter of the target wafer. The calibration module is used to calibrate the target recommended parameters according to the predicted error rate of each process parameter of the target wafer, generating calibrated recommended parameters.

[0155] In some embodiments, the wafer surface treatment process recommendation method further includes: a retrieval module, a second error prediction module, a difference calculation module, and a compensation module. The retrieval module is used to retrieve the reference wafer with the most similar physical properties from the wafer database based on the target physical property parameters; the second error prediction module is used to obtain the target predicted error rate of the reference wafer for each process parameter based on the physical property parameters of the reference wafer, a single hidden layer neural network model, a Transformer encoder, and an error rate prediction model; the difference calculation module is used to obtain a first error rate difference between the target predicted error rate and the actual error rate of each process parameter of the reference wafer; and the compensation module is used to compensate the calibration recommendation parameters based on the first error rate difference to generate compensated recommendation parameters.

[0156] In some embodiments, the acquisition module 51 is further configured to perform linear normalization on the material chemical property parameters to obtain the first feature parameter; perform quantile normalization on the roughness and mechanical property parameters respectively to obtain the second feature parameter and the third feature parameter; and concatenate the first feature parameter, the second feature parameter and the third feature parameter into an input feature matrix to obtain the target physical and chemical property parameters.

[0157] In some embodiments, the second recommendation module 53 is further configured to concatenate the target materialized characteristic parameters and the preliminary recommendation parameters to form a multidimensional joint input vector; map the multidimensional joint input vector to the high-dimensional feature space required by the Transformer encoder through a linear embedding layer; process the embedded features of the high-dimensional feature space through the Transformer encoder to generate a highly expressive hidden representation that integrates multi-source features, and map it to the target recommendation parameters through a regression head.

[0158] In some embodiments, the first error prediction module is further configured to: model the error prediction of mature wafer samples as a time series process, defined as a continuous Ti sequence step error prediction task; input the target recommended parameters and physicochemical property parameters of the mature wafer samples into the error rate prediction model, obtain the prediction error rate of each sequence step output by the error rate prediction model, and obtain the second error rate difference between the prediction error rate and the actual error rate of each sequence step; form a cumulative correction trend based on the second error rate difference corresponding to each sequence step; input the target physicochemical property parameters and target recommended parameters into the error rate prediction model based on a long short-term memory neural network, and output the preliminary prediction error rate of each process parameter; use the cumulative correction trend to correct the preliminary prediction error rate of each process parameter of the target wafer, and obtain the target prediction error rate of each process parameter.

[0159] It should be noted that the wafer surface treatment process recommendation apparatus 50 provided in this application embodiment and the wafer surface treatment process recommendation method provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned wafer surface treatment process recommendation method, and the repeated parts will not be described again.

[0160] In some embodiments, this application also provides an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the recommended method for wafer surface treatment process of any of the above schemes.

[0161] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0162] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0163] This application also provides a computer-readable medium storing a computer program thereon, which, when executed by a processor, implements the recommended method for wafer surface treatment process of any of the above-described schemes. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.

[0164] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0165] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A recommended method for wafer surface treatment, characterized in that, include: The surface physicochemical properties parameters of the target wafer are obtained, and the surface physicochemical properties parameters are preprocessed to obtain the target physicochemical properties parameters; The target physical and chemical properties parameters are input into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for the target wafer in terms of various process parameters; The target materialized characteristic parameters and the preliminary recommendation parameters are concatenated and input into the attention-based Transformer encoder. The multi-head self-attention mechanism is used to model the complex coupling relationship between the input features and extract the deep representation to output the target recommendation parameters. The target physical property parameters and the target recommended parameters are input into an error rate prediction model based on a long short-term memory neural network to obtain the prediction error rate of each process parameter of the target wafer; The target recommended parameters are calibrated based on the prediction error rate of each process parameter of the target wafer to generate calibrated recommended parameters.

2. The recommended method for wafer surface treatment as described in claim 1, characterized in that, The recommended wafer surface treatment process also includes: Based on the target physical and chemical property parameters, determine the reference wafer with the most similar physical and chemical properties from the wafer database; Based on the physical and chemical properties parameters of the reference wafer, the single hidden layer neural network model, the Transformer encoder, and the error rate prediction model, the target prediction error rate of the reference wafer in terms of each process parameter is obtained. Obtain the first error rate difference between the target predicted error rate and the actual error rate of each process parameter of the reference wafer; The calibration recommendation parameters are compensated based on the first error rate difference to generate compensated recommendation parameters.

3. The recommended method for wafer surface treatment as described in claim 1, characterized in that, The surface physicochemical properties parameters include material chemical properties parameters, roughness, and mechanical properties parameters; the preprocessing of the surface physicochemical properties parameters to obtain the target physicochemical properties parameters includes: The chemical property parameters of the material are linearly normalized to obtain the first characteristic parameter; The roughness and the mechanical property parameters are respectively normalized by quantiles to obtain the second characteristic parameter and the third characteristic parameter; The first feature parameter, the second feature parameter, and the third feature parameter are concatenated into an input feature matrix to obtain the target materialization property parameters.

4. The recommended method for wafer surface treatment as described in claim 1, characterized in that, The single-hidden-layer neural network model includes: an input layer, a hidden layer, and an output layer; The number of neurons in the input layer is the same as the number of parameter types included in the surface physical properties parameters; The hidden layer comprises multiple neurons and uses a nonlinear function as the activation function; The number of neurons in the output layer is the same as the number of parameter types included in the process parameters, and the activation function of the output layer is either a sigmoid activation function or a linear activation function. The single-hidden-layer neural network model is trained using a sample set based on historical wafer processing experimental data to adjust the connection weights and bias parameters between the hidden layer and the output layer.

5. The recommended method for wafer surface treatment as described in any one of claims 1 to 4, characterized in that, The process of concatenating the target materialized feature parameters and the preliminary recommendation parameters and inputting them into a Transformer encoder based on an attention mechanism, utilizing a multi-head self-attention mechanism to model the complex coupling relationship between input features, and extracting deep representations to output target recommendation parameters includes: The target materialized characteristic parameters and the preliminary recommended parameters are concatenated to form a multi-dimensional joint input vector; The multidimensional joint input vector is mapped to the high-dimensional feature space required by the Transformer encoder through a linear embedding layer; The Transformer encoder processes the embedded features in the high-dimensional feature space to generate a highly expressive hidden representation that integrates multi-source features, and then maps it to the target recommendation parameters through a regression head.

6. The recommended method for wafer surface treatment as described in claim 1, characterized in that, The step of inputting the target physical property parameters and the target recommended parameters into an error rate prediction model based on a long short-term memory neural network to obtain the prediction error rate of each process parameter of the target wafer includes: Error prediction of mature wafer samples is modeled as a time series process and defined as a continuous Ti sequence step error prediction task; The target recommended parameters and physicochemical property parameters of the mature wafer sample are input into the error rate prediction model to obtain the prediction error rate of each sequence step output by the error rate prediction model, and the second error rate difference between the prediction error rate and the actual error rate of each sequence step is obtained. A cumulative correction trend is formed based on the difference in the second error rate corresponding to each sequence step; The target physical property parameters and the target recommended parameters are input into the error rate prediction model based on the long short-term memory neural network, and the preliminary prediction error rate of each process parameter is output. The cumulative correction trend is used to correct the initial prediction error rate of each process parameter of the target wafer, so as to obtain the target prediction error rate of each process parameter.

7. A recommended apparatus for wafer surface treatment process, characterized in that, include: The acquisition module is used to acquire the surface physicochemical property parameters of the target wafer and preprocess the surface physicochemical property parameters to obtain the target physicochemical property parameters. The first recommendation module is used to input the target physical property parameters into a pre-trained single-hidden-layer neural network model to obtain preliminary recommended parameters for the process parameters. The second recommendation module is used to concatenate the target materialized characteristic parameters and the preliminary recommendation parameters and input them into the Transformer encoder based on the attention mechanism. It uses a multi-head self-attention mechanism to model the complex coupling relationship between the input features and extracts deep representations to output the target recommendation parameters. The first error prediction module is used to input the target physical and chemical characteristic parameters and the target recommended parameters into an error rate prediction model based on a long short-term memory neural network to obtain the prediction error rate of each process parameter of the target wafer. The calibration module is used to calibrate the target recommended parameters based on the prediction error rate of each process parameter of the target wafer, and generate calibration recommended parameters.

8. An electronic device comprising a processor and a memory; said memory having a storage for a computer program, wherein, When executed by the processor, the computer program implements the recommended method for wafer surface treatment process as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the recommended method for wafer surface treatment process as described in any one of claims 1 to 6.

Images