MEMS sensor intelligent design method and system oriented to general design scene

Abstract

The invention discloses an MEMS sensor intelligent design method and system oriented to a universal design scene. The method comprises the steps that a large number of structural parameter-performance parameter data pairs of MEMS sensors are obtained based on finite element simulation; constructing and training a verification network to establish a mapping relation from structure parameters to performance parameters; the fuzzy range, the accurate numerical value and the unconstrained three different design demands are converted into unified interval representation, and a reverse generation strategy is used to construct a data set containing the design demands; constructing and training a search network to realize a mapping relation from a design demand to a feasible structure; and a precision optimizer is constructed to improve the quality of a feasible solution output by the search network. Compared with a traditional intelligent design model, the model expansibility is high, multiple design requirement types can be supported without retraining, the capability of automatically iterating and optimizing precision is achieved, and therefore high-precision and rapid MEMS sensor design is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of MEMS sensor design, and in particular relates to a MEMS sensor intelligent design method and system for general design scenarios. Background Art

[0002] MEMS (Micro-Electro-Mechanical Systems) sensors are sensors made by combining microelectronics and micromechanical technologies. They realize the functions of sensing, converting and outputting signals by integrating mechanical structures, sensor elements, signal processing circuits, etc. at the micrometer 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, smart transportation, smart homes, etc. However, the traditional design method of MEMS sensors relies heavily on time-consuming finite element simulation, and also requires designers to have rich design experience to iteratively fine-tune the design, which makes it difficult to meet the broad and diverse market needs.

[0003] In recent years, data-driven deep learning algorithms have made significant progress in the design of MEMS sensors. Compared with other machine learning-based models, these deep learning models provide extremely high computing speed and prediction accuracy without the need to know the mapping relationship between data in advance. Therefore, researchers have begun to explore the integration of deep learning technology into MEMS design, realizing intelligent design models that directly output feasible structures under given design requirements. However, the design requirements of MEMS sensors in real scenarios are usually vague and diverse, such as: some design parameters are required to be accurate to a certain value, some design parameters need to be constrained within a certain range, and some design parameters can be ignored. Existing intelligent design models usually only focus on specific and precise design requirements. When faced with fuzzy design requirements and design scenarios where some parameters are not constrained, the model has weak scalability and lacks the ability to correct errors when faced with contradictory inputs. Summary of the invention

[0004] In order to solve the problem of generating feasible structures from fuzzy design requirements in MEMS sensor design, the present invention proposes a MEMS sensor intelligent design method and system for general design scenarios, which has strong scalability, can support multiple types of design requirements 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 object, the present invention adopts the following technical solution:

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

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

[0008] Step S2: Based on the MEMS sensor experimental data, construct and train a verification network to realize the mapping relationship from structural parameters to performance parameters;

[0009] Step S3: Generate data based on the verification network, quantify different types of design requirements, and use the reverse generation strategy to construct a data set containing design requirements;

[0010] Step S4: Based on the data set, construct and train a search network to achieve a mapping relationship from design requirements to feasible structures;

[0011] Step S5: construct an accuracy optimizer to optimize the quality of feasible solutions output by the search network.

[0012] Furthermore, in step S1, the experimental data of the MEMS sensor includes structural parameters X and performance parameters Y; wherein X includes continuous variables describing the geometric structure of the sensor, such as parameters such as length, width, thickness, and height, and Y includes continuous variables describing the performance of the sensor, such as parameters such as range, resonant frequency, and sensitivity, which can be expressed 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] Furthermore, in step S2, the verification network is an artificial neural network with an input dimension of m and an output dimension of n; the verification network uses the data set normalized by min-max in step S1 to establish a mapping relationship from the structural parameter X to the performance parameter Y, and continuously iterates the training through the back propagation algorithm, using the mean square error as the loss function of the network training, and its formula is as follows:

[0017]

[0018] Where LVN is the error value used for back propagation in the verification network, yi is the actual performance parameter, It is a performance parameter to verify the network prediction.

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

[0020]

[0021] Among them, A and B represent the constants of the upper or lower limit of the design requirements, respectively, and y i is the constrained i-th 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:

[0022] c r :y i ∈(-∞,+∞)

[0023] Usually, negative infinity and positive infinity are represented by a pair of extremely small and extremely large numbers; in normalized representation, 0 and 1 are used to represent infinitesimal and infinity, respectively.

[0024] Furthermore, step S3 generates training data for the search network by using the reverse generation strategy in the following steps:

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

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

[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 With c xu is the lower and upper bounds of x, c yl With c yu is the lower and upper bounds 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 this group of data is obtainedlu :(c x1l , c x1u , …, c xml , c xmu , c y1l , c y1u , …, c ynl , c ynu ).

[0031] Step S3-3, repeat steps S3-1 and S3-2 to generate the data set (C lu , X, Y).

[0032] Furthermore, in step S4, when the search network is trained, the constraint condition Clu is first input into the search network to obtain the predicted structural parameter X * , and then X * Input into the validation network trained in step S2 to obtain Y * ; Then the error is calculated through the loss function, and the back propagation algorithm is used to iterate the training until the search network error reaches the expected value; the calculation formula of the loss function is as follows:

[0033]

[0034] Among them, LSN is the error value of the search network 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 function T(a,b,c) is as follows:

[0035]

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

[0037] In step S5, the precision optimizer performs the following process to optimize the initial predicted structural parameters generated by the search network: To optimize:

[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 structural parameters Will Input into the trained validation network in step S2 to obtain the current prediction performance parameters Calculate the current error through the loss function of the search network:

[0040] In step S5-3, update Cluc through the following formula:

[0041]

[0042] where C′ luc is the updated value of C luc If Errc < Errf, then set Errf, the final structural parameters the final performance parameters to be Errc, and

[0043] In step S5-4, repeat steps S5-2 and S5-3 to iterate I0 times for the setting of C luc ;

[0044] In step S5-5, output Errf, and

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

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

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

[0048] A dataset construction unit, which generates data based on the verification network, quantifies different types of design requirements, and constructs a dataset containing design requirements using an inverse generation strategy;

[0049] A search network construction and training unit, which constructs and trains a search network based on the dataset to realize the mapping relationship from design requirements to feasible structures;

[0050] A search network optimization unit, which constructs an accuracy optimizer to optimize the quality of the feasible solutions output by the search network.

[0051] Compared with the prior art, the remarkable advantages of the present invention are:

[0052] (1) Stronger practicability: For the design scenario of the fuzzy interval, there is no need to traverse the design space, and the search network can directly output feasible solutions end-to-end;

[0053] (2) Better scalability: For design scenarios where some parameters are not constrained, there is no need to retrain new models to meet various design requirements;

[0054] (3) Better reliability: The verification module can verify the generated solutions and use the precision optimizer to automatically iterate and adjust the solutions with high errors to improve the accuracy.

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

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 It is a schematic diagram of the process of the method shown in the present invention;

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

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

[0060] Figure 4 is an algorithm flow chart of the precision optimizer provided by an embodiment of the present invention;

[0061] Figure 5 It is a schematic diagram of an example of verifying the model architecture according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with its embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

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

[0064] The present invention provides an end-to-end model that uses deep learning to directly map design requirements to feasible structures. The model has the adaptability to fuzzy design requirements and the ability to automatically optimize accuracy. Its high operating accuracy and efficient operating rate solve the limitations of traditional finite element simulation-based design methods and deep learning-based models in terms of efficiency and functionality.

[0065] The MEMS sensor design optimization method based on generative networks is described below in conjunction with specific embodiments.

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

[0067] Step S1: Acquire MEMS sensor experimental data based on finite element simulation.

[0068] 2000 sets of piezoresistive accelerometer experimental data were obtained by finite element simulation method, and their descriptive statistics are shown in Table 1. The experimental data include structural parameters: 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-shirt b And performance parameters: sensor range R s , Sensitivity S s and the resonant frequency f s , which can be expressed as (x1, x2, x3, x4, x5) and (y1, y2, y3) respectively. Among them, the structural parameter L m , T m , L b , L w and T b The values ​​of 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, in order to prevent the weight parameter ω from changing drastically during the training process of the neural network and to accelerate the convergence of the model, the experimental data can be normalized, and the normalization calculation method is as follows:

[0070]

[0071] where x norm is the normalized sensor parameter value, x is the original data, and x max is the maximum value of each characteristic parameter in the original data, x min is the minimum value of each feature parameter in the original data; the experimental data is divided into training set, test set and validation set in the ratio of 6:2:2.

[0072] Step S2: construct and train a verification 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 of the network training. The formula is as follows:

[0074]

[0075] Among them, L VN To verify the error value of the network for back propagation, y i is the actual performance parameter, is the predicted performance parameter.

[0076] Preferably, to ensure the accuracy of the artificial neural network, the network parameters are initialized using the Xavier method, the Learky ReLU function is used as the activation function, the Adam optimizer is used for back propagation, and the ridge regression is used for regularization.

[0077] In order to make full use of the experimental data generated in step S1, the ten-fold cross-validation method was used to finally determine that the number of hidden layers of the verification network is 3, the number of neurons in the hidden layer is 20-20-30, and the optimal batch size is 10.

[0078] After determining the structural parameters of the network, the verification network is trained and the sensor performance parameters R output by the verification network are finally obtained. s , S s and f s The mean absolute error of the proposed method is less than 1%, which enables high-precision verification and the ability to generate training data for search networks.

[0079] Step S3: quantify different types of design requirements and use a reverse generation strategy to construct a data set containing the design requirements;

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

[0081]

[0082] Among them, A and B represent the constants of the upper or lower limit of the design requirements, respectively, and y i is the constrained i-th performance parameter, c r is the rth constraint in the design requirement (0≤r≤8), 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:

[0083] cr :y i ∈[0,1]

[0084] Generate training data for the search network using the reverse generation strategy in the following steps:

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

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

[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 With c xu is the lower and upper bounds of x, c yl With c yu is the lower and upper bounds 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 this group of data is obtained lu :(c x1l , c x1u , …, c xml , c xmu , c y1l , c y1u , …, c ynl , c ynu ).

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

[0092] Step S4: construct and train the search network to achieve the mapping relationship from design requirements to feasible structures;

[0093] like Figure 3As shown in the figure, the input dimension of the search network is 16 and the output dimension is 5; the generated 20,000 sets of training data are divided into training set and validation set in a ratio of 7:3.

[0094] The constraint C lu Input into the search network to get the predicted structural parameter X * , and then X * Input into the validation network trained in step S2 to obtain Y * ; Then the error is calculated through the loss function, and the back propagation algorithm is used to iterate the training until the search network error reaches the expected value; the calculation formula of the loss function is as follows:

[0095]

[0096] Among them, L SN is the error value used by the search network for back propagation, 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] Where a and c must satisfy a≤c.

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

[0100]

[0101] Where N d is the total number of constraints in the data set, which is 160,000, N s The number of parameters that meet the design requirements refers to the number of parameters in the search network output X and the verification network output Y whose error with the design requirements is 0.

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

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

[0104] Step S5: Construct a precision optimizer to improve the quality of the feasible solutions output by the search network.

[0105] like Figure 4 As shown, the precision optimizer performs the following steps to optimize the initial predicted structural parameters generated by the search network: To optimize:

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

[0107] (5-2) C luc Input into the search network trained in step S3 to obtain the current predicted structure parameters Will Input into the trained validation network in step S2 to obtain the current prediction performance parameters Calculate the current error by searching the network's loss function:

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

[0109]

[0110] Err c <Err f , then let Err f , final structural parameters Final performance parameters The values ​​are Err c , and

[0111] (5-4) Repeat steps (5-2) and (5-3) 10 times to make C luc Iterate 10 times;

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

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

[0114] MEMS sensor experimental data acquisition unit, which obtains MEMS sensor experimental data based on finite element simulation;

[0115] Verification network construction and training unit: Based on MEMS sensor experimental data, the verification network is constructed and trained to realize the mapping relationship from structural 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 the design requirements;

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

[0118] The search network optimization unit builds a precision optimizer to optimize the quality of the feasible solutions output by the search network.

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

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

Claims

1. A MEMS sensor intelligent design method for general design scenarios, characterized in that: include: Step S1: Acquire MEMS sensor experimental data based on finite element simulation; Step S2: Based on the MEMS sensor experimental data, construct and train a verification network to realize the mapping relationship from structural parameters to performance parameters; Step S3: Generate data based on the verification network, quantify different types of design requirements, and use the reverse generation strategy to construct a data set containing design requirements; Step S4: Based on the data set, construct and train a search network to achieve a mapping relationship from design requirements to feasible structures; Step S5: construct an accuracy optimizer to optimize the quality of feasible solutions output by the search network.

2. According to claim 1, a MEMS sensor intelligent design method for general design scenarios is characterized in that: The MEMS sensor experimental data in step S1 includes structural parameters X and performance parameters Y; wherein the structural parameters X include continuous variables describing the geometric structure of the sensor, with a dimension of m; and the performance parameters Y include continuous variables describing the performance of the sensor, with a dimension of n.

3. According to claim 1, a MEMS sensor intelligent design method for general design scenarios is characterized in that: 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 data set normalized by min-max in step S1 to establish a mapping relationship from the structural parameter X to the performance parameter Y, and continuously iterates the training through the back propagation algorithm, using the mean square error as the loss function of the network training.

4. The method for intelligent design of MEMS sensors for general design scenarios according to claim 3, characterized in that: The verification network is initialized using the Xavier method, uses the Learky ReLU function as the activation function, uses the Adam optimizer for back propagation, and uses ridge regression for regularization.

5. The method for intelligent design of MEMS sensors for general design scenarios according to claim 1, characterized in that: The quantification method of quantifying different types of design requirements in step S3 is: Among them, A and B represent the constants of the upper or lower limit of the design requirements, respectively, and y i is the constrained i-th 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: c r :y i ∈(-∞,+∞) In normalized expression, 0 and 1 are used to represent infinitesimal and infinity respectively.

6. The method for intelligent design of MEMS sensors for general design scenarios according to claim 1, characterized in that: In step S3, the reverse generation strategy is used to construct a data set containing design requirements, which specifically includes: Step S3-1, using the verification 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 requirement by the following formula; c xl =x-a1 c xu =x+(d1-a1) 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 With c xu is the lower and upper bounds of x, c yl With c yu is the lower and upper bounds 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 this group of data is obtained lu :(c x1l , c x1u , …, c xml , c xmu , c y1l , c y1u , …, c ynl , c ynu ); Step S3-3, repeat steps S3-1 and S3-2 to generate the data set (C lu , X, Y).

7. The method for intelligent design of MEMS sensors for general design scenarios according to claim 1, characterized in that: When training the search network in step S4, first set the constraint C lu Input into the search network to get the predicted structural parameter X * , and then X * Input to the verification network trained in step S2 to obtain Y * ; Then the error is calculated through the loss function, and the back propagation algorithm is used to iterate the training until the search network error reaches the expected value.

8. The method for intelligent design of MEMS sensors for general design scenarios according to claim 7, characterized in that: The loss function is: Among them, L SN is the error value of the search network 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 ; Function T is: Where a and c must satisfy a≤c, and a, b, c are variables of the T function.

9. The method for intelligent design of MEMS sensors for general design scenarios according to claim 8, characterized in that: Constructing a precision optimizer to optimize the quality of the feasible solutions output by the search network specifically includes: Step S5-1: The original constraint condition C lu0 Assign values ​​to the current constraint C respectively luc ; The final error Err f Set to 9999; Set the learning rate η; Step S5-2: C luc Input into the search network trained in step S4 to obtain the current predicted structure parameters Will Input into the validation network trained in step S2 to obtain the current prediction performance parameters Calculate the current error by searching the network's loss function: Step S5-3, update C by the following formula luc : Among them C l ' uc C luc The updated value of Err c <Err f , then let Err f , final structural parameters Final performance parameters The values ​​are Err c , and Step S5-4, repeat steps S5-2 and S5-3 to make C luc Iterate the setting I0 times; Step S5-5, output Err f , and 10. A MEMS sensor intelligent design system for general design scenarios that implements any of the methods described in claims 1-9, characterized in that: include: MEMS sensor experimental data acquisition unit, which obtains MEMS sensor experimental data based on finite element simulation; Verification network construction and training unit: Based on MEMS sensor experimental data, the verification network is constructed and trained to realize the mapping relationship from structural parameters to performance parameters; 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 the design requirements; The search network construction and training unit builds and trains the search network based on the data set to achieve the mapping relationship from design requirements to feasible structures; The search network optimization unit builds a precision optimizer to optimize the quality of the feasible solutions output by the search network.

Images