MEMS sensor intelligent design method and system for general design scenarios

Abstract

The application discloses a kind of general design scene-oriented MEMS sensor intelligent design method and system, which comprises: a large number of "structure parameter-performance parameter" data pairs of MEMS sensor are obtained based on finite element simulation;Mapping relationship from structure parameter to performance parameter is established by constructing and training verification network;Three different design requirements of fuzzy range, accurate numerical value and unconstrained are quantified as unified interval representation, and data set containing design requirements is constructed using reverse generation strategy;Mapping relationship from design requirement to feasible structure is realized by constructing and training search network;Precision optimizer is constructed to improve the quality of feasible solution output by search network.Compared with traditional intelligent design model, the model of the application has strong expansibility, can support multiple design requirement types without retraining, and has the ability to automatically iterate and optimize precision, thereby realizing high-precision, fast MEMS sensor design.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of MEMS sensor design, and particularly relates to a MEMS sensor intelligent design method and system for general design scenarios. BACKGROUND

[0002] MEMS (Micro-Electro-Mechanical Systems) sensors are sensors made by combining microelectronic technology and micromechanical technology, which realize the functions of sensing, converting and outputting signals by integrating mechanical structures, sensing elements, signal processing circuits, etc. on the micron or even nanometer scale. With the further popularization of Internet of Things technology, MEMS devices will play a core role in more scenarios, including smart cities, intelligent transportation, smart homes, etc. However, the traditional design method of MEMS sensors relies heavily on time-consuming finite element simulation and requires designers to have rich design experience to iteratively fine-tune the design, making it difficult to meet the diverse market demand.

[0003] In recent years, data-driven deep learning algorithms have made significant progress in the field of MEMS sensor design. Compared with other machine learning-based models, these deep learning models provide extremely high computational speed and prediction accuracy without prior knowledge of the mapping relationship between data. Therefore, researchers have begun to explore the combination of deep learning technology into MEMS design, realizing an intelligent design model that directly outputs a feasible structure given a design requirement. However, the design requirements of MEMS sensors in real-world scenarios are usually ambiguous and diverse, such as some design parameters requiring accuracy to a certain specific value, some design parameters needing to be constrained within a certain range, and some design parameters being negligible. Existing intelligent design models usually only focus on specific and accurate design requirements, and when faced with ambiguous design requirements and design scenarios where some parameters are not constrained, the model's expansion ability is weak, and it lacks the ability to correct errors when faced with contradictory inputs. SUMMARY

[0004] To solve the problem of generating a feasible structure from ambiguous design requirements for MEMS sensor design, the present application proposes a MEMS sensor intelligent design method and system for general design scenarios, which has strong scalability, supports multiple design requirement types without retraining, and has the ability to automatically iterate and optimize accuracy, thereby realizing high-precision and fast MEMS sensor design.

[0005] To achieve the above purpose, the present application adopts the following technical solutions:

[0006] A MEMS sensor intelligent design method for general design scenarios, comprising the following steps:

[0007] Step S1: obtaining MEMS sensor experimental data based on finite element simulation;

[0008] Step S2: constructing and training a verification network to realize the mapping relationship from structure parameters to performance parameters based on MEMS sensor experimental data;

[0009] Step S3: generating data based on the verification network, quantifying different types of design requirements, and using a reverse generation strategy to construct a dataset containing design requirements;

[0010] Step S4: constructing and training a search network to realize the mapping relationship from design requirements to feasible structures based on the dataset;

[0011] Step S5: constructing a precision optimizer to optimize the quality of the feasible solution output by the search network.

[0012] Further, in step S1, the experimental data of the MEMS sensor includes structure parameters X and performance parameters Y; wherein X contains continuous variables describing the geometry of the sensor, such as length, width, thickness, and other parameters, and Y contains continuous variables describing the performance of the sensor, such as range, resonant frequency, sensitivity, and other parameters, which can be represented as:

[0013] X:(x1,x2,…,x m )

[0014] Y:(y1,y2,…,y n )

[0015] Where m and n are the dimensions of X and Y, respectively.

[0016] Further, in step S2, the verification network is an artificial neural network with input dimension m and output dimension n; the verification network uses the min-max normalized dataset in step S1 to establish the mapping relationship from structure parameters X to performance parameters Y, and is continuously iteratively trained through the backpropagation algorithm, using mean square error as the loss function for network training, which is shown in the formula as follows:

[0017]

[0018] Where LVN is the error value used for backpropagation of the verification network, yi is the true performance parameter, is the performance parameter predicted by the verification network.

[0019] Further, in step S3, the different types of design requirements refer to the common fuzzy range design requirements in MEMS sensor design requirements (such as "range greater than a certain value"), precise numerical design requirements (such as "frequency equal to a certain value"), and types without performance constraints. These different types of design requirements are quantified as intervals with a unified form, and the quantification method is shown as follows:

[0020]

[0021] where A and B represent constants that define the upper or lower limit of the design requirement, respectively, y i is the i-th performance parameter to be constrained, c r is the r-th constraint in the design requirement, 0≤r≤(m+n), c rl is the lower bound of the constraint, c ru is the upper bound of the constraint; for the design requirement without constraints, the quantization method is as follows:

[0022] c r : y i ∈(-∞,+∞)

[0023] Generally, negative infinity and positive infinity are replaced by a pair of minimum and maximum numbers; in the normalized expression, 0 and 1 represent infinitesimal and infinity, respectively.

[0024] Further, step S3 generates the training data of the search network through the inverse generation strategy in the following steps:

[0025] Step S3-1, using the trained verification network in step S2, randomly generates (X, Y) data pairs.

[0026] Step S3-2, for each parameter in the (X, Y) data pair generated in step S3-1, the corresponding design requirement is generated using the following formula:

[0027] c xl = x-a1

[0028] c xu = x+(d1-a1)

[0029]

[0030] where d1 and d2 are random numbers generated in the range [0, 1], a1 and a2 are random numbers generated in the range [0, d]; c xl and c xu are the lower and upper bounds of the constraint of x, c yl and c yu are the lower and upper bounds of the constraint of y, that is, x∈(c xl , c xu ), y∈(c yl , c yu ); if c xl <0, c yl <0, c xu >1 or c yu >1, then regenerate until the design requirement C corresponding to the constraint range of the data pair is obtainedlu :(c x1l ,c x1u ,…,c xml ,c xmu ,c y1l ,c y1u ,…,c ynl ,c ynu )。

[0031] Step S3-3, repeat step S3-1 and S3-2 to generate a data set (C lu , X, Y) required for training the search network.

[0032] Further, in step S4, when training the search network, first input the constraint condition Clu into the search network to obtain the predicted structure parameter X * , and then input X * into the verification network trained in step S2 to obtain Y * ; then calculate the error through the loss function, and continuously iterate the training through the back propagation algorithm until the error of the search network reaches the expectation; wherein the calculation formula of the loss function is as follows:

[0033]

[0034] Wherein, LSN is the error value of the search network used for back propagation, γ is the weight parameter of the error caused by the performance parameter, c xi1 、c xiu 、c yj1 、c yju ∈C lu ; the specific calculation of the function T(a, b, c) is as follows:

[0035]

[0036] Wherein a and c must satisfy a≤c, a, b, c are variables of the T function.

[0037] In step S5, the accuracy optimizer optimizes the initial predicted structure parameter X generated by the search network through the following process:

[0038] Step S5-1, assign the original constraint condition Clu0 to the current constraint condition Cluc respectively; set the final error Errf to 9999; set the learning rate η;

[0039] Step S5-2, input Cluc into the search network trained in step S4 to obtain the current predicted structure parameter X ; input X into the verification network trained in step S2 to obtain the current predicted performance parameter The current error is calculated by searching the loss function of the network:

[0040] In step S5-3, Cluc is updated by the following formula:

[0041]

[0042] Where C' is the updated value of C, Errc < Errf, Errf, and the final structure parameter luc luc are set as Errc,

[0043] In step S5-4, steps S5-2 and S5-3 are repeated to make C luc iterative setting I0 times.

[0044] In step S5-5, Errf,

[0045] A MEMS sensor intelligent design system for general design scenarios includes:

[0046] A MEMS sensor experimental data acquisition unit obtains MEMS sensor experimental data based on finite element simulation;

[0047] A verification network construction and training unit constructs and trains a verification network to realize a mapping relationship from a structure parameter to a performance parameter based on the MEMS sensor experimental data;

[0048] A data set construction unit generates data based on the verification network, quantifies different types of design requirements, and uses a reverse generation strategy to construct a data set containing design requirements;

[0049] A search network construction and training unit constructs and trains a search network to realize a mapping relationship from a design requirement to a feasible structure based on the data set;

[0050] A search network optimization unit constructs a precision optimizer to optimize the quality of the feasible solution output by the search network.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] (1) More practical: for fuzzy interval design scenarios, the search network can directly output a feasible solution end-to-end without traversing the design space;

[0053] ​​​​​(2) Better scalability: The design scenario is not constrained by partial parameters, and there is no need to retrain a new model to meet various design requirements;

[0054] (3) Better reliability: The verification module can verify the generated scheme, and automatically iteratively adjust the precision optimizer for high-error schemes to improve precision;

[0055] (4) High efficiency: The running time of a single search network and precision optimizer is less than milliseconds, and a reliable design scheme can be generated in a very short time. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 is a flowchart of the method shown in the present application;

[0058] Figure 2 is a model structure of the verification network provided by the embodiment of the present application;

[0059] Figure 3 is a model structure of the search network provided by the embodiment of the present application;

[0060] Figure 4 is an algorithm flowchart of the precision optimizer provided by the embodiment of the present application;

[0061] Figure 5 is an example schematic diagram of the model architecture verification provided by the embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present application more clear and explicit, the present application will be further described in detail below in combination with its embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and do not limit the protection scope of the present application.

[0063] The specific embodiments of the present application are as follows:

[0064] The present application provides an end-to-end model which directly maps design requirements to feasible structures using deep learning. The model has the ability to adapt to fuzzy design requirements and automatically optimize precision, and its high running precision and efficient running rate solves the limitations of traditional design methods based on finite element simulation and models based on deep learning in efficiency and function.

[0065] The following describes the design optimization method for MEMS sensors based on generative networks, using specific embodiments.

[0066] Figure 1 A flowchart of the present invention is shown. Exemplarily, this intelligent design method for MEMS sensors for general design scenarios includes the following steps:

[0067] Step S1: Obtain experimental data of MEMS sensors based on finite element simulation.

[0068] Two thousand sets of experimental data for piezoresistive accelerometers were obtained using the finite element method (FEM). The descriptive statistics are shown in Table 1. The experimental data include structural parameters: the side length L of the mass block... m Thickness T m And the length L of the cantilever beam b , width L w Thick T b And performance parameters: Sensor range R s Sensitivity S s and resonant frequency f s They can be represented as (x1, x2, x3, x4, x5) and (y1, y2, y3) respectively. Wherein, the structural parameter L... m T m L b L w and T b The values ​​are random integers in the ranges of 800 to 1000 μm, 150 to 300 μm, 20 to 50 μm, 2500 to 3000 μm, and 400 to 500 μm, respectively.

[0069] Preferably, to prevent drastic changes in the weight parameter ω during the training of the neural network and to accelerate model convergence, the experimental data can be normalized. The normalization calculation method is as follows:

[0070]

[0071] Where x norm Here are the normalized sensor parameter values, and x represents the original data. max x represents the maximum value of each feature parameter in the original data. min The minimum value of each feature parameter in the original data is given; the experimental data is divided into training set, test set and validation set in a ratio of 6:2:2.

[0072] Step S2: Construct and train the validation network to realize the mapping relationship from structural parameters to performance parameters;

[0073] like Figure 2As shown, the input dimension of the verification network is 5, the output dimension is 3, and the mean square error is used as the loss function for network training, and its formula is as follows:

[0074]

[0075] where L VN is the error value used for backpropagation of the verification network, y i is the true performance parameter, is the predicted performance parameter.

[0076] To ensure the accuracy of the artificial neural network, the network parameters are initialized using the Xavier method, the Leaky ReLU function is used as the activation function, the Adam optimizer is used for backpropagation, and the ridge regression is used for regularization.

[0077] To make full use of the experimental data generated in step S1, the number of hidden layers of the verification network is finally determined to be 3, the number of neurons in the hidden layer is 20-20-30, and the optimal batch size Batch Size is 10 through ten-fold cross-validation.

[0078] After determining the structure parameters of the network, the verification network is trained, and the average absolute error of the sensor performance parameters R s , S s and f s output by the verification network is less than 1%, which enables it to have high-precision verification capabilities and generate training data for the search network.

[0079] Step S3: Quantify different types of design requirements and use the inverse generation strategy to build a dataset containing design requirements;

[0080] Different types of design requirements refer to the common fuzzy range design requirements in MEMS sensor design requirements (such as "range greater than a certain value"), precise numerical design requirements (such as "frequency equal to a certain value"), and types without performance constraints. These different types of design requirements are quantified as intervals with a unified form, and the quantification method is as follows:

[0081]

[0082] where A and B represent the constants of the upper or lower limit of the design requirement, y i is the i-th performance parameter to be constrained, c r is the r-th constraint in the design requirement (0≤r≤8), c rl is the lower bound of the constraint, and c ru is the upper bound of the constraint; for the unconstrained design requirement, the quantification method is as follows:

[0083] cr : y i ∈ [0, 1]

[0084] The training data of the search network is generated by the inverse generation strategy in the following steps:

[0085] (3-1) Randomly generate (X, Y) data pairs using the verification network trained in step S2.

[0086] (3-2) For each parameter in the (X, Y) data pairs generated in (3-1), generate the corresponding design requirements using the following formula:

[0087] c xl = x - a1

[0088] c xu = x + (d1 - a1)

[0089]

[0090] Where d1 and d2 are uniformly distributed random numbers generated in the range [0, 1], a1 and a2 are uniformly distributed random numbers generated in the range [0, d]; c xl and c xu are the lower and upper bounds of the constraints of x, c yl and c yu are the lower and upper bounds of the constraints of y, i.e., x ∈ (c xl , c xu ), y ∈ (c yl , c yu ); if c xl < 0, c yl < 0, c xu > 1 or c yu > 1, then regenerate until the design requirements C lu : (c x1l , c x1u , …, c xml , c xmu , c y1l , c y1u , …, c ynl , c ynu ) corresponding to the constraint range of the data pair are obtained.

[0091] (3-3) Repeat steps (3-1) and (3-2) to generate 20,000 sets of data sets (C lu , X, Y) required for training the search network.

[0092] Step S4: Construct and train the search network to realize the mapping relationship from design requirements to feasible structures.

[0093] As Figure 3As shown, the input dimension of the search network is 16 and the output dimension is 5; the generated 20000 sets of training data are divided into training set and validation set in the ratio of 7:3.

[0094] The constraint condition C lu is input into the search network to obtain the predicted structure parameter X * ; then X * is input into the trained validation network in step S2 to obtain Y * ; then the error is calculated by the loss function, and the training is continuously iterated by the back propagation algorithm until the error of the search network reaches the expectation; wherein the calculation formula of the loss function is as follows:

[0095]

[0096] Wherein, wherein L SN is the error value used for back propagation of the search network, and γ is the weight parameter of the error caused by the performance parameter, which is set to 1; the specific calculation of the function T(a, b, c) is as follows:

[0097]

[0098] Wherein a and c must satisfy a≤c.

[0099] The performance of the search network is represented by the design accuracy DA:

[0100]

[0101] Wherein N d is the total number of constraints in the data set, which is 160000, and N s is the number of parameters that meet the design requirements, which refers to the number of parameters in the search network output X and the validation network output Y whose error is 0.

[0102] Through the error of the validation set, it is determined that the number of hidden layers of the search network is 3, the number of neurons in the hidden layer is 40-70-10, and the optimal Batch Size is further determined to be 10.

[0103] In addition, the degree of deviation of the data C lu from the uniform distribution also affects the loss and DA of the model. θ is used to represent the degree of deviation of the data from the uniform distribution. By adjusting θ from 0° to 90°, various different distributions can be obtained, and they are mixed with the uniform distribution data for training the search network. The results show that when the data mixed distribution is 0° and 40°, the loss of the network is the smallest, which is 2.51×10 -4 ; DA is the highest, which is 92.6%.

[0104] Step S5: Constructing the precision optimizer to improve the quality of the feasible solution output by the search network.

[0105] As Figure 4 shown, the precision optimizer optimizes the initial prediction structure parameters generated by the search network through the following steps:

[0106] (5-1) Assign the original constraint condition C lu0 to the current constraint condition C luc respectively; set the final error Err f to 9999; set the learning rate η;

[0107] (5-2) Input C luc to the search network trained in step S3 to obtain the current prediction structure parameter Input to the validation network trained in step S2 to obtain the current performance parameter Calculate the current error through the loss function of the search network:

[0108] (5-3) Update C luc by the following formula:

[0109]

[0110] Err c < Err f , then set Err f , the final structure parameter and the final performance parameter to Err c , and

[0111] (5-4) Repeat steps (5-2) and (5-3) I0 times to make C luc iterate I0 times;

[0112] (5-5) Output Err f , and

[0113] A MEMS sensor intelligent design system for general design scenarios, comprising:

[0114] A MEMS sensor experimental data acquisition unit acquires MEMS sensor experimental data based on finite element simulation;

[0115] A validation network construction and training unit constructs and trains a validation network based on MEMS sensor experimental data to realize the mapping relationship from structure parameters to performance parameters;

[0116] ​A data set construction unit generates data based on the verification network, quantifies different types of design requirements, and uses a reverse generation strategy to construct a data set containing design requirements;

[0117] A search network construction and training unit constructs and trains a search network based on the data set to achieve a mapping relationship from design requirements to feasible structures;

[0118] A search network optimization unit constructs a precision optimizer to optimize the quality of the feasible solution output by the search network.

[0119] Figure 5 An example of verifying the model architecture by the embodiment of the present application is shown. In the face of different types of design requirements, the search model can directly output high-precision feasible solutions, and the calculation time is about 0.1 milliseconds. For schemes that do not meet the design requirements, the precision optimizer in step S5 reduces the design error through iteration, and the single iteration time of the precision optimizer is also 0.1 milliseconds. It proves the end-to-end design capability of the present application in the face of different design scenarios, and the effect of the proposed precision optimizer.

[0120] The above is only a specific embodiment of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A MEMS sensor intelligent design method for general design scenarios, characterized in that, include: Step S1: Obtain experimental data of MEMS sensors based on finite element simulation; Step S2: Based on the experimental data of MEMS sensors, construct and train a verification network to realize the mapping relationship from structural parameters to performance parameters; Step S3: Based on the data generated by the validation network, quantify different types of design requirements, and use a reverse generation strategy to construct a dataset containing design requirements; Step S4: Based on the dataset, construct and train a search network to realize the mapping relationship from design requirements to feasible structures; Step S5: Construct a precision optimizer to improve the quality of feasible solutions output by the search network; The quantification method for different types of design requirements described in step S3 is as follows: where A and B represent constants that define the upper or lower limit of the design requirement, respectively, y i is the ith constrained performance parameter, c r is the rth constraint in the design requirement, 0≤r≤(m+n), c rl is the lower bound of the constraint, c ru is the upper bound of the constraint; For unconstrained design requirements, the quantification method is as follows: c r :y i ∈(-∞,+∞) In the normalized representation, 0 and 1 are used to represent infinitesimal and infinity, respectively; In step S4, when training the search network, the constraint condition C lu is input into the search network to obtain a predicted structure parameter X * , then X * is input into the verification network trained in step S2 to obtain Y * , and then the error is calculated by a loss function, and the training is iterated by a back propagation algorithm until the error of the search network reaches the expectation. The loss function is: where L SN is the error value for the backpropagation of the search network, and γ is a weight parameter for the error caused by the performance parameter, c xi1 , c xiu , c yj1 , c yju ∈ C lu ; the function T is: Where a and c must satisfy a≤c, and a, b, and c are variables of function T. 2.The MEMS sensor intelligent design method for general design scenarios according to claim 1, wherein, The MEMS sensor experimental data mentioned in step S1 includes structural parameters X and performance parameters Y; wherein, structural parameters X contain continuous variables describing the sensor's geometry, with a dimension of m; and performance parameters Y contain continuous variables describing the sensor's performance, with a dimension of n. 3.The MEMS sensor intelligent design method for general design scenarios according to claim 1, wherein, The verification network described in step S2 is an artificial neural network with an input dimension of m and an output dimension of n. The verification network uses the dataset that has been normalized by min-max in step S1 to establish a mapping relationship from structural parameters X to performance parameters Y, and continuously trains iteratively through the backpropagation algorithm, using mean squared error as the loss function for network training.

4. The MEMS sensor intelligent design method for general design scenarios according to claim 3, characterized in that, The validation network is initialized using the Xavier method, uses the Learny ReLU function as the activation function, uses the Adam optimizer for backpropagation, and uses ridge regression for regularization.

5. The MEMS sensor intelligent design method for general design scenarios according to claim 1, characterized in that, Step S3 uses a reverse generation strategy to construct a dataset containing design requirements, specifically including: Step S3-1: Using the validation network trained in step S2, randomly generate (X, Y) data pairs; Step S3-2: For each parameter in the (X, Y) data pair generated in step S3-1, generate the corresponding design requirements using the following formula; c xl = x - a1 c xu = x + (d1 - a1) wherein d1 and d2 are random numbers generated in the range [0, 1], a1 and a2 are random numbers generated in the range [0, d]; c xl With c xu The lower and upper bounds of the constraint of x, c yl With c yu The lower and upper bounds of the constraint of y, that is, x∈(c xl , c xu ), y∈(c yl , c yu ); if c xl <0, c yl <0, c xu >1 or c yu >1, then regenerate until the design requirement C lu :(c x1l , c x1u , …, c xml , c xmu , c y1l , c y1u , …, c ynl , c ynu ) corresponding to the constraint range of the group of data is obtained; Step S3-3, repeat steps S3-1 and S3-2 to generate the dataset (C lu , X, Y) required for training the search network.

6. The MEMS sensor intelligent design method for general design scenarios according to claim 1, characterized in that, Constructing a precision optimizer to improve the quality of feasible solutions output by the search network specifically includes: Step S5-1, assign the original constraint condition C lu0 to the current constraint condition C luc ; set the final error Err f to 9999; set the learning rate η; Step S5-2, inputting the C luc into the trained search network of step S4 to obtain the current predicted structure parameters into the trained validation network of step S2 to obtain the current predicted performance parameters into the trained validation network of step S2 to obtain the current predicted performance parameters The current error is calculated by the loss function of the search network: Step S5-3, update C by the following equation luc : Where C′ luc C luc The updated value, if Err c <Err f Then let Err f Final structural parameters Final performance parameters The values ​​are respectively Err c , and Step S5-4, repeating steps S5-2, S5-3 to make C luc I0 times of iteration are set. Step S5-5, output Err f , and 7. A MEMS sensor intelligent design system for general design scenarios that implements the method of any one of claims 1-6, characterized in that, include: The MEMS sensor experimental data acquisition unit acquires MEMS sensor experimental data based on finite element simulation. Verification network construction and training unit: Based on MEMS sensor experimental data, construct and train a verification network to realize the mapping relationship from structural parameters to performance parameters; The dataset construction unit generates data based on the validation network, quantifies different types of design requirements, and uses a reverse generation strategy to construct a dataset containing design requirements; Search network construction and training units: Based on the dataset, construct and train the search network to realize the mapping relationship from design requirements to feasible structures; Search network optimization units are used to construct a precision optimizer to improve the quality of feasible solutions output by the search network.

Images