Photonic wire bonding parameter prediction method based on neural network and photonic wire bonding control system and method

Abstract

The present invention discloses a method for predicting photon wire bonding parameters based on a neural network, including determining environmental parameters and bonding control parameters that affect two-photon polymerization direct writing technology, and making a data set; normalizing the data set and dividing it into a training set and a test set; constructing a photon wire bonding parameter prediction model based on a fully connected feedforward neural network; optimizing the number of hidden layers and the number of nodes to converge the loss function, and obtaining a target photon wire bonding parameter prediction model after training; the output parameters of a femtosecond laser can be predicted by the model. Furthermore, the present invention also proposes an electronic device, a photon wire bonding control system and a control method. The present invention can effectively improve the coupling efficiency of the photon wire bonding process when interconnecting photonic devices, reduce the cost of manual regulation, and improve the precision of semiconductor processes.

Description

Technical Field

[0001] The present invention belongs to the technical field of semiconductor integrated circuit packaging, and more specifically, relates to a photon wire bonding technology based on neural network. Background Art

[0002] Silicon-based optoelectronic integrated chips have the advantages of high-speed transmission, high integration, and good compatibility. They are widely used in many fields such as optical communication systems, sensor technology, and biomedical imaging and detection. The manufacturing method of silicon-based optoelectronic integrated chips mainly uses photon wire bonding (PWB) to achieve coupling between chips or between chips and optical fibers. The concept of photon wire bonding technology was first proposed by a research team at Karlsruhe University of Technology in Germany. This method controls a high-energy pulsed laser beam to form a three-dimensional polymer waveguide in the photoresist, which initially realizes the coupling between the chip and the optical fiber.

[0003] With the continuous development of silicon-based optoelectronic chips, the preparation process has been gradually improved, and nanochip manufacturing has achieved micron-level cross-regions, which has more stringent requirements for coupling accuracy. Due to the femtosecond laser characteristics of femtosecond lasers and the particle characteristics of photon energy, photon wire bonding technology is extremely difficult. The current photon wire bonding technology at home and abroad usually uses markers as reference objects for alignment in the x and y directions while the chip is being made, which improves the alignment accuracy in the x and y directions, but it is still difficult to achieve precise control in the precise alignment in the z direction. Therefore, at this stage, when the photon wire bonding system is used, manual adjustment is mainly used to control the operation of the equipment. This method is not only time-consuming, but also costly. In addition, due to the lack of high-precision control in the z direction, it is difficult to guarantee the coupling efficiency.

[0004] In summary, in order to improve the intelligence level of photonic wire bonding of silicon-based optoelectronic integrated chips, a new type of photonic wire bonding control technology is urgently needed to achieve precise chip alignment and improve coupling efficiency. Summary of the invention

[0005] In response to the above technical problems, the present invention provides a method for predicting photon wire bonding parameters based on a neural network, introduces a fully connected feedforward neural network to establish a target photon wire bonding parameter prediction model, takes the expected z-direction height and coupling efficiency as input, and achieves precise control of femtosecond laser output parameters, thereby improving coupling efficiency. Furthermore, the present invention also provides an electronic device, a photon wire bonding control system, and a control method based on the above method. The present invention can effectively improve the coupling efficiency of the photon wire bonding process when interconnecting photonic devices, reduce the cost of manual regulation, and improve the precision of semiconductor processes.

[0006] The first aspect of the present invention provides a method for predicting photonic wire bonding parameters based on a neural network, the method comprising the following steps:

[0007] S1: Determine the environmental parameters and bonding control parameters that affect the two-photon polymerization direct writing technology. During the photon wire bonding test, one of the parameters is changed each time the test is performed, and the other parameters remain unchanged. When the measured actual coupling efficiency exceeds the threshold, the corresponding parameter data, the actual coupling efficiency and the actual z-direction height are recorded, and arranged in a preset order as a vector data until all parameters are traversed to form a data set consisting of multiple vector data; the environmental parameters include ambient temperature, ambient pressure, and relative humidity; the bonding control parameters include center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, femtosecond laser wavelength, laser pulse width, laser repetition frequency, expected coupling efficiency, and expected z-direction height;

[0008] S2: performing normalization processing on the vector data in the data set, and dividing the processed data set into a training set and a test set;

[0009] S3: An initial photon wire bonding parameter prediction model is constructed based on a fully connected feedforward neural network, wherein the fully connected feedforward neural network includes an input layer, a hidden layer, and an output layer connected in sequence, wherein the input layer includes 8 input units, which are respectively used to input ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, and expected z-direction height and coupling efficiency; the output layer includes 3 output units, which are respectively used to output predicted values ​​of femtosecond laser wavelength, laser pulse width, and laser repetition frequency;

[0010] S4: using the training set to train the photon wire bonding parameter prediction model, using the test set to test and verify the trained photon wire bonding parameter prediction model, using the control variable method to update and optimize the number of hidden layers and nodes in the model, when the loss function used to evaluate the error between the predicted value and the actual value of the model for the test set tends to converge, the model training is completed, and the desired target photon wire bonding parameter prediction model is obtained;

[0011] S5: using the target photon wire bonding parameter prediction model to predict the output parameters of the femtosecond laser in the photon wire bonding, the output parameters including the femtosecond laser wavelength, the laser pulse width, and the laser repetition frequency.

[0012] A second aspect of the present invention provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is called by the processor, the computer program executes the above-mentioned neural network-based photonic wire bonding parameter prediction method.

[0013] The third aspect of the present invention proposes a photon wire bonding control system, comprising a femtosecond laser, a galvanometer, an objective lens, an XYZ console, a sample stage and a computer. The computer is used to receive the acquired or set ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, expected z-direction height and expected coupling efficiency, and the target photon wire bonding parameter prediction model obtained by the above-mentioned neural network-based photon wire bonding parameter prediction method predicts the femtosecond laser output parameters, and adjust the position of the sample stage through the XYZ console electrically connected thereto; the femtosecond laser outputs a corresponding pulsed laser beam according to the femtosecond laser output parameters; the galvanometer is used to control the laser focus of the pulsed laser beam, and scan the photosensitive resin on the focal plane of the objective lens according to the three-dimensional model shape.

[0014] A fourth aspect of the present invention provides a photon wire bonding control method, based on the above-mentioned photon wire bonding control system, comprising the following steps:

[0015] Fixing a first object and a second object to be interconnected onto the same sample stage; the first object is a photonic chip, and the second object is an optical fiber or a photonic chip;

[0016] embedding the interconnection region of the first object and the second object into a photosensitive resin, and automatically detecting the actual position of the photonic wire bonding waveguide end face and the coupling structure in the photosensitive resin;

[0017] adjusting the three-dimensional model shape of a PWB waveguide used to connect the first object end face to the second object end face according to the waveguide end face position;

[0018] According to the desired z-direction height and the marks on the first object and / or the second object as reference objects, the position of the sample stage is adjusted by the XYZ control console to perform an alignment operation;

[0019] The computer receives the pre-acquired ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, and three-dimensional model shape, sets the expected z-direction height and expected coupling efficiency, and predicts the femtosecond laser output parameters through the target photon wire bonding parameter prediction model;

[0020] The femtosecond laser outputs a corresponding pulsed laser beam according to the predicted femtosecond laser output parameter;

[0021] The galvanometer controls the pulsed laser beam to focus and scan the photosensitive resin according to the three-dimensional model shape of the PWB waveguide to obtain the actual waveguide structure;

[0022] After writing is completed, the PWB waveguide is embedded in a cladding material with a low refractive index, thereby realizing the interconnection between the first object and the second object.

[0023] Compared with the prior art, the advantages of the present invention are:

[0024] (1) The present invention fully considers the environmental parameters and bonding control parameters that affect the two-photon polymerization direct writing technology, performs photon wire bonding tests by sequentially changing the above parameters, and thereby obtains a data set that can be used for training and testing a photon wire bonding parameter prediction model.

[0025] (2) The present invention introduces a fully connected feedforward neural network as the basis of the photon wire bonding parameter prediction model, and trains the model through the data set described in Article (1) to obtain a target photon wire bonding parameter prediction model. Based on the target photon wire bonding parameter prediction model, the output parameters (femtosecond laser wavelength, laser pulse width and laser repetition frequency) of the femtosecond laser under the current environment and bonding conditions can be predicted, so as to make the actual coupling efficiency infinitely close to the expected coupling efficiency by accurately determining the optimal output parameters of the femtosecond laser, thereby improving the coupling efficiency of the optical device connection.

[0026] (3) The present invention automatically controls the height in the z direction of the XYZ console through a computer, and there is no need to manually adjust the XYZ console for alignment operations. Not only can the precise control of the height in the z direction be achieved, but also, combined with the target photon wire bonding parameter prediction model, the optimal femtosecond laser output parameters matching the expected coupling efficiency can be automatically calculated and output at the height, thereby achieving high coupling efficiency for the connection of passive optical devices, ensuring the quality of interconnection of optical devices, and effectively improving the device yield. In addition, the present invention has stronger reliability and practicality, and is easier to operate. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of the process flow of the photon wire bonding control method disclosed in an embodiment of the present invention;

[0028] Figure 2 A schematic diagram of a photon wire bonding data acquisition process is disclosed for an embodiment of the present invention;

[0029] Figure 3 A schematic diagram of the deep neural network structure disclosed in an embodiment of the present invention;

[0030] Figure 4 A schematic diagram of a photon wire bonding model training process is disclosed for an embodiment of the present invention;

[0031] Figure 5 A schematic diagram of the photon wire bonding process disclosed in an embodiment of the present invention;

[0032] Figure 6 This is a schematic diagram of a z-direction control system for photonic wire bonding disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to further understand the present invention, the technical solution of the present invention is further explained below in conjunction with specific embodiments and drawings.

[0034] like Figure 1 As shown, a photon wire bonding (PWB) system disclosed in an embodiment of the present invention mainly includes a femtosecond laser, a galvanometer, an objective lens, an XYZ control console, a potentiometer scanner, an ND filter, a sample stage and a computer. The laser beam is emitted by the femtosecond laser, passes through the ND filter, the potentiometer scanner and the galvanometer in sequence, and is focused by the high-power objective lens into the photosensitive resin. Among them, the computer is mainly used to control the emission of the laser beam, perform optical power control to adjust the laser energy according to demand, and control the XYZ control console to adjust the position of the sample stage in the XYZ direction. The ND filter is mainly used to filter out the generated noise signal. The potentiometer scanner is used to provide the host computer with femtosecond laser wavelength information. The galvanometer is used to control the laser focus of the pulsed laser beam, and scan the photosensitive resin on the focal plane of the objective lens according to the three-dimensional model shape preset by the system. The XYZ control console is mainly used to move the sample stage left and right and lift it, and photon wire bonding can be achieved through linkage control with the galvanometer.

[0035] Based on the above system, the main steps of the photonic wire bonding are as follows:

[0036] Step 1: Use a standard pick-and-place mechanical device to fix the optical devices to be connected (such as optical fibers and photonic chips) on a common sample stage. The transmission ports of the optical fiber and the photonic chip need to be connected to each other through a photonic wire bonding system, and there is a certain height difference between the center distances of the two.

[0037] Step 2, embedding the interconnection area of ​​the optical fiber and the photonic chip into a photosensitive resin, and then using existing 3D machine vision technology to automatically detect the actual position of the photonic wire bonding waveguide end face and the coupling structure in the resin;

[0038] Step 3, according to the detected waveguide end face position, the three-dimensional model shape of the photonic wire bonding waveguide (referred to as "PWB waveguide") is adjusted, and the three-dimensional model shape of the PWB waveguide (referred to as "three-dimensional model shape") is imported into the optical printing platform by the mapping software, and the PWB waveguide is connected from the optical fiber end face to the photonic chip end face;

[0039] Step 4, setting the parameters of the femtosecond laser (including the femtosecond laser wavelength, laser repetition frequency and laser pulse width), so that the femtosecond laser outputs a corresponding pulse laser beam;

[0040] Step 5, using two-photon polymerization direct writing technology (also known as "two-photon 3D printing technology") to define a PWB waveguide structure in the shape of a three-dimensional model in the focus of a pulsed laser beam;

[0041] Step 6: After writing is completed, the unexposed resin material is removed, and the PWB waveguide structure is embedded in a low-refractive-index cladding material, and photonic wire bonding is achieved.

[0042] The embodiment of the present invention discloses a method for predicting photon wire bonding parameters based on a neural network based on a constructed photon wire bonding system, which mainly includes the following steps:

[0043] Step 1: Collect photon wire bonding data and create a data set. Figure 2 As shown, the specific method is as follows:

[0044] The parameters that usually affect the two-photon polymerization direct writing technology mainly include environmental parameters and bonding control parameters. Environmental parameters mainly include ambient temperature, ambient pressure, and relative humidity; bonding control parameters mainly include the relative height difference between the preset photon wire bonding waveguide and the center of the optical device to be connected (referred to as "center relative height difference"), low refractive index cladding material parameters, three-dimensional model shape, femtosecond laser wavelength, laser pulse width, laser repetition frequency, expected coupling efficiency, and expected z-direction height.

[0045] The above parameters are obtained as follows: use a temperature monitor to monitor the ambient temperature; use a pressure chip to measure the ambient pressure; use a high-precision portable relative humidity / temperature calibrator to measure the relative humidity; use 3D machine vision technology to measure the relative height difference between the center of the chip and the center of the optical fiber; use 3D machine vision technology to directly obtain the parameters of the low refractive index coating material; use a potentiometer scanner to obtain the femtosecond laser wavelength; the laser pulse width and laser repetition frequency are computer-generated and controlled parameters and can be directly recorded.

[0046] Based on the established photonic wire bonding system, the environmental parameters and bonding control parameter adjustments are controlled in turn, and two-photon polymerization direct writing technology is used for photonic wire bonding (PWB). The computer-controlled optical test platform is used to test the photonic wire bonding coupling efficiency and obtain the coupling efficiency parameters.

[0047] During the test, only one of the environmental parameters and bonding control parameters is changed each time, and the other parameters remain unchanged. The above process is repeated until all nine parameters, including ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, femtosecond laser wavelength, laser pulse width, and laser repetition frequency, are traversed.

[0048] Each test uses 3D machine vision technology to test the coupling effect and obtain the coupling efficiency and z-direction height value. Specifically, if the coupling efficiency exceeds the threshold (for example, 80%), the z-direction height value is measured using 3D machine vision technology, and other related parameters are measured and recorded at the same time, and the environmental parameters and bonding control parameters of this test, as well as the measured coupling efficiency parameters and z-direction height values ​​are recorded as a set of data; if the measured coupling efficiency does not exceed the threshold, the set of data is discarded.

[0049] Each record data obtained from multiple tests is arranged in the order of ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, femtosecond laser wavelength, laser pulse width, laser repetition frequency, coupling efficiency and z-direction height, and added to the data set as a vector data. The final data set can be used for subsequent model training and testing.

[0050] In order to accurately control the output height of the photon wire bonding in the z direction, it is necessary to accurately adjust the output parameters of the femtosecond laser (including the wavelength of the femtosecond laser, the laser repetition frequency and the laser pulse width). The present invention proposes a fully connected feedforward neural network to predict the real-time output parameters of the femtosecond laser and effectively improve the coupling efficiency of the photon wire bonding.

[0051] Step 2: Preprocess the data in the data set, and divide the data set into a training set and a test set after preprocessing. Before model training, read the data in the data set first, normalize the data of the 8 input units respectively, and obtain a standard vector database to avoid excessive deviation between the data. The specific method is: calculate the mean of the vector data corresponding to each input unit in multiple vector data respectively, that is, sum and divide all the vector data in each input unit to obtain the mean vector, and then subtract the corresponding mean vector from the vector data to obtain the normalized standard vector data that are all around 0, thereby forming a standard vector database.

[0052] The normalized standard vector data is used as sample data, and the obtained sample data is divided randomly: for example, 90% of the data is randomly taken as the training set (data_train) and the remaining data is used as the test set (data_test).

[0053] Step 3: Construct a photonic wire bonding parameter prediction model.

[0054] The embodiment of the present invention adopts a fully connected feedforward neural network (FCNN) to construct an initial photon wire bonding parameter prediction model. By training the model, the required target photon wire bonding parameter prediction model is obtained after the training is completed, which is used to predict the output parameters of the femtosecond laser during the photon wire bonding experimental operation, ensuring the precise control of the z-direction height during the photon wire bonding process, thereby improving the photon wire bonding coupling efficiency.

[0055] like Figure 3 As shown in the figure, the model based on the fully connected feedforward neural network has an input layer, an output layer, and a hidden layer connecting the input layer and the output layer. The input layer contains 8 input units, which correspond to the 8 input parameters of ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, expected coupling efficiency, and expected z-direction height parameters; the number of hidden layers is determined by the size of the data set, and each layer contains multiple neurons (nodes); the output layer contains 3 output units, which correspond to the 3 output parameters of the femtosecond laser, namely, the femtosecond laser wavelength, laser pulse width, and laser repetition frequency.

[0056] Step 4: Train the photon wire bonding parameter prediction model, and after the training is completed, obtain the desired target photon wire bonding parameter prediction model. Figure 4 The specific contents are as follows:

[0057] Use the training set to train the built photon wire bonding parameter prediction model, use the test set to test and verify the trained photon wire bonding parameter prediction model, and calculate the loss function value between the predicted value output by the model during the test and the corresponding actual value in the data set. , continuously update the number of hidden layers and the number of hidden layer nodes (number of neurons) in the optimization model, when the loss function value When it is less than a preset value (for example, 0.001), the training is completed and the desired target photonic wire bonding parameter prediction model is obtained.

[0058] Before training each updated model, all weights and biases in the model are set to random numbers as close to 0 as possible. During the training process, the input data entering the hidden layer is processed in two steps. First, the input data is linearly transformed to calculate the output value of the data entering the node; then the output value is processed by a nonlinear activation function to calculate the output activation value of the hidden node. The activation function used is:

[0059] ;

[0060] In the formula, x Represents the vector data input to each hidden layer node.

[0061] Among them, the processing method from the input layer to the first hidden layer of the fully connected feedforward neural network can be expressed as:

[0062] ;

[0063] ;

[0064] In the formula, represents the weight matrix between the input layer and the first layer, represents the bias vector between the input layer and the first layer, Represents the vector data of the first hidden layer after weight and bias, is the vector data in the first hidden layer after the activation function.

[0065] Among them, the processing method between the current hidden layer and the next hidden layer can be expressed as:

[0066] ;

[0067] ;

[0068] In the formula, M It represents the number of hidden layers of the fully connected feedforward neural network minus 1. represents the weight matrix between the mth layer and the m+1th layer, is the vector data in the mth hidden layer after the activation function, represents the bias vector between the mth layer and the m+1th layer, Represents the vector data in the input m+1 hidden layer after weight and bias, is the vector data in the m+1th hidden layer after the activation function.

[0069] The processing from the terminal hidden layer to the output layer of the fully connected feedforward neural network can be expressed as:

[0070] ;

[0071] In the formula, t represents the number of hidden layers in a fully connected feedforward neural network, represents the weight matrix between the terminal hidden layer and the output layer, represents the bias vector between the terminal hidden layer and the output layer, After the activation function, t The vector data in the hidden layer, Y It is the vector data in the input and output layers after weights and biases;

[0072] Among them, the processing method of the output layer can be expressed as:

[0073] ;

[0074] In the formula, y Represents the output vector of the output layer after the activation function. Y It is the vector data in the input and output layers after weights and biases.

[0075] Finally, the output vector y Each bit of is added with the last three values ​​of the mean vector in order to obtain the output parameter.

[0076] Use the control variable method to update the number of hidden layers and nodes in the optimization model. Specifically, first, under the same number of hidden layers, control the number of nodes from 50 to 500, with an increment of 50, and select MSW <0.001 The number of nodes in the model remains unchanged; then the number of hidden layers is controlled to change from 2 to 20, with an increment of 2. MSW <0.0001 determines the number of hidden layers with the best performance and determines the number of hidden layers and nodes in the deep neural network.

[0077] For the output parameters obtained from the training of the fully connected feedforward neural network model, a difference evaluation function (i.e., loss function) is set to calculate the error between the output parameters and the corresponding data in the data set, which is used to evaluate the model performance. The difference evaluation function value (loss function value) MSW The calculation formula is:

[0078] ;

[0079] In the formula, Indicates the number of vector data in the data set. is the actual value in the data set, are the predictions obtained using a fully connected feed-forward neural network.

[0080] For the initial model after traversing the data, the obtained MSW The value is large, the output value has errors, and the output parameters that are the same as the collected data cannot be obtained. Therefore, it is necessary to further use the gradient descent algorithm and the back propagation algorithm to update the network weights and biases to minimize the loss function.

[0081] The basic formula for gradient descent is:

[0082] ;

[0083] ;

[0084] In the formula, trepresents the number of hidden layers in the fully connected feedforward neural network, m is the total number of hidden layers, η is the learning rate. Here, the learning rate in the neural network is set to 0.00001. L is the loss function MSEv, and They are the partial derivatives of the loss function with respect to weights and biases, respectively.

[0085] The back propagation algorithm starts from the output layer and calculates the gradient layer by layer until the input layer. For each layer, the calculation of the gradient depends on the gradient of the next layer.

[0086] First, calculate the gradient of the output layer:

[0087] ;

[0088] Where y is the output of the output layer activation function, Y is the weighted input of the output layer;

[0089] For each hidden layer, the calculation of the gradient depends on the gradient of the next layer.

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, and are the partial derivatives of the loss function with respect to weight and bias, respectively. is the partial derivative of the loss function with respect to the m-layer weighted input vector, is the transpose of the partial derivative of the loss function with respect to the m-layer weighted input vector.

[0094] Calculate the loss function value for each configuration corresponding to the number of hidden layers and nodes ( MSW ), and finally MSW When the value approaches 0 infinitely, it meets the threshold setting standard and selects MSW The configuration when it is less than 0.0001 is used as the optimal number of hidden layers and the number of nodes in each hidden layer; the number of hidden layers and the number of nodes in each hidden layer are recorded until the traversal is completed. If a fully connected feedforward neural network model that meets the standards is not obtained, continue to adjust the range of the number of hidden layer nodes and the incremental step size until a fully connected feedforward neural network model that meets the standards is obtained. In this embodiment, the optimal number of hidden layers is set to 20 layers, and the number of nodes in each hidden layer is 200. The number of training iterations in each neural network is 5000 times, which also means that the entire training set is used to train the model 5000 times.

[0095] Step S5: using the target photon wire bonding parameter prediction model to predict the output parameters of the femtosecond laser in the photon wire bonding, the output parameters including the femtosecond laser wavelength, the laser pulse width, and the laser repetition frequency.

[0096] Based on the obtained target photon wire bonding parameter prediction model, the embodiment of the present invention further discloses a photon wire bonding method, which mainly includes the following steps:

[0097] Model input parameter acquisition: The computer receives parameters such as ambient pressure value, ambient temperature value, relative humidity value, three-dimensional model shape, low refractive index cladding material parameters, and center relative height difference, and the user inputs the expected coupling efficiency and expected z-direction height into the computer. Among the parameters, the ambient pressure value, ambient temperature value, relative humidity value, and the center relative height difference between the preset photon wire bonding waveguide and the optical device to be connected are real-time parameters currently measured and acquired, and the specific acquisition method has been involved in the above embodiments.

[0098] Photonic wire bonding parameter prediction: The ambient pressure value, ambient temperature value, relative humidity value, relative height difference between the preset photonic wire bonding waveguide and the center of the optical device to be connected, low refractive index cladding material parameters, three-dimensional model shape, input expected coupling efficiency and expected z-direction height parameters are input into the trained target photonic wire bonding parameter prediction model for parameter prediction to obtain the predicted values ​​of femtosecond laser wavelength, laser pulse width and laser repetition frequency.

[0099] Photon wire bonding: The computer controls the femtosecond laser wavelength, laser pulse width and laser repetition frequency of the high-energy pulsed laser beam. The laser pulse beam is emitted by the femtosecond laser, filtered out the noise signal through the ND filter, and transmitted to the galvanometer after the potentiometer scanner. The galvanometer controls the laser focus to scan on the focal plane of the objective lens, and then focuses it into the photoresist through the high-power objective lens. Photon polymerization occurs at a specific position of the photoresist to form a three-dimensional polymer waveguide. The specific process can refer to the steps of photon wire bonding in the previous article.

[0100] It is worth noting that during this process, the computer can control the XYZ console in real time to move the sample stage left and right and lift it up and down according to the desired z-direction height and the mark on the first object used as a reference object, so as to achieve alignment in the x and y directions and the desired z-direction height. Thus, photonic wire bonding is achieved through the linkage control of the galvanometer and the XYZ console, so that the optical fiber and the PWB waveguide as well as the PWB waveguide and the end face of the optical chip can be accurately connected, thereby improving the coupling efficiency between the two and realizing efficient interconnection technology between the optical fiber and the photonic chip, thereby realizing precise control of the z-direction height of the photonic wire bonding polymer and the waveguide coupling efficiency.

[0101] Furthermore, an embodiment of the present invention also discloses an electronic device, which mainly includes a processor and a memory, and the processor is mainly used to call the computer program stored in the memory and execute the neural network-based photon wire bonding parameter prediction method provided in the above embodiment. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the electronic device and the program code in the computer-readable storage medium described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0102] Finally, it should be noted that although the embodiments of the present invention are described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments and application fields, and the above-mentioned specific embodiments are only illustrative and instructive, rather than restrictive. Under the guidance of this specification, a person of ordinary skill in the art can also make many forms without departing from the scope of protection of the claims of the present invention, all of which belong to the scope of protection of the present invention.

Claims

1. A method for predicting photonic wire bonding parameters based on neural network, characterized in that: include: S1: Determine the environmental parameters and bonding control parameters that affect the two-photon polymerization direct writing technology. During the photon wire bonding test, one of the parameters is changed each time the test is performed, and the other parameters remain unchanged. When the measured actual coupling efficiency exceeds the threshold, the corresponding parameter data, the actual coupling efficiency and the actual z-direction height are recorded, and arranged in a preset order as a vector data until all parameters are traversed to form a data set consisting of multiple vector data; the environmental parameters include ambient temperature, ambient pressure, and relative humidity; the bonding control parameters include center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, femtosecond laser wavelength, laser pulse width, laser repetition frequency, expected coupling efficiency, and expected z-direction height; S2: performing normalization processing on the vector data in the data set, and dividing the processed data set into a training set and a test set; S3: An initial photon wire bonding parameter prediction model is constructed based on a fully connected feedforward neural network, wherein the fully connected feedforward neural network includes an input layer, a hidden layer, and an output layer connected in sequence, wherein the input layer includes 8 input units, which are respectively used to input ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, and expected z-direction height and coupling efficiency; the output layer includes 3 output units, which are respectively used to output predicted values ​​of femtosecond laser wavelength, laser pulse width, and laser repetition frequency; S4: using the training set to train the photon wire bonding parameter prediction model, using the test set to test and verify the trained photon wire bonding parameter prediction model, using the control variable method to update and optimize the number of hidden layers and nodes in the model, when the loss function used to evaluate the error between the predicted value and the actual value of the model for the test set tends to converge, the model training is completed, and the desired target photon wire bonding parameter prediction model is obtained; S5: using the target photon wire bonding parameter prediction model to predict the output parameters of the femtosecond laser in the photon wire bonding, the output parameters including the femtosecond laser wavelength, the laser pulse width, and the laser repetition frequency.

2. A method for predicting photonic wire bonding parameters based on a neural network as claimed in claim 1, characterized in that: During the model training process, the processing method of input data entering the hidden layer is divided into two steps: first, a linear transformation is performed on the input data to calculate the output value of the data entering the node; then a nonlinear activation function is performed on the output value to calculate the output activation value of the hidden node.

3. A method for predicting photonic wire bonding parameters based on a neural network as claimed in claim 2, characterized in that: The nonlinear activation function is: ； In the formula, x Represents the vector data input to each hidden layer node.

4. The method for predicting photonic wire bonding parameters based on a neural network according to claim 1, characterized in that: The control variable method is used to update and optimize the number of hidden layers and the number of nodes in the model, specifically including: first, under the same number of hidden layers, controlling the number of nodes to change from 50 to 500, with an incremental step of 50, and selecting the number of nodes of the model to remain unchanged when the loss function value is less than 0.001; then controlling the number of hidden layers to change from 2 to 20, with an incremental step of 2, and selecting the number of hidden layers of the model when the loss function value is <0.0001, thereby obtaining the optimal number of hidden layers and the optimal number of nodes of the model.

5. The method for predicting photonic wire bonding parameters based on a neural network as claimed in claim 4, characterized in that: The optimal number of hidden layers is 20; the optimal number of nodes is 200.

6. The method for predicting photonic wire bonding parameters based on a neural network according to claim 1, characterized in that: The expression of the loss function is as follows: ； In the formula, MSE is the loss function value, Indicates the number of vector data in the data set. is the actual value in the data set, are the predictions obtained using a fully connected feed-forward neural network.

7. A method for predicting photonic wire bonding parameters based on a neural network according to any one of claims 1 to 6, characterized in that: Use the gradient descent algorithm and back-propagation algorithm to update the network weights and biases to minimize the loss function.

8. An electronic device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is called by the processor, the neural network-based photonic wire bonding parameter prediction method according to any one of claims 1 to 7 is executed.

9. A photon wire bonding control system, comprising a femtosecond laser, a galvanometer, an objective lens, an XYZ console, a sample stage and a computer, characterized in that: The computer is used to receive the acquired or set ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, three-dimensional model shape, expected z-direction height and expected coupling efficiency, and the target photon wire bonding parameter prediction model obtained by the neural network-based photon wire bonding parameter prediction method according to any one of claims 1 to 7 to predict the output parameters of the femtosecond laser, and to adjust the position of the sample stage through an XYZ console electrically connected thereto; the femtosecond laser outputs a corresponding pulsed laser beam according to the output parameters of the femtosecond laser; The galvanometer is used to control the laser focus of the pulse laser beam and scan the photosensitive resin on the focal plane of the objective lens according to the shape of the three-dimensional model.

10. A photon wire bonding control method, characterized in that: The system according to claim 9 comprises the following steps: Fixing a first object and a second object to be interconnected onto the same sample stage; the first object is a photonic chip, and the second object is an optical fiber or a photonic chip; embedding the interconnection region of the first object and the second object into a photosensitive resin, and automatically detecting the actual position of the photonic wire bonding waveguide end face and the coupling structure in the photosensitive resin; adjusting the three-dimensional model shape of a PWB waveguide used to connect the first object end face to the second object end face according to the waveguide end face position; According to the desired z-direction height and the marks on the first object and / or the second object as reference objects, the position of the sample stage is adjusted by the XYZ control console to perform an alignment operation; The computer receives the pre-acquired ambient temperature, ambient pressure, relative humidity, center relative height difference, low refractive index cladding material parameters, and three-dimensional model shape, sets the expected z-direction height and expected coupling efficiency, and predicts the femtosecond laser output parameters through the target photon wire bonding parameter prediction model; The femtosecond laser outputs a corresponding pulsed laser beam according to the predicted femtosecond laser output parameter; The galvanometer controls the pulsed laser beam to focus and scan the photosensitive resin according to the three-dimensional model shape of the PWB waveguide to obtain the actual waveguide structure; After writing is completed, the PWB waveguide is embedded in a cladding material with a low refractive index, thereby realizing the interconnection between the first object and the second object.

Images