Artificial neural network-based casting layer stress prediction method and system, and storage medium

Abstract

The application discloses a kind of based on artificial neural network's casting layer stress prediction method, system and storage medium, comprising: with silicon chip as electroforming base, one side of electroforming base is polished, and the other side is provided with conductive layer;Configuration electrocasting groove, collimating light source and wavefront sensor, adjust the relative position of collimating light source, wavefront sensor and electrocasting groove, make the light beam of collimating light source be reflected into wavefront sensor through the polished surface of silicon chip;With direct current or pulse current, electrocasting is carried out in electrocasting groove;Through data acquisition module, electrocasting process data and laser wavefront curvature radius are collected;According to the stress value calculated according to laser wavefront curvature radius, and stress value data and electrocasting process data are converted into neural network adaptation format;Select network type, activation function and hyperparameter, train artificial neural network model to fit the nonlinear relationship between electrocasting parameter and stress value;Stress prediction: input electrocasting parameter to artificial neural network model, output predicted casting layer stress value.

Description

Technical Field

[0001] The present invention relates to the technical field of quality inspection, and in particular to a casting layer stress prediction method, system and storage medium based on an artificial neural network. Background Art

[0002] In polymer micron-scale manufacturing, microinjection molding is often used for the mass production of functional polymer microdevices such as microfluidics, micro-optics, and anti-stick surfaces due to its high throughput, low manufacturing cost, and high replication accuracy. In injection molding of polymer microdevices, precision metal microstructured core fabrication is one of the key technologies for achieving mass production of polymer microdevices. Currently, the use of micro-electroforming, a method with high manufacturing precision (atomic-level deposition) and low cost, to prepare metal microstructured cores is one of the main approaches to core micromanufacturing internationally.

[0003] Artificial neural networks (ANNs) are mathematical models that mimic the behavioral characteristics of animal neural networks. They are used for distributed parallel information processing and can be used to analyze complex nonlinear problems. Essentially, an ANN is a computational model composed of a large number of interconnected nodes (neurons). ANNs have demonstrated excellent intelligent properties in fields such as pattern recognition, intelligent robotics, automation, biology, medicine, and economics, and have successfully solved many practical problems.

[0004] During the electroforming process, microstructured cores are prone to generating large macroscopic stresses within the casting layer, leading to defects such as warping, blistering, and cracking. Therefore, the key to achieving mass production and industrial application of polymer microdevices is to cost-effectively measure the nonlinear relationship between casting layer stress and basic electroforming process parameters online, thereby obtaining low-stress cores and reducing the occurrence of related defects. Currently, existing technologies in this field primarily focus on the development of online casting layer stress detection methods and devices. However, no existing technology has yet combined artificial neural networks (ANNs) to transform the casting layer stress detection process into a digital model for online prediction of internal stress. Summary of the Invention

[0005] The present invention aims to address at least one of the technical problems existing in the prior art. To this end, the present invention proposes a method for predicting cast layer stress based on an artificial neural network. This method is capable of obtaining electroformed mold cores in a low-stress state. This method improves the manufacturing quality of electroformed mold cores while effectively leveraging the advantages of the nonlinear fitting capabilities of artificial neural networks, reducing processing costs and improving processing efficiency.

[0006] According to an embodiment of the first aspect of the present invention, a casting layer stress prediction method based on an artificial neural network includes:

[0007] S100, electroforming substrate pretreatment: using a silicon wafer as an electroforming substrate, polishing one side of the electroforming substrate, and providing a conductive layer on the other side;

[0008] S200, adjusting the in-situ stress detection platform: configuring an electroforming tank and an optical measurement module, wherein the optical measurement module includes a collimated light source and a wavefront sensor, and adjusting the relative positions of the collimated light source, the wavefront sensor, and the electroforming tank so that the light beam of the collimated light source is reflected by the polished surface of the silicon wafer and enters the wavefront sensor;

[0009] S300, electroforming experiment: electroforming is performed in the electroforming tank using a direct current or a pulse current;

[0010] S400, data acquisition: including a data acquisition module, through which electroforming process data and laser wavefront curvature radius are collected;

[0011] S500, data processing: calculating a stress value according to the laser wavefront curvature radius, and converting the stress value data and the electroforming process data into a neural network adaptation format;

[0012] S600, establishing an artificial neural network model: selecting a network type, an activation function, and hyperparameters, and training the artificial neural network model to fit the nonlinear relationship between electroforming parameters and stress values;

[0013] S700, stress prediction: inputting electroforming parameters into the trained artificial neural network model, and outputting predicted casting layer stress values.

[0014] According to some embodiments of the present invention, in step S500, calculating the stress value includes the following steps:

[0015] S501, calculating the curvature radius of the sample;

[0016] S502, calculating an average stress value according to the curvature radius of the sample, wherein the calculation of the average stress value includes a linear scenario and a nonlinear scenario;

[0017] The formula for the radius of curvature of the sample is:

[0018]

[0019] in: is the radius of curvature of the sample, is the radius of curvature of the laser wavefront, is the distance between the sample and the sensor, is the refractive index of the electrolyte;

[0020] In a linear scenario, the theoretical model formula of the average stress value is:

[0021]

[0022] wherein: is the average stress value, is the radius of curvature before electrodeposition of the substrate, is the radius of curvature after electrodeposition of the substrate;

[0023] , , , ;

[0024] , , , are the biaxial elastic modulus, the elastic modulus, the Poisson's ratio and the thickness of the substrate, respectively;

[0025] , , , are the biaxial elastic modulus, the elastic modulus, the Poisson's ratio and the thickness of the deposited layer, respectively;

[0026] In the nonlinear scenario, the theoretical model formula of the average stress is:

[0027]

[0028] wherein, is the average stress value, and are the substrate deformation curvatures in the x-axis and y-axis directions, respectively.

[0029] According to some embodiments of the present application, in step S100, the step of disposing the conductive layer on the electroforming substrate comprises the following steps:

[0030] S101. Plating a layer of metal titanium or metal chromium as a bonding layer on the electroforming substrate;

[0031] S102. Plating a layer of metal platinum as a conductive layer on the bonding layer.

[0032] According to some embodiments of the present application, in step S101, the sputtering conditions are a vacuum degree of 5x10 - ² mbar, a current of 100-150 mA, and a time of 100-200 s;

[0033] In step S102, the sputtering conditions are a vacuum degree of 5x10 - ² mbar, a current of 35 mA, and a time of 100-350 s.

[0034] According to some embodiments of the present invention, in step S100, providing the conductive layer on the electroformed substrate includes the following steps: pasting a layer of conductive tape on the electroformed substrate.

[0035] According to some embodiments of the present invention, in step S300:

[0036] The current density of the direct current is 0.4-2A / dm²;

[0037] The frequency of the pulse current is 0.1-1200 Hz, the average current density is 0.2-2 A / dm², and the duty cycle is 0.1-0.9.

[0038] According to some embodiments of the present invention, in step S600: establishing the artificial neural network model includes the following steps:

[0039] S601, selecting a deep neural network according to the type of the collected data;

[0040] S602, select a suitable activation function, the number of neural network layers and the number of neurons in each layer;

[0041] S603: Select a loss function for the regression task.

[0042] According to some embodiments of the present invention, in step S700: the generated process parameter step size is 0.01, and the stress value is output after inputting the model to optimize the electroforming parameter selection.

[0043] According to an embodiment of the second aspect of the present invention, a casting layer stress prediction system includes a memory and a processor, wherein the memory stores a determination program of a casting layer stress prediction method based on an artificial neural network, and the processor runs the determination program of the casting layer stress prediction method based on an artificial neural network, so that the casting layer stress prediction system executes the casting layer stress prediction method based on an artificial neural network.

[0044] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium includes: a determination program of a casting layer stress prediction method based on an artificial neural network is stored on the computer-readable storage medium, and when the determination program of the casting layer stress prediction method based on an artificial neural network is executed by a processor, the casting layer stress prediction method based on an artificial neural network is implemented.

[0045] The casting layer stress prediction method, system and computer-readable storage medium based on an artificial neural network according to the embodiments of the present invention have at least the following beneficial effects:

[0046] 1. The cost of detecting casting layer stress is reduced: the traditional casting layer stress detection method adopts offline mode and requires destructive sampling, which cannot realize real-time monitoring of the internal stress of the casting layer, which undoubtedly reduces the fault tolerance and increases the experimental and production manufacturing cost. The method can significantly reduce the production manufacturing cost on the basis of ensuring the acquisition of low stress state casting layer by in-situ optical measurement and model prediction, using ANN to map the relationship between casting layer stress and electroforming basic process parameters.

[0047] 2. Real-time feedback of parameter setting in the electroforming core process is provided: suitable electroforming process parameters can obtain a good casting layer with good forming quality. Through learning experimental data, the network model can make feedback on a new set of electroforming basic process parameters, which is beneficial to the selection of parameters and reduces the workload of process engineers.

[0048] 3. The universality of the method is improved: the experimental data of different process conditions and different substrate materials are continuously learned, which expands the application range of a single model. Combined with the continuous development of deep learning, a network with stronger learning expression ability can be used to establish a casting layer stress prediction model.

[0049] Additional aspects and advantages of the application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0050] The application will be further described below in conjunction with the drawings and examples, wherein:

[0051] Figure 1 is a network model diagram constructed for direct current electroforming according to an embodiment of the present application;

[0052] Figure 2 is a flowchart of a casting layer stress prediction method according to an embodiment of the present application;

[0053] Figure 3 is a structure diagram of an in-situ stress detection platform according to an embodiment of the present application;

[0054] Figure 4 is a scatter plot showing the influence of different hidden layer numbers and hidden layer node numbers on the determination coefficient in the casting layer stress prediction method according to an embodiment of the present application;

[0055] Figure 5 is a loss curve diagram of the training of the network model in the casting layer stress prediction method according to an embodiment of the present application.

[0056] LIST OF REFERENCE NUMBERS

[0057] Collimated light source 100; wavefront sensor 200; computer 300; electroforming tank 400; cathode 500; anode 600; power supply 700. DETAILED DESCRIPTION

[0058] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0059] In the description of the present invention, it should be understood that descriptions involving orientation, such as the orientation or positional relationship indicated by up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0060] In the description of the present invention, "a plurality" refers to more than two. The use of "first" or "second" is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of the indicated technical features, or implicitly indicating the order of the indicated technical features.

[0061] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0062] refer to Figure 1 As shown, the circles represent the neuron nodes of the neural network, and the numbers 2, 14, 14, and 3 represent the number of nodes in the input layer, hidden layer, hidden layer, and output layer, respectively. The number of nodes is obtained based on experience and experimental verification. Under this structure, the obtained effect is better. The present invention adopts a feedforward neural network multi-layer perceptron (MLP) model, and uses the backpropagation algorithm to continuously update the weights and biases of the hidden layer so that the loss function obtains a minimum value within a certain error range. For DC electroforming, the input of the MLP model is the casting layer thickness and current density; for pulse electroforming, the input is the casting layer thickness, average current density, frequency, and duty cycle. The output of both models is the casting layer stress, which includes the average stress values ​​in the x-axis and y-axis directions calculated by Freund's formula and the average stress value calculated by the nonlinear formula. The activation function selects the ReLU function to speed up the convergence speed and reduce the computational cost. The number of hidden layers and the number of nodes are determined by automatic machine learning. The specific construction method of the network structure is as follows:

[0063] First, build the input and output layers of the network according to the number of inputs and outputs. Select 1, 2, 4, and 8 hidden layers in sequence. The number of nodes in each layer is selected starting from 2, with a step size of 2 and an end point of 16. Since this is a regression task, the loss function can be a common function of regression tasks, such as the mean square error (MSE), and the evaluation index can be the coefficient of determination (R 2 ) or MSE.

[0064] Traditional methods for detecting stress in metal mold cores make it difficult to measure stress in the cast layer online, which reduces production efficiency, yield, and quality. The quality of metal mold cores is crucial for ensuring the quality and performance of polymer microdevices. Artificial neural networks are a powerful tool for capturing complex nonlinear relationships. Combined with high-quality data collected during the electroforming process, they can train highly accurate predictive models, thus contributing to the production of high-quality electroformed mold cores. The neural network structure and dataset quality have a significant impact on the model's predictive performance.

[0065] The present invention proposes a neural network stress prediction method for producing low-stress metal mold cores. The stress detection mechanism of an in-situ electroforming layer stress detection platform is incorporated into a neural network model, providing parameter guidance for the manufacture of low-stress mold cores. This method effectively leverages the advantages of the neural network stress prediction method while reducing mold core warpage and thickness non-uniformity, shortening the processing cycle and improving processing efficiency.

[0066] Reference Figure 2 and Figure 3 As shown, the present invention discloses a casting layer stress prediction method based on artificial neural network, comprising:

[0067] S100, electroforming substrate pretreatment: using a silicon wafer as an electroforming substrate, polishing one side of the electroforming substrate and providing a conductive layer on the other side.

[0068] Specifically, the electroformed substrate is a silicon wafer with a length of 80 mm and a width of 15 mm. A chromium (Cr) bonding layer is sputtered on the non-polished side (sputtering conditions: vacuum degree 5×10 - ² mbar, current 130 mA, time 170 s) and platinum (Pt) conductive layer (sputtering conditions: vacuum 5×10 - 2 mbar, current 35 mA, time 330 s). The substrate was ultrasonically cleaned (ethanol solution for 15 minutes, deionized water for 15 minutes) and dried. The titanium layer strengthens the bond between the substrate and the conductive layer, while the platinum layer ensures uniform current distribution during electroforming.

[0069] S200, adjusting the in-situ stress detection platform: configuring an electroforming tank and an optical measurement module, the optical measurement module including a collimated light source and a wavefront sensor, and adjusting the relative positions of the collimated light source, the wavefront sensor, and the electroforming tank so that the light beam of the collimated light source is reflected by the polished surface of the silicon wafer and enters the wavefront sensor;

[0070] Specifically, refer to Figure 3 As shown, the pre-treated silicon wafer, wire and cathode fixture are bonded with conductive tape so that the wire is between the silicon wafer and the fixture. The anode nickel plate is suspended on the anode fixture with titanium wire. The cathode and anode fixtures are placed in the electroforming tank so that the deposition side of the silicon wafer is facing the anode, the distance between the anode and the cathode is 20-30mm, and the polished side of the silicon wafer is facing the boundary of the electroforming tank. Ensure that the distance between the cathode and the wall is small enough to reduce the measurement error caused by the electroforming liquid. Connect the power supply and the anode and cathode. The electroforming liquid in the electroforming tank is mainly nickel sulfamate; adjust the relative positions of the electroforming tank, laser emitter and wavefront sensor, rotate the laser emitter lens to adjust the laser spot size, the laser spot should be slightly smaller than the width of the silicon wafer, and is best located in the center of the silicon wafer immersed in the electroforming liquid. After the laser is reflected by the polished side of the silicon wafer, it needs to enter the incident lens of the wavefront sensor. After adjusting the relative positions of the above three, measure the distance L between the sample and the sensor and record it.

[0071] Open the GEV Player software on the detection platform, configure the sensor's IP address, ensure data transmission between the sensor and the computer, open the SID4 software, turn on the camera, adjust the sensor's exposure to around 3000, fine-tune the relative positions of the electroforming tank, laser emitter, and wavefront sensor, so that the camera's light spot is located in the center of the field of view. Define a circular analysis area with a radius of 1.8 mm in the center of the light spot, block the laser emitter, and save the current field of view background to eliminate the influence of external light sources. Pre-collect three substrate images to eliminate the influence of initial silicon wafer deformation on the results. Set parameters to officially start data acquisition.

[0072] Install the cathode fixture (holding the silicon wafer, with the polished side facing the side of the electroforming tank) and the anode fixture (suspending the nickel plate), adjust the distance between the cathode and anode to 20 mm, and prepare the nickel sulfamate electroforming solution.

[0073] A collimated light source emits a parallel laser beam with a diameter of 14 mm, which is irradiated onto the polished side of the substrate. The reflected beam passes through the electroforming liquid and enters the wavefront sensor, specifically a quadri-wave lateral shearing interferometry sensor (QWLSIS).

[0074] Adjust the relative positions of the light source, sensor and electroforming tank to ensure that the reflected light spot is centered.

[0075] The STM32 microcontroller is connected to the ammeter, voltmeter, pH meter, and temperature sensor through the Modbus protocol to collect process parameters in real time.

[0076] The wavefront sensor data is used to record the curvature radius through the SID4 software.

[0077] S300, electroforming experiment: Use direct current or pulse current to perform electroforming in the electroforming tank.

[0078] Specifically, a pulse power supply is connected to the anode and cathode, and direct current and pulse current are used for electroforming respectively. Pulse current is essentially an on-off current, and the electroforming time is 30 minutes. The anode-cathode spacing is selected within a specific range, the electroforming temperature is between 25-65°C, and the pH of the electroforming liquid is between 3.5-4.5.

[0079] S400, data acquisition: including a data acquisition module, through which electroforming process data and laser wavefront curvature radius are collected;

[0080] Specifically, the basic process parameters of electroforming are collected through a hardware data acquisition module with a single-chip microcomputer as the control core. The voltmeter and ammeter are connected to the power supply, and the pH meter and temperature sensor are placed in the electroforming liquid. When electroforming begins, the EV screen recording software is opened to record the laser wavefront curvature radius collected by the SID4 software.

[0081] S500, data processing: calculating the stress value according to the laser wavefront curvature radius, and converting the stress value data and electroforming process data into a neural network adaptation format;

[0082] Specifically, the stress values ​​in the corresponding directions are calculated based on the laser wavefront curvature radius in the horizontal and vertical directions, and the stress values ​​corresponding to the number of network outputs and the remaining process parameter data are made into a data sample for model training, verification and testing.

[0083] S600, establishing an artificial neural network model: selecting a network type, activation function, and hyperparameters, and training the artificial neural network model to fit the nonlinear relationship between electroforming parameters and stress values;

[0084] Specifically, we use the PyChram integrated development platform to build an MLP model based on the Pytorch framework, using MSE as the loss function and ReLU as the activation function. 2 As an evaluation indicator; train and verify the model, and select a set of hyperparameters that meet the verification loss requirements.

[0085] S700, stress prediction: input electroforming parameters into the trained artificial neural network model, and output the predicted casting layer stress value.

[0086] Specifically, Python is used to generate a set of basic electroforming process parameters to determine the initial value, step size, and final value; by inputting the basic electroforming process parameters into the network, the corresponding stress output can be obtained, thereby achieving a specific accuracy of casting layer stress prediction.

[0087] Traditional offline stress testing requires destructive sampling. This method, through in-situ optical measurement and model prediction, can significantly reduce costs. Traditional methods require downtime for sampling, each of which is time-consuming. This method, however, enables real-time online monitoring, with data collection and prediction completed simultaneously, improving efficiency. The model can quickly recommend low-stress parameter combinations, reducing trial-and-error experiments and shortening debugging cycles. By adjusting the training data, the model can be adapted to different substrate materials (such as copper and nickel alloys) and electroforming liquid formulations, expanding its application scenarios. The neural network model has low prediction error, and the optical measurement curvature resolution is high, meeting the requirements of micron-level mold core manufacturing.

[0088] In some embodiments of the present invention, processing the collected data includes the following steps:

[0089] Import the collected data into an Excel file and use Python custom functions to calculate, trim and read the data;

[0090] Based on the Pytorch deep learning framework, a custom class inherited from the Dataset class is used to create a dataset instance, implement dataset partitioning, data type conversion, and special model validation methods;

[0091] Load the dataset through Dataloader and define the hyperparameters of the dataset.

[0092] In some embodiments of the present invention, in step S500, the calculation of the stress value includes the following steps:

[0093] S501, calculating the curvature radius of the sample;

[0094] S502. Calculate an average stress value according to the curvature radius of the sample, where the average stress value includes a linear scenario and a nonlinear scenario;

[0095] The formula for the radius of curvature of the sample is:

[0096]

[0097] in: is the radius of curvature of the sample, is the radius of curvature of the laser wavefront, is the distance between the sample and the sensor, is the refractive index of the electrolyte;

[0098] “+” is used when the casting layer is under compressive stress, that is, to measure the curvature of the concave mirror; “-” is used when the casting layer is under tensile stress, that is, to measure the curvature of the convex mirror.

[0099] In a linear scenario, the theoretical model formula for the average stress value is:

[0100]

[0101] in: is the average stress value, is the curvature radius of the substrate before electrodeposition, is the radius of curvature of the substrate after electrodeposition;

[0102] , , , ;

[0103] , , , are the biaxial elastic modulus, elastic modulus, Poisson's ratio, and thickness of the substrate, respectively;

[0104] , , , are the biaxial elastic modulus, elastic modulus, Poisson’s ratio and thickness of the deposited layer, respectively;

[0105] In nonlinear scenarios, the theoretical model formula for mean stress is:

[0106]

[0107] in, is the average stress value, and are the base deformation curvatures in the x-axis and y-axis directions, respectively.

[0108] In this example, the collimated light source emits a laser with a wavelength of 532 nm and a refractive index of 1.48. The distance between the sensor and the substrate is 23 mm. The current density is 1 A / dm², and the data read from the software is: r x :-6.675m -1 , r y :-9.860m -1 This value is derived from experiments and formulas:

[0109]

[0110] in: ris the radius of curvature of the laser wavefront, L is the distance between the sample and the sensor, R 0 is the curvature radius of the substrate before electrodeposition, R is the radius of curvature of the substrate after electrodeposition.

[0111] Silicon wafer elastic modulus: 170 GPa, Poisson's ratio: 0.25, thickness: 300 μm;

[0112] Deposited nickel parameters: elastic modulus: 210 GPa, Poisson's ratio: 0.31, thickness: 6.15 μm (thickness obtained by calculation).

[0113] Curvature radius after electroforming: R x :-19.826 m -1 , R y :-29.254 m -1 (R0 has been excluded in the software, and the measured data has taken into account the influence of initial deformation).

[0114] The reason for the negative sign is that the direction of deformation of the sample plane is opposite to the polished side (because the sensor measures this side). Therefore, a negative sign is added before the formula.

[0115] The Freund formula calculates the stress in the x-direction and the y-direction respectively, that is, substitute R x , R y The calculated value (in this embodiment, after adding a negative sign, the radius of curvature substituted into the formula is a positive value) should be a positive value, that is, the casting layer is tensile stress.

[0116] Substituting the data, we get σ x =30.316MPa, σ y =20.546MPa.

[0117] The nonlinear formula substitutes both together to obtain a mean stress.

[0118] Substitute the data, σ =41.240MPa.

[0119] The nonlinear model takes into account the difference in biaxial curvature and predicts stress values ​​under large deformation conditions that are significantly higher than those of the linear model and closer to the measured values.

[0120] The linear model is suitable for small deformation and isotropic scenes, and has the characteristics of fast calculation speed. The nonlinear model is suitable for large deformation and anisotropic scenes, with higher accuracy and support for complex structure analysis. Furthermore, the linear model can be used for rapid estimation in the early stage of electroforming, and then switched to the nonlinear model for fine correction in the later stage. The model is adjusted by E s 、ν s and other parameters, suitable for various substrate materials such as silicon, copper, and nickel alloys.

[0121] In some embodiments of the present invention, in step S100, providing a conductive layer on the electroformed substrate includes the following steps:

[0122] S101, plating a layer of metal titanium or metal chromium on the electroformed substrate as a bonding layer;

[0123] S102, plating a layer of metal platinum on the bonding layer as a conductive layer.

[0124] Titanium and chromium have high chemical stability and strong interface bonding ability, and can form chemical bonds with the silicon substrate (such as Ti-Si or Cr-Si bonds), while providing a flat attachment surface for subsequent conductive layers. Platinum has an extremely low resistivity (10.6×10 -8 Ω·m) and electrochemical corrosion resistance to ensure uniform current distribution during electroforming and reduce local overheating or dendrite growth.

[0125] The specific steps are as follows;

[0126] A single-side polished silicon wafer (80 mm × 15 mm) was immersed in an acetone solution for ultrasonic cleaning for 10 minutes to remove organic contaminants from the surface. The wafer was then rinsed with deionized water, dried with nitrogen, and placed on the sample stage of the sputtering chamber.

[0127] Set the vacuum degree of the chamber to 5×10 - 2 mbar, argon was introduced as the working gas (flow rate of approximately 20 sccm). The chromium target power supply was turned on, the sputtering current was adjusted to 130 mA, and the sputtering time was 170 s to form a chromium bonding layer with a thickness of approximately 75 nm.

[0128] Replace the platinum target, maintaining the vacuum level and argon flow rate. Adjust the sputtering current to 35 mA and the sputtering time to 330 s to form a platinum conductive layer with a thickness of approximately 100 nm.

[0129] The sputtered substrate was immersed in an ethanol solution for ultrasonic cleaning for 15 minutes to remove surface particulate contaminants. After rinsing with deionized water, it was dried with nitrogen and set aside.

[0130] The above treatment significantly reduces the risk of interface delamination during the electroforming process, lowering defect rates. It also ensures uniform electroforming current distribution, reduces variations in deposited layer thickness, and reduces cracks or warping in the cast layer caused by stress concentration, thereby improving product qualification rates.

[0131] In some embodiments of the present invention, in step S101, the sputtering condition is a vacuum degree of 5×10 - ² mbar, current 100-150 mA, time 100-200 s;

[0132] A high vacuum environment reduces oxygen and impurities in the chamber, preventing oxidation of titanium (Ti) or chromium (Cr) during the sputtering process and ensuring film purity and chemical stability. Current directly influences the target sputtering rate. A current of 100-150 mA forms a dense, pore-free film. Film thickness (50-100 nm) is adjusted over time to balance bonding strength and substrate stress. A short duration (100 s) reduces thermal stress accumulation, while a longer duration (200 s) enhances interfacial adhesion.

[0133] Before electroforming, the pre-treated silicon wafer and the wire are bonded with conductive tape to ensure full contact between the wire and the non-polished side of the silicon wafer. The cathode fixture and the silicon wafer with the wire bonded are bonded in the same way, so that the wire is between the silicon wafer and the fixture. During electroforming, the anode and cathode fixtures are installed in the electroforming tank, and the knobs on the side of the fixture are tightened to fix the position of the fixture. The electroforming process current and electroforming time are set on the same dedicated pulse power supply, and finally the current is connected. The length of each electroforming section is determined by dividing the layer thickness by the electroforming rate, which is 12µm / (A / dm2).

[0134] In step S102, the sputtering condition is a vacuum degree of 5×10 - ² mbar, current 35 mA, time 100-350 s.

[0135] Maintaining the same vacuum environment as the bonding layer prevents uneven platinum (Pt) film structure due to pressure fluctuations. The sputtering rate of the Pt target is sensitive to current. At a current of 35 mA, Pt is deposited at a rate of 0.3-0.5 nm / s, forming nanoscale grains (<20 nm) and reducing surface roughness (Ra ≤ 5 nm). A short time (100 s) to form a 30 nm conductive layer meets basic requirements; a longer time (350 s) to form a 106 nm thick layer enhances corrosion resistance and current carrying capacity.

[0136] In some embodiments of the present invention, in step S100 , providing a conductive layer on the electroformed substrate includes the following steps: attaching a layer of conductive tape on the electroformed substrate.

[0137] In this embodiment, the method of sputtering a metal layer on the substrate surface can cause the cast layer to fall off during electroforming due to substrate warping and deformation, resulting in failure to complete electroforming and, consequently, failure to measure stress. In this embodiment, conductive tape was applied to the substrate, and no fall-off was observed during the experiment. Only slight warping was observed at the substrate edges, significantly improving the success rate.

[0138] In some embodiments of the present invention, in step S300:

[0139] The current density of DC current is 0.4-2A / dm²;

[0140] The frequency of the pulse current is 0.1-1200 Hz, the average current density is 0.2-2 A / dm², and the duty cycle is 0.1-0.9.

[0141] Direct current provides a continuous electric field that drives metal ions (such as Ni² + ) is uniformly deposited on the cathode (silicon substrate) surface. Current density directly affects the deposition rate: at low currents, ion migration is slow, the deposited layer has finer grains, and residual stress is low. High current densities result in faster deposition rates, but also greater internal stress in the cast layer, significant warping, and poor flatness.

[0142] The pulse current achieves the following effects through periodic on-off (Ton / Toff):

[0143] Ton stage: High instantaneous current (e.g., peak value of 2 A / dm²) accelerates ion migration and fills microscopic depressions. Toff stage: Current interruption restores the solution concentration gradient, reduces concentration polarization, and suppresses hydrogen bubbles and defects.

[0144] Low frequency is suitable for thick layer deposition, with sufficient Toff time to reduce internal stress. High frequency is suitable for nano-scale fine structures, and short pulses can suppress grain growth.

[0145] At low duty cycles, the Toff time can be extended, improving the density of the deposited layer. At high duty cycles, it approaches a DC effect, improving deposition efficiency.

[0146] Pulsed current mode is particularly effective in reducing residual stress and improving surface quality, while DC mode remains valuable for high-efficiency deposition. By optimizing parameter selection, a reliable solution can be provided for the manufacture of high-precision, low-stress metal mold cores.

[0147] In some embodiments of the present invention, data is collected through a single-chip microcomputer-based hardware module, including an STM32 single-chip microcomputer, an ammeter, a voltmeter, a pH meter, a temperature sensor, and a display screen. The host STM32 communicates with other slaves through the Modbus communication protocol, and the collection of voltage, current, pH, temperature, and time can be achieved.

[0148] In some embodiments of the present invention, in step S600: establishing an artificial neural network model includes the following steps:

[0149] S601, selecting a deep neural network according to the type of collected data;

[0150] S602, select a suitable activation function, the number of neural network layers and the number of neurons in each layer;

[0151] S603: Select a loss function for the regression task.

[0152] In this embodiment, a deep neural network is selected based on the type of data being collected. The neural network type is not fixed and includes, but is not limited to, feedforward neural networks, convolutional neural networks, and recurrent neural networks. For one-dimensional data, a feedforward neural network is generally chosen, suitable for time-independent scalar data such as current density, temperature, and pH. For two-dimensional data, a convolutional neural network can be selected; and for cases where time is also input, a recurrent neural network is generally chosen.

[0153] Select a suitable activation function, the number of neural network layers and the number of neurons in each layer that match the complexity of the model. These parameters can be obtained through automatic machine learning.

[0154] Activation function selection: ReLU (Rectified Linear Unit): used in hidden layers to prevent gradient vanishing and accelerate convergence. Sigmoid: mainly used for classification problems, compressing the results to the [0, 1] range.

[0155] Network Structure Configuration: Grid search was used to determine the optimal combination. For example, a 3-node input layer, 2 hidden layers (14 neurons each), and a 3-node output layer. Automated Machine Learning (AutoML): Using the scikit-learn and skorch libraries, the goal was to minimize the mean squared error (MSE) on the validation set.

[0156] The loss function can be a common function for regression tasks, such as mean square error (MSE) and mean absolute error (MAE), and the evaluation indicator can be the coefficient of determination (R2) or MSE, MAE.

[0157] Mean Squared Error (MSE): used for regression tasks, calculates the squared difference between the predicted stress and the true value. Mean Absolute Error (MAE): more robust to outliers.

[0158] Compared with the traditional multiple linear regression model, it improves the prediction accuracy, optimizes the training efficiency, and enhances the generalization ability.

[0159] In some embodiments of the present invention, in step S700 : the generated process parameter step size is 0.01, and the stress value is output after inputting the model to optimize the electroforming parameter selection.

[0160] In this implementation, experimental data acquired from the in-situ stress detection platform is processed to generate a series of data, either through code or other means. This series of data is then fed into a trained model with a step size of 0.01. The model then predicts the stress under the corresponding process conditions. Based on the resulting series of predicted values, the one with the lowest stress is selected, thus determining the corresponding process conditions as optimal. A process parameter step size of 0.01 allows sufficient simulation experimental data to be obtained to select the process parameters that meet the requirements, while also preventing the data from becoming excessively large and resulting in high processing costs.

[0161] In some embodiments, the current density is 0.4-2.0 A / dm², with a step size of 0.01 A / dm², generating 160 candidate values. The pH value is 3.5-4.5, with a step size of 0.01, generating 100 candidate values. The temperature is 25-65°C, with a step size of 0.1°C, generating 400 candidate values. The full parameter combination is generated by Cartesian product (160×100×400=6.4×10 6 ), Latin Hypercube Sampling was used to extract 5000 groups of representative data to reduce the amount of calculation.

[0162] Input format: Convert parameter combinations into PyTorch tensors (shape 5000 × 3). Batch inference: Use the trained MLP model for parallel predictions, with a single prediction time of approximately 2–3 seconds. Output filtering: Extract the top 10 parameter combinations with the lowest predicted stress values. Select the parameter combination with the lowest predicted stress (e.g., current density 1.23 A / dm², pH = 3.98, temperature 52.5°C). Conduct electroforming experiments with the selected parameters and verify the error.

[0163] The step size of 0.01 (current density) and 0.1°C (temperature) covers the micro-process window, and the optimization accuracy is improved by 10 times compared with the traditional step size (0.1 A / dm², 1°C).

[0164] The present invention also discloses a casting layer stress prediction system, including a memory and a processor, wherein the memory stores a determination program of a casting layer stress prediction method based on an artificial neural network, and the processor runs the determination program of the casting layer stress prediction method based on an artificial neural network, so that the casting layer stress prediction system executes the casting layer stress prediction method based on an artificial neural network.

[0165] refer to Figure 4 As shown, Figure 4This is a scatter plot showing the effect of different numbers of hidden layers and nodes on the coefficient of determination in the cast layer stress prediction method. As the number of hidden layers and nodes per layer varies, within the current range, the coefficient of determination approaches 0.5, indicating that the model's predictive ability is superior to the mean prediction. If the model's predictive ability is close to the mean prediction, the coefficient of determination will be close to 0.

[0166] refer to Figure 5 As shown, Figure 5 It is the loss curve of the network model training in the casting layer stress prediction method. Figure 5 The results show that using a 5-fold cross-validation strategy, the mean squared error for each training fold approaches 0 after a certain number of iterations, indicating that the model has learned the patterns in the data. The inconsistent convergence speed is due to the use of a 5-fold cross-validation strategy, in which the training and validation data for each fold are slightly different.

[0167] from Figure 4 and Figure 5 It can be seen that the method for predicting casting layer stress based on artificial neural network of the present invention has great application potential for predicting stress values.

[0168] The present invention also discloses a computer-readable storage medium, comprising: a determination program of a casting layer stress prediction method based on an artificial neural network is stored on the computer-readable storage medium; when the determination program of the casting layer stress prediction method based on an artificial neural network is executed by a processor, the casting layer stress prediction method based on an artificial neural network is implemented.

[0169] Since the casting layer stress prediction system and computer-readable storage medium adopt all the technical solutions of the casting layer stress prediction method based on artificial neural network in the above embodiment, they at least have all the beneficial effects brought by the technical solutions of the above embodiment, which will not be repeated here.

[0170] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the scope of the present invention.

Claims

1. A casting layer stress prediction method based on artificial neural network, characterized in that: include: S100, electroforming substrate pretreatment: using a silicon wafer as an electroforming substrate, polishing one side of the electroforming substrate, and providing a conductive layer on the other side; S200, adjusting the in-situ stress detection platform: configuring an electroforming tank and an optical measurement module, wherein the optical measurement module includes a collimated light source and a wavefront sensor, and adjusting the relative positions of the collimated light source, the wavefront sensor, and the electroforming tank so that the light beam of the collimated light source is reflected by the polished surface of the silicon wafer and enters the wavefront sensor; S300, electroforming experiment: electroforming is performed in the electroforming tank using a direct current or a pulse current; S400, data acquisition: including a data acquisition module, through which electroforming process data and laser wavefront curvature radius are collected; S500, data processing: calculating a stress value according to the laser wavefront curvature radius, and converting the stress value data and the electroforming process data into a neural network adaptation format; S600, establishing an artificial neural network model: selecting a network type, an activation function, and hyperparameters, and training the artificial neural network model to fit the nonlinear relationship between electroforming parameters and stress values; S700, stress prediction: inputting electroforming parameters into the trained artificial neural network model, and outputting predicted casting layer stress values.

2. The casting layer stress prediction method based on artificial neural network according to claim 1, characterized in that: In step S500, the calculation of the stress value includes the following steps: S501, calculating the curvature radius of the sample; S502, calculating an average stress value according to the curvature radius of the sample, wherein the calculation of the average stress value includes a linear scenario and a nonlinear scenario; The formula for the radius of curvature of the sample is: in: is the radius of curvature of the sample, is the radius of curvature of the laser wavefront, is the distance between the sample and the sensor, is the refractive index of the electrolyte; In a linear scenario, the theoretical model formula of the average stress value is: in: is the average stress value, is the curvature radius of the substrate before electrodeposition, is the curvature radius of the substrate after electrodeposition; ， ， ， ； , , , are the biaxial elastic modulus, elastic modulus, Poisson's ratio, and thickness of the substrate, respectively; , , , are the biaxial elastic modulus, elastic modulus, Poisson’s ratio and thickness of the deposited layer, respectively; In a nonlinear scenario, the theoretical model formula of the mean stress is: in, is the average stress value, and are the base deformation curvatures in the x-axis and y-axis directions, respectively.

3. The casting layer stress prediction method based on artificial neural network according to claim 1 is characterized in that: In step S100, providing a conductive layer on the electroformed substrate includes the following steps: S101, plating a layer of metal titanium or metal chromium on the electroformed substrate as a bonding layer; S102, plating a layer of metal platinum on the bonding layer as a conductive layer.

4. The method for predicting casting layer stress based on artificial neural network according to claim 3, wherein: In step S101, the sputtering condition is a vacuum degree of 5×10 - ² mbar, current 100-150 mA, time 100-200 s; In step S102, the sputtering condition is a vacuum degree of 5×10 - ² mbar, current 35 mA, time 100-350 s.

5. The casting layer stress prediction method based on artificial neural network according to claim 1 is characterized in that: In step S100 , providing the conductive layer on the electroformed substrate includes the following steps: pasting a layer of conductive tape on the electroformed substrate.

6. The casting layer stress prediction method based on artificial neural network according to claim 1, characterized in that: In step S300: The current density of the direct current is 0.4-2A / dm²; The frequency of the pulse current is 0.1-1200 Hz, the average current density is 0.2-2 A / dm², and the duty cycle is 0.1-0.

9.

7. The casting layer stress prediction method based on artificial neural network according to claim 1 is characterized in that: In step S600: establishing the artificial neural network model includes the following steps: S601, selecting a deep neural network according to the type of collected data; S602, select an activation function, the number of neural network layers and the number of neurons in each layer; S603: Select a loss function for the regression task.

8. The casting layer stress prediction method based on artificial neural network according to claim 1 is characterized in that: In step S700 : the generated process parameter step size is 0.01, and the stress value is output after inputting the model to optimize the electroforming parameter selection.

9. A casting layer stress prediction system, characterized by: It includes a memory and a processor, wherein the memory stores a determination program of the casting layer stress prediction method based on an artificial neural network, and the processor runs the determination program of the casting layer stress prediction method based on an artificial neural network, so that the casting layer stress prediction system executes the casting layer stress prediction method based on an artificial neural network as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that include: The computer-readable storage medium stores a determination program of a casting layer stress prediction method based on an artificial neural network. When the determination program of the casting layer stress prediction method based on an artificial neural network is executed by a processor, the casting layer stress prediction method based on an artificial neural network according to any one of claims 1 to 8 is implemented.

Images