A TMR sensing chip thin film coating parameter optimization method and device
By combining deep learning models with intelligent optimization algorithms, the inefficiency and accuracy problems of traditional TMR sensor chip thin film coating process parameter optimization are solved, realizing efficient and accurate parameter combination search, and improving process development efficiency and cost-effectiveness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID ZHEJIANG ELECTRIC POWER CO MARKETING SERVICE CENT
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241425A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of chips, and in particular relates to a method and apparatus for optimizing thin film coating parameters of TMR sensor chips. Background Technology
[0002] Tunneling magnetoresistance (TMR) is a core physical effect in magnetoresistive random access memory (MRAM) and magnetic sensors. The performance of TMR sensor chips, such as magnetoresistive ratio, sensitivity, and thermal stability, is highly dependent on the precise control of thin-film deposition process parameters (such as sputtering power, deposition time, substrate temperature, gas pressure, and target composition). Traditional process parameter optimization mainly relies on engineers' experience and a "trial and error" approach, conducting a limited number of experiments through orthogonal experimental design. This is not only time-consuming and resource-intensive, but also makes it difficult to capture the complex nonlinear relationships between parameters, thus failing to achieve global optimization.
[0003] In related technologies, there have been attempts to apply machine learning to process optimization, but most of them use shallow models (such as support vector machines and random forests). For complex processes like TMR thin film deposition, which involve multiple parameters, strong coupling, nonlinearity, and time-series characteristics, the fitting ability and generalization performance of shallow models are limited. In addition, existing methods focus more on predictive model construction and lack a systematic solution to effectively combine models with optimization algorithms to directly output the optimal parameter combination. Summary of the Invention
[0004] In view of this, the present invention discloses a method and apparatus for optimizing thin film coating parameters of TMR sensor chips, which can solve the shortcomings of related technologies.
[0005] To achieve the above objectives, the present invention discloses the following technical solution: According to a first aspect of the present invention, a method for optimizing thin-film coating parameters of a TMR sensor chip is proposed, comprising: Acquire historical parameter data of the thin film coating of the TMR sensor chip and the corresponding chip performance index data, and generate a training dataset; A deep learning model is constructed and trained based on the training dataset to obtain a mapping model. The mapping model is used to establish a nonlinear mapping relationship between coating process parameters and chip performance indicators. The mapping model includes: a fully connected neural network module for extracting static nonlinear features of the coating process parameters; a convolutional neural network module for extracting spatial correlation features between coating process parameters arranged according to thin film structure; a long short-term memory network module for extracting dynamic features of coating process parameters with time series characteristics; and a splicing module for splicing the outputs of multiple neural network modules included in the mapping model. A parameter optimization problem is constructed based on the mapping model and solved within the feasible region of the coating process parameters to obtain the target parameter combination that optimizes the chip performance.
[0006] According to a second aspect of the present invention, a device for optimizing thin-film coating parameters of a TMR sensor chip is provided, the device comprising: Acquisition Unit: Acquires historical parameter data of the thin-film coating of the TMR sensor chip and the corresponding chip performance index data, and generates a training dataset; Training Unit: Constructs and trains a deep learning model based on the training dataset to obtain a mapping model; wherein, the mapping model is used to establish a nonlinear mapping relationship between coating process parameters and chip performance indicators, and the mapping model includes: a fully connected neural network module for extracting the static nonlinear features of the coating process parameters, a convolutional neural network module for extracting the spatial correlation features between coating process parameters arranged according to the thin film structure, a long short-term memory network module for extracting the dynamic features of coating process parameters with time series characteristics, and a splicing module for splicing the outputs of multiple neural network modules included in the mapping model; Solving unit: Constructs a parameter optimization problem based on the mapping model and solves it within the feasible region of the coating process parameters to obtain the target parameter combination that optimizes the chip performance.
[0007] According to a third aspect of the present invention, an electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor implements the steps of the method as described in the first aspect by running the executable instructions.
[0008] According to a fourth aspect of the invention, a computer-readable storage medium is provided having computer instructions stored thereon that, when executed by a processor, implement the steps of the method as described in the first aspect.
[0009] As can be seen from the above technical solutions, the TMR sensor chip thin film coating parameter optimization method disclosed in this invention is as follows: On the one hand, deep learning models incorporating fully connected neural networks, convolutional neural networks, and long short-term memory networks are used to uncover the complex nonlinear relationships, spatial correlations, and time-series characteristics between process parameters and performance indicators. Compared to traditional shallow models, these models exhibit stronger fitting capabilities and prediction accuracy, thereby enhancing the accuracy of parameter optimization. On the other hand, combining deep learning models with intelligent optimization algorithms enables automatic and rapid searching of the process parameter space to find globally optimal or near-optimal solutions. This avoids the blindness and inefficiency of traditional trial-and-error methods, shortens the process development cycle, and not only improves parameter optimization efficiency but also reduces experimental costs. Attached Figure Description
[0010] Figure 1 This is a flowchart of an exemplary embodiment of a method for optimizing thin-film coating parameters of a TMR sensor chip; Figure 2 This is a schematic diagram of a mapping model structure provided in an exemplary embodiment; Figure 3 This is a schematic diagram illustrating a parameter optimization solution provided in an exemplary embodiment; Figure 4 This is a schematic diagram illustrating a comparison of experimental results provided in an exemplary embodiment; Figure 5 This is a schematic structural diagram of a device provided in an exemplary embodiment; Figure 6 This is a block diagram of an exemplary embodiment of a device for optimizing thin-film coating parameters of a TMR sensor chip. Detailed Implementation
[0011] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of the present invention as detailed in the appended claims.
[0012] It should be noted that in other embodiments, the corresponding methods are not necessarily performed in the order shown and described in this invention. The method comprises steps. In some other embodiments, the method may include more or fewer steps than those described in this invention. Furthermore, a single step described in this invention may be broken down into multiple steps in other embodiments; and multiple steps described in this invention may be combined into a single step in other embodiments.
[0013] Tunneling magnetoresistance (TMR) is a core physical effect in magnetoresistive random access memory (MRAM) and magnetic sensors. The performance of TMR sensor chips, such as magnetoresistive ratio, sensitivity, and thermal stability, is highly dependent on the precise control of thin-film deposition process parameters (such as sputtering power, deposition time, substrate temperature, gas pressure, and target composition). Traditional process parameter optimization mainly relies on engineers' experience and a "trial and error" approach, conducting a limited number of experiments through orthogonal experimental design. This is not only time-consuming and resource-intensive, but also makes it difficult to capture the complex nonlinear relationships between parameters, thus failing to achieve global optimization.
[0014] In related technologies, there have been attempts to apply machine learning to process optimization, but most of them use shallow models (such as support vector machines and random forests). For complex processes like TMR thin film deposition, which involve multiple parameters, strong coupling, nonlinearity, and time-series characteristics, the fitting ability and generalization performance of shallow models are limited. In addition, existing methods focus more on predictive model construction and lack a systematic solution to effectively combine models with optimization algorithms to directly output the optimal parameter combination.
[0015] To address the shortcomings in related technologies, this invention proposes a method and apparatus for optimizing thin-film coating parameters of TMR sensor chips.
[0016] Figure 1 This is a flowchart illustrating an exemplary embodiment of a method for optimizing thin-film coating parameters of a TMR sensor chip. (Example:) Figure 1 As shown, the method may include the following steps: Step 101: Obtain historical parameter data of the thin film coating of the TMR sensor chip and the corresponding chip performance index data, and generate a training dataset.
[0017] Historical parameter data is a combination of historical coating process parameters, which may include at least: base vacuum, working gas pressure, sputtering power of each layer, sputtering time of each layer, gas ratio, and substrate temperature.
[0018] The chip performance index data refers to the core performance indicators of the TMR chip prepared under the corresponding parameter combination, which may include at least: magnetoresistive ratio, sensitivity, resistance-area area, and coercivity.
[0019] Step 102: Construct and train a deep learning model based on the training dataset to obtain a mapping model; wherein, the mapping model is used to establish a nonlinear mapping relationship between coating process parameters and chip performance indicators, and the mapping model includes: a fully connected neural network module for extracting the static nonlinear features of the coating process parameters, a convolutional neural network module for extracting the spatial correlation features between coating process parameters arranged according to the thin film structure, a long short-term memory network module for extracting the dynamic features of coating process parameters with time series characteristics, and a splicing module for splicing the outputs of multiple neural network modules included in the mapping model.
[0020] The output of the mapping model is one or more chip performance indicators predicted for the combination of input coating process parameters.
[0021] Specifically, the fully connected neural network module includes multiple fully connected layers that use linear rectified functions as activation functions, and dropout layers set between the fully connected layers; the convolutional neural network module includes one-dimensional convolutional layers, pooling layers, and flattening layers; and the long short-term memory network module includes at least one LSTM layer.
[0022] Step 103: Construct a parameter optimization problem based on the mapping model and solve it within the feasible region of the coating process parameters to obtain the target parameter combination that optimizes the chip performance.
[0023] In this embodiment, on the one hand, a deep learning model incorporating fully connected neural networks, convolutional neural networks, and long short-term memory networks is employed to mine the complex nonlinear relationships, spatial correlations, and time-series characteristics between process parameters and performance indicators. Compared to traditional shallow models, this model exhibits stronger fitting capabilities and prediction accuracy, thereby enhancing the accuracy of parameter optimization. On the other hand, combining the deep learning model with intelligent optimization algorithms enables automatic and rapid searching of the process parameter space to find the globally optimal or near-optimal solution. This avoids the blindness and inefficiency of traditional trial-and-error methods, shortens the process development cycle, and not only improves parameter optimization efficiency but also reduces experimental costs.
[0024] In one embodiment, the method further includes: performing data cleaning operations on the acquired historical parameter data and chip performance index data; wherein the cleaning operations include at least one of the following: outlier detection and processing, missing value imputation, unit unification; normalizing or standardizing the cleaned data; and dividing the processed data into training datasets, validation datasets, and test datasets according to a predetermined ratio.
[0025] The collected data is cleaned, normalized / standardized, and divided into training, validation, and test sets. To improve development efficiency and reproducibility, this invention employs an automated data preprocessing pipeline to perform data cleaning, feature transformation, and normalization / standardization. Specifically, it utilizes the Pipeline mechanism from Python's scikit-learn library, encapsulating missing value imputation, outlier detection, and normalization operations into a single process. This ensures that the same processing flow is automatically triggered each time data is updated or new data is added, avoiding manual repetition of steps.
[0026] Data cleaning can include outlier handling, missing value handling, and unification of units and dimensions.
[0027] Outlier handling: Screen outlier data, for example, by using box plots or the 3σ principle to identify outliers; for outlier records caused by equipment failure or human error, delete or replace them.
[0028] Missing value handling: Imput missing data, for example, by using the mean, median, interpolation method, or more intelligent methods such as K-nearest neighbors (KNN) to fill in missing values; if some samples have too many missing values and cannot be salvaged, then the sample points are removed.
[0029] Unit and dimension consistency: Standardize the unit conversion for different physical quantities (such as power, air pressure, temperature, etc.) to avoid affecting model training due to inconsistent dimensions.
[0030] Data normalization processing can include: Normalization: Scaling the eigenvalues to the range of [0, 1] or [-1, 1] by the maximum and minimum values. It is suitable for parameters with strict limits on their numerical range.
[0031] Standardization: Calculates the mean and standard deviation of a feature and scales it to a mean of 0 and a variance of 1. It is suitable for Gaussian distribution or feature data with infinite boundaries.
[0032] All normalization / standardization parameters are computed on the training set and applied to both the validation and test datasets to ensure that no data leakage is introduced during the testing phase.
[0033] The dataset can be divided into training, validation, and test datasets in a ratio of 7:2:1 or 6:2:2. Training Set: Used for learning model parameters. Validation Set: Used to monitor for overfitting during training and to adjust model structure and hyperparameters. Test Set: Used to independently evaluate the model's generalization ability, providing an objective assessment of the final prediction accuracy.
[0034] Formatting: The preprocessed dataset is stored in a formatted structure (CSV file). In one embodiment, constructing and training a deep learning model based on the training dataset includes: constructing a deep neural network, the deep neural network comprising an input layer, at least one hidden layer, and an output layer; wherein the number of nodes in the input layer corresponds to the number of coating process parameters, and the number of nodes in the output layer corresponds to the number of chip performance indicators; determining a loss function and selecting an optimizer; training the deep neural network based on the training dataset, and using the validation dataset to monitor the training process to adjust hyperparameters until the model converges.
[0035] like Figure 2 As shown, a deep neural network model is constructed. The number of nodes in the input layer of this model corresponds to the number of coating parameters, and the number of nodes in the output layer corresponds to the number of performance metrics (supporting multi-objective optimization).
[0036] The deep neural network model may include fully connected layers and dropout layers to prevent overfitting. For more complex relationships, convolutional neural networks can be used to handle spatial correlations between parameters, or long short-term memory networks can be introduced to handle process parameters with time-series characteristics.
[0037] Define a loss function (such as mean squared error MSE or mean absolute error MAE) and select an appropriate optimizer (such as Adam).
[0038] The model is trained using the training set, and the training process is monitored and hyperparameters are tuned using the validation set until the model converges. Finally, the prediction accuracy of the model is evaluated using the test set.
[0039] The fully connected neural network module (MLP, Multi-Layer Perceptron) is used to process numerical scalar features (such as sputtering power, substrate temperature, etc.). This module contains the following layers: Input Layer: The input dimension is the number of coating process parameters.
[0040] Hidden Layers: These consist of multiple fully connected layers that use the ReLU activation function to achieve non-linear feature extraction. For example, three layers can be set up with 128, 64, and 32 nodes respectively.
[0041] Dropout layer: Insert a Dropout layer (e.g., 0.2~0.5) between hidden layers to reduce the risk of overfitting.
[0042] Output Layer: Outputs the predicted value of the corresponding performance metric, with the activation function being either linear or sigmoid (set according to the range of the predicted metric).
[0043] Each hidden layer extracts high-order nonlinear features, achieving a complex mapping from input parameters to performance metrics through layer-by-layer transformation. This network structure is suitable for approximate modeling of static parameters and performance.
[0044] Convolutional Neural Network (CNN) modules are used to handle the spatial structural relationships of parameters. For example, in multilayer thin films, the process parameters are arranged layer by layer according to physical hierarchy, and there is a lateral dependency.
[0045] CNNs can extract the patterns of how local parameter combinations affect performance: 1D Convolutional Layer (Conv1D): The input is a one-dimensional parameter sequence arranged according to the thin film structure, and the convolutional kernel window (such as 3 or 5) extracts the local parameter relationship.
[0046] Pooling layer: Max pooling is used to reduce feature size and enhance the model's robustness to noise.
[0047] Flatten layer: Flattens the extracted spatial features and fuses them with the output of the MLP module.
[0048] The output of CNN layers supplements spatial structure-related features, enabling the model to not only focus on independent parameter values, but also capture the interactive effects of parameter combinations between layers.
[0049] Long-Short Term Memory (LSTM) network modules are suitable for processing time-series process parameters, such as dynamically changing annealing temperatures and gas flow rate ratios.
[0050] LSTM can automatically remember the sequential impact of parameter changes on the final performance: LSTM layer (single or stacked double layers): The input is a sequence of process parameters on the time axis, and the output is a time-dependent summary (hidden state vector).
[0051] Dense layer integration: The temporal dynamic features extracted by LSTM are input into the subsequent fully connected layer to achieve global fusion.
[0052] The advantage of LSTM is that it preserves "historical dependence" and reflects the cumulative effect of time-dimensional intervention on the final performance of thin film structures in the actual process.
[0053] Further, determining the loss function includes: when the mapping model is used to predict a single performance index, determining the mean squared error loss function as the loss function of the mapping model; when the mapping model is used to predict multiple performance indexes, determining the weighted multi-objective loss function as the loss function of the mapping model, wherein the weighted multi-objective loss function is a weighted sum of the mean squared error losses of the multiple performance indexes; and when there are process constraints, introducing a penalty term into the loss function of the mapping model, wherein the penalty term is used to constrain process parameters to remain within a preset physically reasonable range.
[0054] When the model outputs only a single performance metric (such as magnetoresistance ratio), using mean squared error (MSE) as the loss function is more suitable for continuous numerical prediction and can amplify samples where the predicted value deviates significantly from the actual value, thereby improving the overall model's fitting accuracy.
[0055] Its mathematical definition is: ; in, For the number of performance parameters, Let the weight be the weight of the k-th task. This represents the mean squared error for the corresponding task.
[0056] When multiple performance parameters (such as reluctance ratio, RA value, resistance, coercivity, etc.) need to be predicted simultaneously, this invention employs a weighted multi-objective loss function. Each prediction output is assigned a weight coefficient to balance the importance or numerical scale of different tasks: ; To further constrain specific process conditions (such as resistance range, current directionality, etc.), a penalty term can be introduced based on actual process requirements. For example, if a target needs to be maintained within a physically reasonable range, a penalty term can be introduced: ; in, For constraint strength coefficients, For parameters A function to determine whether a rule is violated.
[0057] The final total loss function is: ; This invention utilizes the Adam optimizer to optimize the loss function, sets a learning rate decay strategy, and incorporates an early stopping mechanism to prevent overfitting. Furthermore, during multi-objective loss training, the gradient normalization (GradNorm) method can be combined to further improve the co-training performance of imbalanced tasks.
[0058] The outputs of the three modules are concatenated to form a unified feature representation, which is then passed through one or more fully connected layers to obtain the performance prediction result. The hybrid model is trained using the training set, and the training process is monitored and hyperparameters are tuned using the validation set until the model converges. Finally, the prediction accuracy of the model is evaluated using the test set. In one embodiment, the step of constructing a parameter optimization problem based on the mapping model and solving it within the feasible region of the coating process parameters includes: determining a parameter vector composed of coating process parameters and using chip performance indicators as the optimization objective function; determining upper and lower bound constraints for each coating process parameter to form a feasible region; and selecting gradient ascent, global optimization algorithm, or hybrid optimization strategy to solve the optimization problem based on the differentiability of the deep learning model and the complexity of the objective function to obtain the target parameter combination.
[0059] Furthermore, the step of selecting gradient ascent, global optimization algorithm, or hybrid optimization strategy to solve the optimization problem based on the differentiability of the deep learning model and the complexity of the objective function includes: when the deep learning model is differentiable and the objective function is a convex function, using gradient ascent to iteratively update the process parameters, and projecting the parameters into the feasible region after each iteration through a pruning operation until convergence; when the objective function is not differentiable or there are multiple local optima, using a genetic algorithm or particle swarm optimization algorithm for global search, and iteratively solving the problem through fitness evaluation, selection, crossover, mutation, or velocity and position update operations; when the system contains discrete parameters or has complex constraints, first using a genetic algorithm or particle swarm optimization algorithm for global search to obtain candidate solutions, and then using gradient ascent to perform local fine-tuning in the neighborhood of the candidate solutions.
[0060] like Figure 3 As shown, an optimization problem is constructed and solved based on a trained deep learning model.
[0061] Objective: To find a set of coating parameters This improves the model's predictive performance. To achieve the optimal (e.g., maximizing the magnetoresistive ratio while keeping the resistance-area area within a certain range).
[0062] Mathematical expression: , ; Where F is the trained deep learning model, P is the coating parameter vector, and the constraint is the physical feasible region of the parameters.
[0063] Since optimization problems involve multiple objectives and both discrete and continuous input parameters, this invention, based on the differentiability of deep neural network models and the complexity of objective functions, adopts a hybrid solution framework that combines gradient ascent and global optimization algorithms to achieve an adaptive optimization process from global search to local fine-tuning.
[0064] The process mainly includes five stages: problem modeling, initialization, optimization solution, constraint handling, and result selection.
[0065] Problem modeling: The process parameters are represented as a vector: ; These include the base vacuum, working gas pressure, sputtering power of each layer, sputtering time, gas ratio, and substrate temperature.
[0066] The performance metrics (such as magnetoresistance ratio MR, RA value, and film resistance) are recorded as model outputs: ; The prediction is given by a pre-trained deep neural network model.
[0067] Define an optimization objective function, which can be a single objective or a weighted multi-objective form: ; in, For the first One performance metric, For weights.
[0068] Algorithm selection strategy: When the model is differentiable, gradient ascent is used for fast convergence; when the model is not differentiable or has non-convexity or multiple local optima, a global heuristic algorithm (genetic algorithm, particle swarm optimization algorithm) is used; if the system is complex or contains discrete parameters, a hybrid strategy can be used: a global algorithm searches for candidate solutions, and gradient method is used for local fine-tuning.
[0069] Gradient method solution: Initialize parameter vector In the model Calculate the gradient of the performance index: ; Update parameters (gradient ascent):
[0070] in, The learning rate; Perform constrained projection: ; Iterate until convergence or the performance improvement threshold is reached, then output the optimal solution. .
[0071] Global optimization solution process: When the objective function is non-differentiable or has multiple local optima, a genetic algorithm (GA) or particle swarm optimization (PSO) is used for global search. Genetic Algorithm (GA) Steps: Encoding Representation, Parameters Encode as chromosomes; initialize the population, randomly generate several candidate solutions; evaluate fitness, calculate the fitness of each individual in the model. Performance Selection operation: Select superior individuals based on fitness probability; Crossover and mutation: Generate new solutions while maintaining population diversity; Constraint repair and screening: Eliminate or repair individuals that do not meet process constraints; Termination condition: Output the optimal solution when the performance improvement is less than a threshold or the maximum number of iterations is reached.
[0072] Particle Swarm Optimization (PSO) Steps: Initialize particle swarm positions and speed Calculate the fitness of each particle; update the individual extreme values. With global extrema Repeat until convergence, then output the optimal solution.
[0073] Update the particle state according to the following formula: ; .
[0074] Hybrid optimization strategy: First, a global search is performed using GA / PSO to obtain multiple high-potential candidate solutions. Then, in the neighborhood of these solutions, local fine-tuning is performed using gradient ascent, ultimately outputting the optimal solution set and forming the Pareto compromise front.
[0075] The following detailed explanation uses the optimized magnetoresistance ratio of the CoFeB / MgO / CoFeB structure TMR as an example: 1. Data preparation: Input parameters (8): antiferromagnetic layer sputtering power, pinned layer sputtering power, MgO barrier layer sputtering power, free layer sputtering power, annealing temperature, annealing time, annealing magnetic field strength, and background vacuum.
[0076] Output metric (1): Reluctance ratio.
[0077] We collected 200 sets of historical experimental data to form the initial dataset. The data was Z-score standardized and divided into training, validation, and test sets in a 7:2:1 ratio.
[0078] 2. Model building and training: Construct a fully connected neural network with 4 hidden layers, with the number of nodes in each layer being 128, 64, 32, and 16 respectively. Use ReLU as the activation function and a linear activation function for the output layer.
[0079] The loss function is MSE, the optimizer is Adam, and the learning rate is set to 0.001.
[0080] The model was trained for 300 epochs on the training set with a batch size of 32. During training, the loss function continuously decreased on the validation set and eventually converged. The coefficient of determination (R²) between the predicted and measured values on the test set exceeded 0.92, indicating that the model has extremely high prediction accuracy.
[0081] 3. Parameter optimization: Optimization objective: Maximize the predicted magnetoresistance ratio.
[0082] Constraints: All input parameters are within the physical limits allowed by their device.
[0083] Solution method: Gradient ascent method is used. A set of parameters is randomly initialized, input into the model to calculate the predicted value of the magnetoresistance ratio, and then backpropagation is used to calculate the gradient of the predicted value with respect to the input parameters. The parameters are updated according to the gradient direction, iterating 1000 times or until the parameter change is less than a threshold.
[0084] Output: A set of recommended optimal parameters .
[0085] 4. Experimental verification: according to TMR chips are fabricated on a magnetron sputtering device.
[0086] The measured magnetoresistance ratio is 15% higher than the historical best level.
[0087] Add this new data point to the dataset, fine-tune the model, and update the model weights.
[0088] The above embodiments confirm the combination Figure 4 This embodiment can effectively and accurately optimize the thin film coating parameters of the TMR sensor chip and improve product performance.
[0089] Figure 5 This is a schematic structural diagram of a device provided in an exemplary embodiment. Please refer to... Figure 5At the hardware level, the device includes a processor 502, an internal bus 504, a network interface 506, memory 508, and non-volatile memory 510, and may also include other hardware required for its functions. One or more embodiments of the present invention can be implemented in software, for example, the processor 502 reads the corresponding computer program from the non-volatile memory 510 into memory 508 and then runs it. Of course, in addition to software implementation, one or more embodiments of the present invention do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.
[0090] Please refer to Figure 6 A device for optimizing thin-film coating parameters of a TMR sensor chip can be applied to applications such as... Figure 6 The device shown, in order to implement the technical solution of the present invention, includes: The acquisition unit 601 is used to acquire historical parameter data of the thin film coating of the TMR sensor chip and the corresponding chip performance index data, and generate a training dataset. Training unit 602 is used to construct and train a deep learning model based on the training dataset to obtain a mapping model; wherein, the mapping model is used to establish a nonlinear mapping relationship between coating process parameters and chip performance indicators, and the mapping model includes: a fully connected neural network module for extracting static nonlinear features of the coating process parameters, a convolutional neural network module for extracting spatial correlation features between coating process parameters arranged according to thin film structure, a long short-term memory network module for extracting dynamic features of coating process parameters with time series characteristics, and a splicing module for splicing the outputs of multiple neural network modules included in the mapping model; The solution unit 603 is used to construct a parameter optimization problem based on the mapping model and solve it within the feasible domain of the coating process parameters to obtain the target parameter combination that makes the chip performance index optimal.
[0091] The fully connected neural network module includes multiple fully connected layers that use linear rectified functions as activation functions, and dropout layers set between the fully connected layers; the convolutional neural network module includes one-dimensional convolutional layers, pooling layers, and flattening layers; the long short-term memory network module includes at least one LSTM layer.
[0092] Optionally, the device further includes: The cleaning unit 604 is used to perform data cleaning operations on the acquired historical parameter data and chip performance index data; wherein the cleaning operation includes at least one of the following: outlier detection and processing, missing value imputation, and unit unification; Processing unit 605 is used to normalize or standardize the cleaned data. The partitioning unit 606 is used to divide the processed data into training dataset, validation dataset and test dataset according to a predetermined ratio.
[0093] Furthermore, the training unit 602 is specifically used for: A deep neural network is constructed, comprising an input layer, at least one hidden layer, and an output layer; wherein the number of nodes in the input layer corresponds to the number of coating process parameters, and the number of nodes in the output layer corresponds to the number of chip performance indicators. Determine the loss function and select the optimizer; The deep neural network is trained using the training dataset, and the training process is monitored using the validation dataset to adjust hyperparameters until the model converges.
[0094] Furthermore, the training unit 602 is specifically used for: When the mapping model is used to predict a single performance metric, the mean squared error loss function is determined as the loss function of the mapping model. When the mapping model is used to predict multiple performance indicators, the weighted multi-objective loss function is determined as the loss function of the mapping model, and the weighted multi-objective loss function is the weighted sum of the mean squared error losses of the multiple performance indicators; When process constraints exist, a penalty term is introduced into the loss function of the mapping model. The penalty term is used to constrain process parameters to remain within a preset physically reasonable range.
[0095] Optionally, the solving unit 603 is specifically used for: Determine the parameter vector composed of coating process parameters, and use chip performance indicators as the optimization objective function; Determine the upper and lower bound constraints of each coating process parameter to form a feasible region; Based on the differentiability of the deep learning model and the complexity of the objective function, the gradient ascent method, global optimization algorithm, or hybrid optimization strategy is selected to solve the optimization problem in order to obtain the combination of objective parameters.
[0096] Furthermore, the solving unit 603 is specifically used for: When the deep learning model is differentiable and the objective function is a convex function, the gradient ascent method is used to iteratively update the process parameters. After each iteration, the parameters are projected into the feasible region through a pruning operation until convergence. When the objective function is non-differentiable or has multiple local optima, a global search is performed using a genetic algorithm or particle swarm optimization algorithm, and the solution is solved iteratively through fitness evaluation, selection, crossover, mutation or velocity and position update operations. When the system contains discrete parameters or has complex constraints, a global search is first performed using a genetic algorithm or particle swarm optimization algorithm to obtain candidate solutions, and then the gradient ascent method is used to perform local fine-tuning in the neighborhood of the candidate solutions.
[0097] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which can take the form of a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email sending and receiving device, game console, tablet computer, wearable device, or any combination of these devices.
[0098] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0099] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0100] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0101] For any computer-readable medium (or computer-readable storage medium) as described above or otherwise, computer instructions may be stored thereon, which, when executed by a processor, implement one or more of the above embodiments, thereby realizing the technical solution of the present invention.
[0102] The present invention also proposes a computer program that, when executed by a processor, implements one or more of the embodiments described above, thereby realizing the technical solution of the present invention. This computer program may be specifically recorded on the above-described or other computer-readable media, and the present invention does not impose any limitations on this.
[0103] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0104] The foregoing has described specific embodiments of the invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.
[0105] The terminology used in one or more embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in one or more embodiments of the invention and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items.
[0106] It should be understood that although the terms first, second, third, etc., may be used to describe various information in one or more embodiments of the present invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of one or more embodiments of the present invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."
[0107] The above description is merely a preferred embodiment of one or more embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of the present invention should be included within the protection scope of one or more embodiments of the present invention.
Claims
1. A method for optimizing thin-film coating parameters of a TMR sensor chip, characterized in that, include: Acquire historical parameter data of the thin film coating of the TMR sensor chip and the corresponding chip performance index data, and generate a training dataset; A deep learning model is constructed and trained based on the training dataset to obtain a mapping model. The mapping model is used to establish a nonlinear mapping relationship between coating process parameters and chip performance indicators. The mapping model includes: a fully connected neural network module for extracting static nonlinear features of the coating process parameters; a convolutional neural network module for extracting spatial correlation features between coating process parameters arranged according to thin film structure; a long short-term memory network module for extracting dynamic features of coating process parameters with time series characteristics; and a splicing module for splicing the outputs of multiple neural network modules included in the mapping model. A parameter optimization problem is constructed based on the mapping model and solved within the feasible region of the coating process parameters to obtain the target parameter combination that optimizes the chip performance.
2. The method according to claim 1, characterized in that, The fully connected neural network module includes multiple fully connected layers that use linear rectified functions as activation functions, and dropout layers set between the fully connected layers; the convolutional neural network module includes one-dimensional convolutional layers, pooling layers, and flattening layers; the long short-term memory network module includes at least one LSTM layer.
3. The method as described in claim 1, characterized in that, The method further includes: Data cleaning operations are performed on the acquired historical parameter data and chip performance index data; wherein, the cleaning operations include at least one of the following: outlier detection and processing, missing value imputation, and unit unification; Normalize or standardize the cleaned data; The processed data is divided into training dataset, validation dataset, and test dataset according to a predetermined ratio.
4. The method as described in claim 3, characterized in that, The step of constructing and training a deep learning model based on the training dataset includes: A deep neural network is constructed, comprising an input layer, at least one hidden layer, and an output layer; wherein the number of nodes in the input layer corresponds to the number of coating process parameters, and the number of nodes in the output layer corresponds to the number of chip performance indicators. Determine the loss function and select the optimizer; The deep neural network is trained using the training dataset, and the training process is monitored using the validation dataset to adjust hyperparameters until the model converges.
5. The method according to claim 4, characterized in that, The determination of the loss function includes: When the mapping model is used to predict a single performance metric, the mean squared error loss function is determined as the loss function of the mapping model. When the mapping model is used to predict multiple performance indicators, the weighted multi-objective loss function is determined as the loss function of the mapping model, and the weighted multi-objective loss function is the weighted sum of the mean squared error losses of the multiple performance indicators; When process constraints exist, a penalty term is introduced into the loss function of the mapping model. The penalty term is used to constrain process parameters to remain within a preset physically reasonable range.
6. The method according to claim 1, characterized in that, The step of constructing a parameter optimization problem based on the mapping model and solving it within the feasible region of the coating process parameters includes: Determine the parameter vector composed of coating process parameters, and use chip performance indicators as the optimization objective function; Determine the upper and lower bound constraints of each coating process parameter to form a feasible region; Based on the differentiability of the deep learning model and the complexity of the objective function, the gradient ascent method, global optimization algorithm, or hybrid optimization strategy is selected to solve the optimization problem in order to obtain the combination of objective parameters.
7. The method according to claim 6, characterized in that, The step of selecting gradient ascent, global optimization algorithm, or hybrid optimization strategy to solve the optimization problem based on the differentiability of the deep learning model and the complexity of the objective function includes: When the deep learning model is differentiable and the objective function is a convex function, the gradient ascent method is used to iteratively update the process parameters. After each iteration, the parameters are projected into the feasible region through a pruning operation until convergence. When the objective function is non-differentiable or has multiple local optima, a global search is performed using a genetic algorithm or particle swarm optimization algorithm, and the solution is solved iteratively through fitness evaluation, selection, crossover, mutation or velocity and position update operations. When the system contains discrete parameters or has complex constraints, a global search is first performed using a genetic algorithm or particle swarm optimization algorithm to obtain candidate solutions, and then a gradient ascent method is used to perform local fine-tuning in the neighborhood of the candidate solutions.
8. A device for optimizing thin-film coating parameters of a TMR sensor chip, characterized in that, The device includes: Acquisition Unit: Acquires historical parameter data of the thin-film coating of the TMR sensor chip and the corresponding chip performance index data, and generates a training dataset; Training Unit: Constructs and trains a deep learning model based on the training dataset to obtain a mapping model; wherein, the mapping model is used to establish a nonlinear mapping relationship between coating process parameters and chip performance indicators, and the mapping model includes: a fully connected neural network module for extracting the static nonlinear features of the coating process parameters, a convolutional neural network module for extracting the spatial correlation features between coating process parameters arranged according to the thin film structure, a long short-term memory network module for extracting the dynamic features of coating process parameters with time series characteristics, and a splicing module for splicing the outputs of multiple neural network modules included in the mapping model; Solving unit: Constructs a parameter optimization problem based on the mapping model and solves it within the feasible region of the coating process parameters to obtain the target parameter combination that optimizes the chip performance.
9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor implements the steps of the method as described in any one of claims 1-7 by running the executable instructions.
10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.