Precision etching process parameter intelligent optimization method and system based on deep learning

Abstract

The invention relates to the technical field of semiconductor manufacturing, and discloses a precision etching process parameter intelligent optimization method and system based on deep learning, and the method comprises the steps: firstly constructing a static basic process model; during online operation, a process drift state vector is generated by calculating a prediction residual error between an actual etching result and a static model reference prediction result and performing statistical coding on a historical residual error sequence; inputting the state vector into a submerged space correction operator to generate a submerged space disturbance vector and a correction confidence score; the submerged space disturbance vector is used for constructing a dynamic correction prediction model reflecting a current equipment state, and the correction confidence score is used for dynamically setting an exploration-utilization balance strategy of an optimization engine; and an optimization engine calls the dynamic model to execute search and outputs optimal process parameters. According to the method, the basic model and the drift correction are decoupled, the process drift compensation is realized through the confidence-driven optimization strategy, and the accuracy of the optimization result is improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, specifically to a method and system for intelligent optimization of precision etching process parameters based on deep learning. Background Technology

[0002] Plasma etching is one of the core processes in integrated circuit manufacturing, used to precisely etch micro- and nano-scale patterns onto the wafer surface, such as the gate structure of transistors. The quality of the etching result, including critical dimensions, cross-sectional morphology, and uniformity, directly affects the performance and yield of the final chip. This process is influenced by dozens of process parameters, such as gas flow rate, RF power, and chamber pressure, and these parameters have complex nonlinear relationships with the etching result. Therefore, establishing an accurate process model to guide the optimization of process parameters is crucial for achieving precise process control.

[0003] In actual production, the condition of etching equipment changes as the number of wafers processed increases. Factors such as deposit accumulation on the inner walls of the reaction chamber and component aging can cause process characteristics to deviate from their initial baseline state; this phenomenon is commonly referred to as process drift. Static process models built based on data collected under the equipment's baseline state gradually lose their predictive accuracy due to process drift. If such models continue to be used for optimization, the output process parameters will be difficult to achieve the expected process targets in actual production, leading to a decrease in product yield.

[0004] To address the problems caused by process drift, existing technologies typically employ periodic recalibration of the entire process model. This approach requires interrupting normal production and conducting extensive process experiments to collect new datasets, resulting in long equipment downtime and high production costs. Other methods attempt to perform simple compensation corrections to the model using online data, but these often fail to effectively capture the complex dynamics and multivariate coupling characteristics of the drift process. Furthermore, directly using a small amount of online data to fine-tune the entire model may lead to the model forgetting the fundamental process rules learned from large amounts of offline data—a problem known as "catastrophic forgetting"—affecting the model's stability and generalization ability. Therefore, developing a process parameter optimization method that can compensate for process drift in real time and accurately while avoiding large-scale model recalibration is a pressing technical challenge in this field. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that, in the prior art, the precision etching process model is usually a static model, which is difficult to adapt in real time to the state drift caused by factors such as equipment aging during the production process. This results in low accuracy and poor timeliness of the process parameter optimization results, and the cost of frequent global model recalibration is high.

[0006] To address the aforementioned technical problems, the first aspect of this invention provides a method for intelligent optimization of precision etching process parameters based on deep learning. This method constructs and solidifies a static basic process model describing the process behavior of the equipment under a baseline state. During online system operation, it continuously monitors the residual between the actual production results and the baseline prediction results of this static model, thereby quantifying the drift state of the equipment.

[0007] This method statistically encodes historical residual sequences to form a process drift state vector characterizing the current drift amplitude, internal correlations, and dynamic trends. This vector is then input into an online trainable latent space correction operator to generate a perturbation vector acting on the latent space of the static basic process model. By correcting at the latent space level, the behavior of the basic model is adjusted, thereby constructing a dynamic correction prediction model that reflects the current equipment state.

[0008] This method generates a modified confidence score simultaneously with the latent space perturbation vector, and uses this score to dynamically adjust the search strategy of the optimization process. When the modified confidence score indicates high reliability of the modification, the optimization engine adopts an exploit-based search strategy for local optimization; when the score indicates low reliability of the modification, the optimization engine adopts an exploration-based search strategy to cover a wider parameter space.

[0009] Through the above mechanism, the method of the present invention achieves compensation for process drift while maintaining the stability of the static basic model, and ensures the effectiveness of the optimization results through a confidence-aware optimization strategy.

[0010] A second aspect of this invention provides a deep learning-based intelligent optimization system for precision etching process parameters, which realizes all the technical concepts of the aforementioned methods. The system includes:

[0011] The static basic process model module contains a static basic process model and is configured to output predicted etching results based on the input process parameter vector and latent space vector.

[0012] The online drift sensing and correction module is configured to acquire the actual production process parameters and actual etching results, calculate the prediction residual with the prediction results of the static basic process model, generate a process drift state vector based on the historical prediction residual sequence, and generate a latent space perturbation vector and a corrected confidence score based on the state vector.

[0013] The dynamic correction prediction module is configured to apply the latent space disturbance vector to the static basic process model to construct a dynamic correction prediction model.

[0014] The confidence-driven optimization engine module is configured to set its exploration-utilization balance strategy based on the modified confidence score and invoke the dynamically modified prediction model to perform optimization search to output optimal process parameters.

[0015] This invention provides a method and system for intelligent optimization of precision etching process parameters based on deep learning. It has the following beneficial effects:

[0016] 1. This invention constructs and solidifies a static basic process model while simultaneously monitoring and predicting residuals online to generate latent space perturbation vectors, thereby building a dynamically corrected prediction model in real time. This mechanism separates the basic law model from the time-varying drift correction, allowing the system to continuously adapt to changes in equipment status without interrupting production for large-scale experiments and model recalibration, significantly reducing equipment downtime and lowering maintenance costs.

[0017] 2. The optimization process of this invention is based on a dynamic correction prediction model, which is formed by applying a latent space perturbation vector representing the current drift state to a fixed static basic process model. Compared to compensation at the output end, correction at the latent space level can more accurately reflect the impact of equipment drift on the intrinsic correlation of the process, thereby enabling the optimization engine to search based on more accurate process response predictions and improving the accuracy of the final output optimal process parameters.

[0018] 3. This invention introduces the concept of a corrected confidence score and dynamically sets the exploration-utilization balance parameter of the optimization engine based on this score, achieving adaptive adjustment of the optimization strategy. When the system has high confidence in the drift correction, the optimization process focuses on utilization to accelerate convergence; when the confidence is low, it focuses on exploration to avoid getting trapped in local optima caused by model uncertainty. This mechanism makes the optimization process more robust to potential biases in model predictions and improves the efficiency of optimization. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the functional modules of a precision etching process parameter intelligent optimization system according to an embodiment of the present invention;

[0020] Figure 2 This is a flowchart of a deep learning-based intelligent optimization method for precision etching process parameters, according to an embodiment of the present invention.

[0021] Among them, 100 is the static basic process model module; 200 is the online drift perception and correction module; 300 is the dynamic correction and prediction module; and 400 is the confidence-driven optimization engine module. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0023] See attached document Figure 1 , Figure 1 This is a schematic diagram of the functional modules of a precision etching process parameter intelligent optimization system according to an embodiment of the present invention. The present invention provides a precision etching process parameter intelligent optimization system, which is deployed in a central control computer and configured to communicate with precision etching equipment and its data acquisition interface. The system may include:

[0024] The static basic process model module 100 contains a static basic process model that characterizes the fundamental laws of the etching process. This module is configured to output the corresponding predicted etching results based on the input process parameters and the latent space vector used to adjust the model output.

[0025] An online drift sensing and correction module 200 is connected to the static basic process model module 100. This module is configured to acquire process parameters and etching results from actual production, calculate the residuals between these parameters and the predictions from the static basic process model, and generate a process drift state vector by extracting statistical features from historical residual sequences. Furthermore, based on this state vector, the module generates a latent space perturbation vector for correcting the static basic process model and simultaneously outputs a correction confidence score, which quantifies the uncertainty of this correction.

[0026] The dynamic correction prediction module 300 connects the static basic process model module 100 and the online drift sensing and correction module 200. This module is configured to invoke the static basic process model when performing parameter optimization tasks, and to construct a dynamic correction prediction model reflecting the current state of the equipment in real time by combining the latest latent space disturbance vector generated by the online drift sensing and correction module 200 with the reference latent space vector.

[0027] The confidence-driven optimization engine module 400 connects the online drift sensing and correction module 200 and the dynamic correction prediction module 300. This module is configured to receive an externally set optimization objective and obtain the correction confidence score generated by the online drift sensing and correction module 200. Based on this score, the module dynamically sets the exploration of its internal optimization algorithm—utilizing a balancing strategy—and calls the prediction model provided by the dynamic correction prediction module 300 to search in the process parameter space, ultimately outputting the optimal process parameters.

[0028] See attached document Figure 2 , Figure 2This is a flowchart of a deep learning-based intelligent optimization method for precision etching process parameters according to an embodiment of the present invention. The present invention provides an intelligent optimization method for precision etching process parameters, which includes the following steps:

[0029] See attached document Figure 2 , Figure 2 This is a flowchart of a deep learning-based intelligent optimization method for precision etching process parameters according to an embodiment of the present invention. The present invention provides an intelligent optimization method for precision etching process parameters, which includes the following steps:

[0030] S101, under the equipment's baseline state, collects multiple sets of process parameters and their corresponding etching results to form an offline training dataset.

[0031] S102 uses an offline training dataset to train a variational autoencoder model containing an encoder and a decoder.

[0032] S103. After training is completed, the network weights of the decoder part in the variational autoencoder model are solidified, and the decoder is used as a static basic process model.

[0033] S104, after the system goes online, acquires the process parameters and corresponding actual etching results for each actual production.

[0034] S105, call the static basic process model, input the actual production process parameters and the preset reference latent space vector, obtain the reference prediction result, and calculate the prediction residual between it and the actual etching result.

[0035] S106, Based on the historical prediction residual sequence within the preset sliding window, the process drift state vector is generated by extracting its statistical features.

[0036] S107, the process drift state vector is input into an online trainable model that serves as a latent space correction operator, which generates a latent space perturbation vector and a corrected confidence score. This latent space correction operator uses the data points obtained in step S104 to update its parameters.

[0037] S108, Construct a dynamic correction prediction model. This model is formed by summing the latent space disturbance vector and the reference latent space vector and using it as the latent space input of the static basic process model.

[0038] S109, When an optimization task is received, the exploration-utilization balance parameter of the optimization engine is determined based on the corrected confidence score.

[0039] S110, the optimization engine calls the dynamic correction prediction model and performs a search based on the exploration-balance parameters determined in step S109 until the termination condition is met, and outputs the optimal process parameters.

[0040] To further clarify the present invention, the specific implementation methods of each module of the above system and each step of the method will be described in detail below.

[0041] In step S101, to construct a static basic process model, an offline training dataset is first collected using experimental design methods, assuming the etching equipment is in a stable and repeatable baseline state. This dataset consists of... The process data consists of a vector of input process parameters for each set of data. and the corresponding etching result vector .

[0042] Process parameter vector This can include process settings such as gas flow rate, RF power, chamber pressure, and etching time. Etching result vector. Performance metrics can include etching rate, sidewall angle, linewidth loss, uniformity, and other parameters.

[0043] In step S102, in order to effectively learn the complex nonlinear mapping relationship between process parameters and etching results, and to obtain a latent space with good structure and controllability, this embodiment of the invention uses a variational autoencoder (VAE) architecture for model training. The model includes an encoder and a decoder.

[0044] The encoder is a deep neural network whose function is to process the etched result of the input. Mapped to latent space The probability distribution parameters. In one embodiment, the encoder consists of multiple fully connected layers, and its output layer outputs two vectors: the mean vector of the latent space distribution. Sum of logarithmic variance vector .

[0045] The decoder is also a deep neural network, and its function is to process process parameters. and vectors sampled from latent space As input, the predicted etching result is reconstructed. Once trained, this decoder becomes the core of the static basic process model module 100.

[0046] The training process of a variational autoencoder involves minimizing a loss function using an optimizer (such as the Adam optimizer). To adjust the network parameters of the encoder and decoder and The loss function Reconstruction loss and KL divergence loss Composed of two weighted components:

[0047] ;

[0048] in, It is a weight hyperparameter used to balance the two parts of the loss.

[0049] Reconstruction loss Used to measure the prediction results of the decoder output. Compared with actual etching results The difference between them. In regression problems where the etching result is a continuous value, this loss can be specifically realized as the negative log-likelihood, which, assuming the output follows a Gaussian distribution, is equivalent to the mean squared error (MSE):

[0050] ;

[0051] in, It is the posterior probability distribution learned by the encoder. This is the latent space vector sampled from this distribution. To enable the gradient to backpropagate through the sampling step, a reparameterization technique is used here, i.e. ,in From the standard normal distribution Random noise in the sampled data.

[0052] KL divergence loss As a regularization term, it is used to constrain the latent space distribution learned by the encoder. Approximating the pre-set standard normal distribution prior This helps to form a structured latent space. The specific calculation formula is as follows:

[0053] ;

[0054] in, It is a latent space The dimension; It is the latent space mean vector The One component; It is the latent space variance vector The Each component has a value determined by... Calculated.

[0055] The specific network structure of the encoder and decoder can be designed by those skilled in the art according to the data dimension and complexity of the specific application. The selection and configuration of the encoder and decoder are well-known technologies in the field and will not be elaborated here.

[0056] After the training process in step S102 converges or meets the preset termination condition, step S103 performs a solidification operation on the model to form a static basic process model.

[0057] Specifically, this solidification operation applies to the decoder portion of the trained variational autoencoder model. The network parameters of the decoder (i.e., weights and biases) are denoted as... Their values ​​are saved after training. These parameters are then used when the system is deployed for online optimization. It is loaded and set to non-training mode. This means that in all subsequent computations, the system will not calculate gradients for these parameters or update them via backpropagation, thus keeping their values ​​constant.

[0058] After the solidification process, the decoder constitutes an independent static basic process model (FPM), which, for a given input, can produce a definite output. The mathematical function of this model can be expressed as:

[0059] ;

[0060] In this function, It is the input process parameter vector; It is the input latent space vector; It is the predicted etching result vector output by the model.

[0061] The technical purpose of this solidification step is to preserve the knowledge learned from the benchmark dataset that characterizes the fundamental physicochemical laws of the etching process in the form of model parameters. The resulting FPM can serve as a stable reference benchmark, providing a foundation for accurately quantifying equipment state drift in subsequent online stages. This design, which separates the fundamental law model from the time-varying correction model, avoids the catastrophic forgetting problem that may occur when fine-tuning the entire large model online.

[0062] After the model is solidified, the encoder part of the original variational autoencoder is no longer used in the subsequent online optimization stage because it is only used to assist in the learning of the latent space during the offline training stage.

[0063] After the static basic process model is constructed and solidified, the system enters the online operation phase.

[0064] In step S104, the system acquires data for each actual production run. Specifically, for each time... For each completed production run, obtain the vector of process parameters used. and the vector of actual etching results obtained through measurement .

[0065] Step S105 quantifies the deviation of the current state of the device from the reference state. This step is achieved by calculating the prediction residual. (The time interval is missing from the original text.) Actual process parameters , and the preset reference latent space vector Together, they are fed into the solidified static basic process model (FPM) as input. Reference latent space vector. For the prior distribution of the latent space The mathematical expectation of the value is typically a zero vector in the implementation. The FPM outputs the baseline prediction result. Predicting residuals This is the actual etching result. Compared with the baseline prediction results The difference between them:

[0066] ;

[0067] In step S106, to obtain a comprehensive description of the equipment drift state, the system generates a process drift state vector based on historical residual sequence encoding. The system internally maintains a length of sliding window This window stores the most recent Predicted residual sequence generated from the next production run:

[0068] ;

[0069] This step involves modifying the window. Statistical analysis is performed on the data to extract a set of features that characterize the current drift properties, and these features are combined into a process drift state vector. .

[0070] In one specific embodiment, the extracted statistical features may include:

[0071] 1. Residual mean vector : Calculate the average of all residual vectors within the window, which reflects the magnitude and direction of the systematic deviation caused by recent equipment drift:

[0072] ;

[0073] 2. Residual covariance matrix Calculate the covariance matrix of the residual vectors within the window, which describes the correlation of drift between different etching result metrics. For example, does a drift in etching rate always accompany a drift in the same or opposite direction of uniformity?

[0074] 3. Residual Trend Vector Linear regression is performed on the residual time series of each dimension (corresponding to an etching result indicator) within the window, and the regression coefficients (slopes) are used to construct a trend vector. This vector reflects the recent speed and direction of drift development.

[0075] Finally, the calculated feature vectors are concatenated to form the process drift state vector. For example, the covariance matrix can be... Vectorization Then, the parts are assembled.

[0076] ;

[0077] This process drifts the state vector. It is a quantitative representation of the current device drift state, which includes the current drift amplitude, internal structure and dynamic change information, and serves as the input for the latent space correction operator in the next stage.

[0078] In step S107, the system generates the process drift state vector in step S106. As input, a latent space perturbation vector for correcting the static basic process model is calculated using a latent space correction operator (LCO). Simultaneously, a corrected confidence score corresponding to this correction is generated. .

[0079] In one embodiment of the present invention, the latent space correction operator is specifically implemented as a Bayesian neural network (BNN), and its network parameters are denoted as follows: The reason for choosing a Bayesian neural network is that its network weights are probability distributions rather than deterministic scalar values. This means that when the network performs forward inference, its output itself is also a probability distribution, thus providing a quantification of the uncertainty of the prediction result and a direct basis for calculating the corrected confidence level. This Bayesian neural network... The specific network structure, such as the number of hidden layers and the number of neurons in each layer, can be determined by those skilled in the art based on the input vector. Dimensions and latent space The design is based on these dimensions, which will not be elaborated upon here.

[0080] The working process of the latent space correction operator is as follows:

[0081] process drift state vector Input this Bayesian neural network The forward propagation process of the network outputs parameters distributed by a Gaussian distribution, i.e., the mean vector. Covariance Matrix :

[0082] ;

[0083] Latent space perturbation vector We take the mean of this output distribution because it represents the most likely latent space correction direction and magnitude in the current drift state:

[0084] ;

[0085] Corrected confidence score The covariance matrix is ​​calculated based on the uncertainty of the output distribution; the higher the uncertainty, the lower the confidence level. traces This reflects the magnitude of the overall prediction variance of the latent space perturbation vector. Therefore, the corrected confidence score can be defined as a decreasing function of this overall variance:

[0086] ;

[0087] in, It is a preset positive hyperparameter used to adjust the sensitivity of the confidence score to variance.

[0088] The latent space correction operator is an online-trainable model. Whenever the system acquires a new valid production data point... Then, the parameters of LCO can be set. An update is performed. The training objective is to find the model that makes the revised static basic process model predict results closest to the actual values. Specifically, the loss function for online training. Defined as the mean square error between the corrected prediction and the actual result:

[0089] ;

[0090] in It is calculated from the current LCO. The system calculates the loss function. Regarding the parameters of the latent space modified operator network The gradient, and use optimization algorithms (such as Adam) to calculate it. It performs small-step updates. This online learning mechanism enables the LCO to continuously track and adapt to slow changes in the device's drift state.

[0091] In step S108, when the system receives a process parameter optimization task, the dynamic correction prediction module 300 constructs a dynamic correction prediction model in real time based on the latest equipment drift information. This model provides the subsequent optimization engine with process behavior predictions reflecting the current equipment state.

[0092] The model is constructed as follows:

[0093] The online drift sensing and correction module 200 obtains the current time... The latest latent space perturbation vector generated is calculated. The disturbance vector is compared with the preset reference latent space vector. Perform vector addition on (usually the zero vector) to obtain the corrected latent space vector. :

[0094] ;

[0095] Subsequently, this corrected latent space vector This serves as the latent space input for the static basic process model (FPM). In this way, a dynamic correction prediction function is formed that is effective during this optimization session. Its mathematical representation is:

[0096] ;

[0097] In this function, It is a vector of process parameters that serves as independent variables, and its range of values ​​is explored by the optimization engine; It is a constant vector calculated and fixed at the beginning of this optimization task; FPM is a fixed static basic process model.

[0098] The function of this dynamic correction prediction model is to, for any set of process parameter vectors to be evaluated given by the optimization engine,... It can output a predicted etching result. This predicted result integrates static basic process knowledge with the dynamic drift information of the current equipment, and is used as an objective function for the optimization engine module 400 to call during the optimization process.

[0099] In step S109, before the optimization task begins, the system determines the internal search strategy of the confidence-driven optimization engine module 400 based on the currently calculated corrected confidence score.

[0100] In one embodiment, the confidence-driven optimization engine module 400 employs a Bayesian optimization algorithm. Bayesian optimization balances the exploration of the parameter space and the utilization of known optimal regions by constructing a surrogate model of the objective function and using a sampling function to select the next evaluation point. For the minimization optimization problem of this invention, the embodiment uses a confidence lower bound (LCB) as the sampling function. Its definition is:

[0101] ;

[0102] in, and These are Bayesian optimization of the internal Gaussian process surrogate model for the objective function at the point... The predicted mean and standard deviation are given. This Gaussian process surrogate model is based on dynamically correcting the prediction model during the optimization process. It is constructed and updated using data points obtained from multiple sampling evaluations. These are the exploration-exploitation balance parameters used to regulate the balance between exploration and exploitation. The optimizer finds parameters that allow for the optimization of the exploration-exploitation balance. smallest To determine the next evaluation point.

[0103] The key to this method lies in exploring and utilizing equilibrium parameters. It is not a fixed value, but rather a revised confidence score generated in step S107 at the beginning of each optimization session. Dynamic settings are implemented. This dynamic setting mechanism uses preset functions. accomplish:

[0104] ;

[0105] This function The specific implementation is a piecewise function that depends on two preset confidence thresholds: a high confidence threshold and a low confidence threshold. and low confidence threshold ,in .

[0106] ;

[0107] In this function:

[0108] , and These are three pre-defined nonnegative constants that satisfy the relation .

[0109] This setting allows the optimization engine to switch its optimization behavior based on the confidence level of the operators adjusted according to the latent space:

[0110] 1. When Greater than the high confidence threshold When this occurs, it indicates that the system believes the prediction model should be dynamically adjusted. It can accurately reflect the current status of the equipment. At this time, Set to a smaller value This will reduce the uncertainty term in the acquisition function. The weights are adjusted so that the optimization search is focused on the vicinity of the currently predicted optimal solution, i.e., the "exploitation" priority strategy is implemented.

[0111] 2. When Less than or equal to the low confidence threshold This indicates that the dynamically corrected prediction model may have a significant bias. Set to a larger value This increases the weight of the uncertainty term, prompting the optimizer to explore regions of high uncertainty in its internal surrogate model. This "exploration-first" strategy helps avoid the optimization process getting trapped in local optima caused by inaccurate prediction models.

[0112] 3. In other cases, when Between and In between, the system uses a balanced parameter. A search strategy that balances the exploration and utilization of standards.

[0113] In step S110, the confidence-driven optimization engine module 400 performs the optimization process for process parameters. This process is based on the dynamically corrected prediction model constructed in step S108. And the exploration using equilibrium parameters determined in step S109 Expand.

[0114] In one embodiment, the optimization process specifically includes the following sub-steps:

[0115] S110-1, Initialize the optimization engine. Taking the Bayesian optimization engine as an example, its internal Gaussian process surrogate model, used to fit the dynamically corrected prediction model, is initialized. Simultaneously, process parameters are defined. The search space, i.e., the upper and lower limits of the constraints for each process parameter.

[0116] S110-2, Iterative search execution. The optimization engine iteratively performs the following operations in a loop:

[0117] 1) Select the next evaluation point: based on the predicted mean given by the current Gaussian process surrogate model. and the predicted standard deviation Based on this, and combined with the established exploration-utilization balance parameters The next vector of process parameters to be evaluated is selected by solving the following minimization problem. :

[0118] ;

[0119] This formula is a confidence lower bound (LCB) acquisition function defined for minimizing the problem, consistent with the strategy defined in step S109.

[0120] 2) Evaluate the objective function: Call the dynamic correction prediction model to evaluate the selected process parameter vector. An evaluation was conducted to obtain the predicted etching results. :

[0121] ;

[0122] 3) Update the proxy model: Apply the newly acquired data to... The observed dataset is added to the optimization engine, and its internal Gaussian process surrogate model is retrained using the updated dataset.

[0123] S110-3, Determine the termination condition. After each iteration, check whether the preset termination condition is met. The termination condition may include: reaching the preset maximum number of iterations; or, the improvement of the optimal target value is lower than the preset threshold (i.e., convergence) for a consecutive number of steps. If either condition is met, stop the iteration; otherwise, return to S110-2.

[0124] S110-4, Output the optimal result. After the iteration process terminates, the optimization engine finds the process parameter vector that optimizes the objective function from its historical observation dataset and uses it as the final result of this optimization. Output. For example, in a minimization problem, It is the corresponding observation among all evaluated X values. The one with the smallest value.

[0125] Example:

[0126] The technical solution of the present invention will be described below with reference to a specific application embodiment. This embodiment aims to illustrate the parameter optimization process of the plasma etching process for polycrystalline silicon gates in semiconductor manufacturing.

[0127] The goal of this process is to precisely control the critical dimension (CD) and sidewall angle (SWA) of the gate, while maximizing the etch rate (ER) to improve production efficiency. During continuous production, the walls of the etching chamber gradually age due to the deposition of reaction byproducts, causing process results to drift over time. The objective of this embodiment is to quickly and accurately calculate compensatory optimal process parameters for new production targets when equipment drift occurs.

[0128] 1. Establishment of the static basic process model (steps S101~S103):

[0129] First, on an etching machine that has just undergone standard cleaning and maintenance (i.e., in baseline condition), step S101 is performed. Through experimental design, the process parameter vector is systematically varied. Its components include CHF3 gas flow rate, O2 gas flow rate, and RF bias power. For each set of process parameters... Process the wafer and measure its corresponding etching result vector. Its components include etch rate (ER), gate critical size (CD), and sidewall angle (SWA). This results in a set containing 200 groups. The data is an offline training dataset.

[0130] Subsequently, step S102 is performed to train a variational autoencoder model using the dataset. The decoder part of this model is designed to receive process parameters. and latent space vector Output the predicted etching results .

[0131] After training is complete, step S103 is executed to solidify the parameters of the decoder network, forming a static basic process model (FPM). This FPM represents the ideal process response characteristics of the cleanroom.

[0132] 2. Online drift sensing and correction (steps S104-S107):

[0133] After the system went online and processed 50 wafers, the equipment status drifted. At this point, steps S104 to S107 were executed.

[0134] During the processing of the 51st wafer (time) Record the process parameters used. And the actual etching results measured after processing .assumed The CD value is 46.2nm.

[0135] Execute step S105, and and zero vector Input FPM to obtain baseline prediction results. Assuming its CD value is 45.1 nm, the prediction residual for the current CD dimension is calculated. =46.2-45.1=+1.1nm.

[0136] Execute step S106, the system retrieves the most recent... The residual sequence of wafers (32nd to 51st) was used to calculate the process drift state vector. This vector contains the mean of the residuals (e.g., CD is 0.9 nm larger on average), covariance (e.g., there is a strong correlation between a larger CD and a smaller SWA), and trend (e.g., the CD deviation increases at a rate of 0.05 nm per wafer), thus comprehensively characterizing the current drift state.

[0137] Execute step S107, and The input is the latent space correction operator (LCO) (Bayesian neural network). Because recent drift data patterns are clear, the LCO can make inferences with high certainty, thus outputting a latent space perturbation vector. It also provides a high corrected confidence score. =0.92. Simultaneously, the system uses data from the 51st wafer to perform an online update of the LCO's network parameters, enabling it to better track subsequent drift.

[0138] 3. Confidence-driven optimization process (steps S108-S110):

[0139] At this point, the system receives a new optimization task: to find the optimal process parameters for the next batch of products, so that it can achieve the target of CD of 48.0nm and SWA of 90 degrees, while maximizing ER.

[0140] Execute step S108, the system will update the latest and Adding them together yields the corrected latent space vector. The dynamic correction prediction model used in this optimization session was then constructed. This model represents the behavior of currently drifting devices.

[0141] Execute step S109 to optimize the engine reading and correcting the confidence score. =0.92. This value is higher than the preset high confidence threshold. (e.g., 0.90), the optimization engine will explore and utilize the balance parameters. Set to a smaller value This instructs the optimization engine to trust the predictions of the current dynamically revised model and adopt a "utilize"-first search strategy.

[0142] Finally, step S110 is executed, and the Bayesian optimization engine starts. It minimizes the comprehensive objective function. The goal is to iteratively optimize within its scope. Among these optimizations, For positive weighting coefficients, minimize This term is equivalent to maximizing ER. In each iteration, it calls... Conduct a rapid assessment. Because... When the value is small, the search will focus around the optimal solution predicted by the dynamic model. After dozens of iterations, the optimization process converges and outputs a set of optimal process parameters. This set of parameters The current device drift has been implicitly compensated, and when applied to actual production, it is expected to achieve the new 48nm CD target.

[0143] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent optimization of precision etching process parameters based on deep learning, characterized in that, Includes the following steps: A fixed static basic process model is provided, which is configured to output the corresponding predicted etching results based on the input process parameter vector and latent space vector; During online operation, the actual production process parameters and actual etching results are obtained. Based on the actual production process parameters and the preset benchmark latent space vector, the static basic process model is called to obtain the benchmark prediction result, and then the prediction residual between the actual etching result and the benchmark prediction result is calculated. Based on the historical predicted residual sequence, the process drift state vector is encoded and generated. The process drift state vector is input into the latent space correction operator to generate a latent space perturbation vector for correcting the static basic process model; The latent space perturbation vector is applied to the static basic process model to construct a dynamic correction prediction model; An optimization search is performed based on the dynamic correction prediction model to obtain the optimal process parameters.

2. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 1, characterized in that, The step of providing a solidified static basic process model includes: Collect offline training datasets under baseline device conditions; A variational autoencoder model containing an encoder and a decoder is trained using the offline training dataset. The network weights of the decoder part in the variational autoencoder model are solidified, and the decoder is used as the static basic process model.

3. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 1, characterized in that, The step of encoding and generating process drift state vectors based on historical predicted residual sequences includes: Within a preset sliding window, one or more sets of statistical features are extracted from the historical prediction residual sequence to form the process drift state vector.

4. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 3, characterized in that, The statistical characteristics include: The historical predicted residual sequence includes at least one of the mean vector, covariance matrix, and trend vector.

5. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 1, characterized in that, The step of inputting the process drift state vector into the latent space correction operator further includes: The latent space correction operator generates the latent space perturbation vector and simultaneously generates a correction confidence score that quantifies the uncertainty of the correction.

6. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 5, characterized in that, Before the step of performing the optimization search based on the dynamically corrected prediction model, the method further includes: Based on the corrected confidence score, the exploration-utilization balance parameters of the optimization engine are determined; The optimization engine performs the optimization search based on the exploration-utilization of balance parameters.

7. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 6, characterized in that, The steps for determining the optimization engine using balancing parameters include: When the corrected confidence score is greater than the preset high confidence threshold, the exploration-utilization balance parameter is set to the preset utilization value; When the corrected confidence score is less than or equal to a preset low confidence threshold, the explore-utilization balance parameter is set to a preset explore value.

8. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 5, characterized in that, The latent space correction operator is a Bayesian neural network model, whose output is a probability distribution. The latent space perturbation vector is the mean of this probability distribution, and the corrected confidence score is calculated based on the variance of this probability distribution.

9. The intelligent optimization method for precision etching process parameters based on deep learning according to claim 8, characterized in that, The method further includes: The Bayesian neural network model is updated online using the actual production process parameters and actual etching results obtained during online operation. The update objective is to minimize the error between the prediction results of the static basic process model after correction by the latent space perturbation vector and the actual etching results.

10. A deep learning-based intelligent optimization system for precision etching process parameters, used to execute the deep learning-based intelligent optimization method for precision etching process parameters as described in any one of claims 1-9, characterized in that, include: The static basic process model module contains a static basic process model and is configured to output predicted etching results based on the input process parameter vector and latent space vector. The online drift sensing and correction module is configured to acquire the actual production process parameters and actual etching results, calculate the prediction residuals with the prediction results of the static basic process model, generate a process drift state vector based on the historical prediction residual sequence, and generate a latent space perturbation vector and a corrected confidence score based on the state vector. The dynamic correction prediction module is configured to apply the latent space disturbance vector to the static basic process model to construct a dynamic correction prediction model. The confidence-driven optimization engine module is configured to set its exploration-utilization balance strategy based on the modified confidence score and invoke the dynamically modified prediction model to perform optimization search to output optimal process parameters.

Images