An intelligent wafer etching process

Abstract

The present invention relates to an intelligent wafer etching process in the field of information technology, comprising: obtaining etching rates of different materials under various chemical solution formulations, constructing an etching rate database covering various material systems, screening out material combinations with high etching selectivity by comparing etching rate differences in the database, and determining a targeted optimization direction for the chemical solution formulation; installing a multi-parameter online monitoring unit in the etching equipment, using an optical fiber spectrometer to measure chemical solution concentration in real time, and using a thermocouple and a pH electrode to measure chemical solution temperature and pH in real time; when the monitored parameters exceed a set range, starting a chemical solution ratio adjustment program, and realizing dynamic regulation of chemical solution properties by controlling liquid inlet and outlet volumes, thereby ensuring the stability and consistency of the etching process.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an intelligent wafer etching process. Background Art

[0002] Improving etching uniformity and controlling chemical parameters in real time are two closely related technical challenges in semiconductor etching. Traditional immersion etching struggles to ensure uniformity across large wafer areas, and while spray etching can improve this issue, it still suffers from uneven etching in localized areas.

[0003] These problems are primarily due to the uneven flow and diffusion of the chemical solution across the wafer surface. Furthermore, the concentration, temperature, and pH of the chemical solution change over time during the etching process, affecting the etching rate and selectivity. Designing an efficient spray system and stirring device to ensure uniform distribution of the chemical solution across the wafer surface and to monitor and control chemical parameters in real time are key issues that require urgent resolution. This requires comprehensive consideration of factors such as nozzle layout, spray angle, and the shape and rotational speed of the stirring blades. Furthermore, a quantitative relationship must be established between chemical solution parameters and etching performance, and an intelligent online monitoring and feedback control system must be developed. Resolving these interrelated technical challenges is crucial for improving the stability and consistency of the etching process. Summary of the Invention

[0004] The present invention provides an intelligent wafer etching process, which includes the following steps:

[0005] S101, obtaining the etching rates of different materials under various chemical solution formulations, building an etching rate database covering multiple material systems, and screening out material combinations with high etching selectivity by comparing the etching rate differences in the database to determine the targeted optimization direction of the chemical solution formulation;

[0006] S102: For the selected material combinations, first-principles calculations are used to study the interaction mechanism between each component in the solution and the material surface, obtain the kinetic parameters of the surface reaction, and perform molecular dynamics simulations to reveal the microscopic mechanism of the solution diffusion and mass transfer process. A multi-scale coupled etching process model is established, including a molecular-scale surface reaction model, a mesoscopic-scale diffusion and mass transfer model, and a macroscopic-scale etching morphology evolution model. Through the coupling of these models, the rate-controlling steps and chemical reaction pathways of material removal are clarified, guiding the optimization of the solution formulation.

[0007] S103 uses microfluidic chip technology to design and integrate multiple parallel etching reaction chambers to achieve high-throughput screening of drug solution formulas. By controlling the flow rate of the sampling pump and the drug solution ratio, a concentration gradient is formed within the chip. Spectroscopy is used to monitor the drug solution composition at the chip outlet in real time to obtain etching rate data for different formulations. Machine learning algorithms are used to establish a quantitative relationship between formulation and etching performance, resulting in a drug solution formula with high etching selectivity.

[0008] S104. Based on the microfluidic chip screening results, a spray head array and stirring device are designed for a solution formulation with high etching selectivity to form a uniform solution spray and turbulent flow field. By optimizing the nozzle shape, number of nozzles, spray angle, and stirring blade parameters, the mass transfer process between the solution and the material surface is enhanced, improving etching uniformity. The solution recovery device uses dual parameter control of temperature and conductivity to adjust the recovery time in real time according to changes in the solution properties, reducing solution waste.

[0009] S105: Install a multi-parameter online monitoring unit in the etching equipment. Use a fiber optic spectrometer to measure the concentration of the chemical solution in real time. Use a thermocouple and pH electrode to measure the temperature and pH of the chemical solution in real time. When the monitored parameters exceed the set range, the chemical solution ratio adjustment program is activated. By controlling the liquid inlet and outlet volumes, the chemical solution properties are dynamically regulated to ensure the stability and consistency of the etching process.

[0010] S106: Characterize and evaluate the etching results after dynamic control, using etching depth, surface roughness, and etching residue as evaluation indicators. Optical microscopy and atomic force microscopy are used to characterize the etching results. The etching performance of different material systems and solution formulations is quantitatively evaluated by calculating the uniformity index and selectivity index. The uniformity index is calculated as the ratio of the standard deviation of the etching depth to the average value, and the selectivity index is calculated as the ratio of the etching rates of the two materials. For the performance evaluation results, gray correlation analysis is used to determine key process parameters to guide the learning and iterative update of the intelligent optimization system.

[0011] S107, based on process optimization experimental data, constructs a nonlinear mapping model between solution formula, process parameters and etching results. The model adopts a multi-layer neural network structure, with the input layer being the solution formula and process parameters, and the output layer being the etching depth, uniformity and selectivity indicators. The obtained etching laws and control strategies are transformed into dynamically updated knowledge reasoning rules, and an intelligent decision-making system based on case reasoning and fuzzy logic is developed to realize the automation and intelligence of solution formula design, process parameter selection and etching defect diagnosis.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0013] The present invention discloses an intelligent wafer etching process, which constructs a multi-material etching rate database, combines first-principles calculations and molecular dynamics simulations, establishes a multi-scale coupled etching process model, uses microfluidic chip technology to achieve high-throughput screening of liquid formulations, and uses machine learning algorithms to optimize the formulations. The present invention designs dynamically controlled etching equipment, a spray head array, and a stirring device, and also includes a uniform spray system and a multi-parameter online monitoring unit, which realizes real-time control of the etching process. By characterizing and evaluating the etching results, a nonlinear mapping model is established, and an intelligent decision-making system based on case reasoning and fuzzy logic is developed. The present invention realizes the automation and intelligent optimization of the etching process, significantly improves the etching uniformity and selectivity, and provides an innovative solution for the etching process in microelectronics manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The figure is a flow chart of an intelligent wafer etching process of the present invention.

[0015] Figure 2 A schematic diagram of an intelligent wafer etching process of the present invention.

[0016] Figure 3 This is another schematic diagram of an intelligent wafer etching process of the present invention. DETAILED DESCRIPTION

[0017] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1-3 In this embodiment, an intelligent wafer etching process may specifically include:

[0019] S101. Obtain the etching rates of different materials under various chemical solution formulations and construct an etching rate database covering multiple material systems. By comparing the etching rate differences in the database, select material combinations with high etching selectivity and determine the direction for targeted chemical solution formulation optimization.

[0020] S101 also includes: obtaining etching rate data covering a preset number of types of material systems and liquid formulations, and establishing the etching rate database; using a data analysis algorithm to analyze the etching rate data in the etching rate database, and calculating the etching rate differences of different material combinations under the same liquid formulation; judging the selectivity of the material combination based on the etching rate differences, and screening out material combinations with high etching selectivity; for the material combination with high etching selectivity, using a machine learning algorithm to analyze the relationship between the material combination and the liquid formulation parameters, and obtaining key liquid formulation parameters that affect the etching selectivity; determining the liquid formulation optimization direction based on the key liquid formulation parameters, and optimizing the liquid formulation by adjusting the value range of the key liquid formulation parameters; applying the optimized liquid formulation to the material combination with high etching selectivity, and verifying the optimization effect through etching experiments. If the preset optimization target is not achieved, returning to the liquid formulation optimization step, and continuing the optimization until the optimal liquid formulation is obtained to achieve high-selective etching of the material.

[0021] For example, we first need to clarify the different material systems and chemical solution formulas. Here are two examples of material systems. One is the silicon-based material commonly used in integrated circuit manufacturing, including single crystal silicon, polycrystalline silicon, silicon dioxide, silicon nitride, etc.; the other is a new two-dimensional material system, such as molybdenum disulfide, tungsten diselenide, etc. The chemical solution formula depends on different materials and etching requirements. Common ones include acidic solutions, alkaline solutions, and solutions containing oxidants and chelating agents. For example, hydrofluoric acid solution is often used to etch silicon dioxide, and potassium hydroxide solution is used to etch single crystal silicon. These are all selective wet etching. Phosphoric acid, nitric acid, and acetic acid are mixed in different proportions. Single crystal silicon, silicon dioxide, and silicon nitride, commonly used in the integrated circuit field, are placed in mixed solutions of different proportions and etched at the same temperature and time. The thickness change is then measured. The thickness change and process time under different process parameters and environments are calculated to obtain the etching rate. The relevant data is collected and classified uniformly. This allows for the acquisition of etching rate data for various process parameters, different materials, different formulations, different temperatures, and different times. Subsequently, these acquired etching rate data will be stored in a dedicated database. As experimental data continues to accumulate, the database content will continue to grow, forming an etching rate database covering different material systems, solution formulations, temperatures, times, and other parameters. This constructed etching rate database will include etching rates for different materials. Different materials will exhibit different etching rates under different combinations of chemical formulations and parameters, making it easier for users to understand the precautions and key points that need to be avoided in process design. With such a comprehensive and detailed etch rate database, data analysis methods and algorithms can be used to deeply explore trends reflected in the database data and conduct in-depth analysis of the data. For example, suppose that under the same chemical solution formulation, the etch rate of material A is 10 nm / min, while the etch rate of material B is 1 nm / min. By dividing the etch rates of different material combinations under the same chemical solution formulation, we can intuitively compare the difference in etch rates between the two materials under that formulation, i.e., the etch selectivity. In this example, the etch selectivity of material A to material B is 10:1, indicating that under this chemical solution formulation, material A is etched 10 times faster than material B. If the primary goal is to etch material A while preserving material B, and A and B are located in different parts of the same device or product, a 10:1 etch selectivity of material A to material B is clearly a good result. This value indicates that material A is effectively etched while minimally affecting material B. For the material combinations with high etching selectivity screened by the above method, machine learning methods can be further used to optimize key parameters. Machine learning methods such as multivariate linear regression and artificial neural networks can be used.The input is material and chemical information, such as silicon dioxide in a solution of phosphoric acid, nitric acid, and acetic acid in a specific ratio. A large amount of recipe data with different ratios is used to build a model. The key component concentrations and temperature in these solutions are then extracted as key parameters. The output is the silicon dioxide etching rate. The trained model can then be used for subsequent recipe prediction and optimization. For example, if the desired silicon dioxide etching rate is 8 nanometers / minute, and the current recipe yields an etching rate of 6 nanometers / minute, with 8 nanometers / minute as the target model output, the optimization model can identify parameters, such as the phosphoric acid content and temperature, that can increase the etching rate to 8 nanometers / minute without affecting other performance indicators, such as corrosion of metals exposed to the etchant. The model can also provide an approximate parameter range or value, making it possible to achieve targeted etching rate optimization in a relatively short period of time. Based on the key solution recipe parameters and guidance for optimization derived from machine learning model analysis, new solutions are formulated and experimented with. For example, if the model suggests increasing the phosphoric acid content and raising the temperature in the solution, one can then try modifying one or both of these parameters in a specific ratio. After optimizing the chemical formula, for example, by increasing the phosphoric acid concentration to a certain percentage and setting the temperature to a certain degree Celsius, the new formula is applied to the aforementioned material combinations with high etch selectivity. Experimentation allows precise determination of the actual etch rate for each material using this new formula. After etching, the film thickness change is measured to accurately calculate the etch rate. The surface morphology is then examined to determine whether the new parameters or ranges are reasonable. After obtaining this data, a comprehensive evaluation is conducted on the optimization results. For example, the silicon dioxide etch rate has indeed increased to approximately 8 nanometers / minute as expected. If the optimization results meet expectations, the chemical formula has been successfully improved. On the other hand, if the actual results fall short of expectations or if other negative effects, such as metal corrosion, occur, the machine learning model analysis step needs to be re-evaluated. Further learning is required to fit the newly acquired data, verify the differences between the new experimental data and the previous data, and then proceed with further parameter adjustment guidance or optimization target definition, refine the chemical formula, and conduct further experimental verification. Through the above means, after repeated iterations and optimizations, the optimal liquid medicine formula is finally obtained. Under the guidance of the optimal medicine formula, this etching process can be used in the production of integrated circuits or chips, improving processing efficiency and yield, and promoting the development of advanced processes.

[0022] S102. For the selected material combinations, first-principles calculations were used to investigate the interaction mechanisms between the components in the solution and the material surface, obtaining the kinetic parameters of the surface reaction. Molecular dynamics simulations were performed to reveal the microscopic mechanisms of the solution's diffusion and mass transfer processes. A multi-scale coupled etching process model was established, including a molecular-scale surface reaction model, a mesoscopic-scale diffusion and mass transfer model, and a macroscopic-scale model of the etched morphology evolution. By coupling these models, the rate-controlling steps and chemical reaction pathways of material removal were elucidated, guiding the optimization of the solution formulation.

[0023] S102 also includes: constructing a molecular model of the interaction between the material surface and the liquid components for the screened material combination; using first-principles calculations to obtain the energy changes and reaction rate constant kinetic parameters of the surface reaction in the molecular model; establishing a diffusion and mass transfer model of the liquid near the material surface at the mesoscopic scale to calculate the concentration distribution and mass transfer flux of the liquid components; achieving a coordinated simulation of molecular-scale surface reactions, mesoscopic-scale diffusion and mass transfer, and macroscopic-scale etching morphology evolution through the coupling of multi-scale models; and revealing the rate-controlling steps and chemical reaction paths of the material removal process based on the results of the coordinated simulation.

[0024] Exemplary, according to the selected material combination, a molecular model of the interaction between the material surface and the liquid component is constructed, and the kinetic parameters such as the energy change and reaction rate constant of the surface reaction are obtained by first-principles calculation. For example, for the interaction between silicon-based materials and hydrofluoric acid solution, the adsorption model of the silicon surface and the hydrofluoric acid molecule can be constructed, and the energy change of the adsorption energy, dissociation energy and reaction path of the hydrofluoric acid molecule on the silicon surface can be calculated using first-principles calculation methods such as density functional theory (DFT). Through these calculations, the activation energy and reaction rate constant of the silicon surface and the hydrofluoric acid reaction can be obtained, and these parameters are the basis for understanding the etching process. Molecular dynamics simulation is used to study the adsorption, desorption and diffusion behavior of each component in the liquid on the material surface, and mass transfer parameters such as the microscopic mechanism and diffusion coefficient of the diffusion mass transfer process are obtained. Taking silicon dioxide and phosphoric acid solution as an example, through molecular dynamics simulation, the adsorption and desorption process of phosphoric acid molecules on the silicon dioxide surface, as well as its diffusion behavior on the surface can be observed. By using the molecular motion trajectories during statistical simulation, the diffusion coefficient of phosphoric acid on the silica surface can be calculated. These data help understand the mass transfer mechanism of chemical solution components on the material surface. Based on surface reaction kinetic parameters and diffusion mass transfer parameters, a molecular-scale surface reaction model is established to describe the chemical reaction process and reactant concentration changes on the material surface. For example, based on the kinetic parameters of the reaction between silicon and hydrofluoric acid calculated above, a reaction model can be constructed to describe the time-varying hydrofluoric acid concentration on the silicon surface. This model can simulate the chemical reaction rate and reactant consumption on the silicon surface under different hydrofluoric acid concentrations. At the mesoscopic scale, considering the flow and mass transfer processes of the chemical solution near the material surface, a diffusion mass transfer model is established to calculate the concentration distribution and mass transfer flux of the chemical solution components. Taking tungsten diselenide and nitric acid solution as an example, computational fluid dynamics (CFD) methods can be used to simulate the flow field and concentration distribution of nitric acid solution near the tungsten diselenide surface. This simulation can determine the concentration changes and mass transfer flux of nitric acid at different locations, thereby understanding the mass transfer process of the chemical solution on the material surface. At the macroscale, an etching morphology evolution model is established based on the chemical reactions on the material surface and the mass transfer process of the liquid solution. This model simulates the changes in the material surface morphology over time and predicts the etching depth and etching rate. For example, for the etching process of single-crystalline silicon with potassium hydroxide solution, a macromodel that includes chemical reactions and mass transfer processes can be established to simulate the morphological changes on the single-crystalline silicon surface at different time points and predict the etching depth and etching rate. By coupling multi-scale models, a coordinated simulation of molecular-scale surface reactions, mesoscopic-scale diffusion mass transfer, and macroscale etching morphology evolution is achieved, revealing the rate-controlling steps and chemical reaction pathways of the material removal process. Taking the etching of polycrystalline silicon with acidic solution as an example, by coupling the molecular-scale reaction model, the mesoscopic-scale mass transfer model, and the macroscale morphology evolution model, the etching process of polycrystalline silicon in acidic solution can be systematically simulated, revealing the controlling steps and main chemical reaction pathways of the etching rate.Based on the coupled simulation results, the effects of process parameters such as the concentration of chemical components and temperature on the etching rate and etching morphology are analyzed, and the chemical solution formulation and process conditions are optimized to improve the selectivity and uniformity of material removal. For example, through simulation, it was found that increasing the hydrofluoric acid concentration can increase the etching rate of silicon, but too high a concentration will lead to uneven etching. By optimizing the hydrofluoric acid concentration and temperature, the uniformity and selectivity of etching can be improved while maintaining a high etching rate. Through the above multi-scale simulation and analysis, we can deeply understand the microscopic mechanism of the interaction between materials and chemical solutions, guide the optimization of chemical solution formulation and process conditions, and ultimately improve the efficiency of the etching process and product quality. This method is not only applicable to silicon-based materials in integrated circuit manufacturing, but can also be extended to the etching process research of other new materials, and has broad application prospects.

[0025] S103. Using microfluidic chip technology, we designed and integrated multiple parallel etching reaction chambers to achieve high-throughput screening of chemical solution formulations. By controlling the flow rate of the sampling pump and the chemical solution ratio, a concentration gradient is formed within the chip. Spectroscopy is used to monitor the chemical solution composition at the chip outlet in real time, acquiring etching rate data for different formulations. Machine learning algorithms are used to establish a quantitative relationship between formulation and etching performance, resulting in chemical solution formulations with high etching selectivity.

[0026] S103 also includes: obtaining the number of multiple parallel etching reaction chamber structures formed in the chip, and configuring the initial component ratios of the corresponding number of liquid medicines at the entrances of different reaction chambers according to the number; using spectroscopy to detect the preset number of types of liquid medicines corresponding to the outputs of different reaction chambers at the chip outlet, and judging whether there is a liquid medicine component of a certain reaction chamber in the detected liquid medicine that meets the preset spectral characteristic peak threshold range, and if so, obtaining the position data of the etching reaction chamber corresponding to the liquid medicine on the chip; according to the position data, obtaining the liquid medicine concentration information and flow rate control information of the corresponding sampling pump, and obtaining the liquid medicine component information data of a preset number of types and different concentrations, and converting the liquid medicine component information data into the liquid medicine component information data. The information data is associated one by one with the numbers of the corresponding multiple reaction chambers on the chip; according to the different component ratios of the liquid medicine corresponding to the reaction chambers at different positions on the same chip, the performance data of different etching rates corresponding to the liquid medicine with different component ratios are determined through the association to obtain an etching performance data set; a random forest algorithm is used to analyze and process the obtained liquid medicine component data of different concentrations and the etching performance data to establish an etching performance prediction model; the liquid medicine etching formula is screened by optimizing the input variables to determine whether the output variables obtained from the input formula data set of the etching performance prediction model meet the preset conditions. If so, the corresponding liquid medicine formula is determined as a liquid medicine formula with high etching selectivity.

[0027] Exemplarily, according to the different flow rate changes of the injection pump within the set time, different component ratios of the liquid at the inlet of the chip are configured. For example, assuming that an etching liquid containing hydrofluoric acid (HF) and nitric acid (HNO3) needs to be configured, the injection pump can deliver different volumes of HF and HNO3 respectively within the set time by precisely controlling the flow rate, thereby realizing the configuration of liquids with different component ratios. Assuming that the set time is 10 minutes, the injection pump delivers HF at a flow rate of 10 ml per minute in the first 5 minutes, and delivers HNO3 at a flow rate of 5 ml per minute in the last 5 minutes, and finally obtains a liquid with a volume ratio of HF and HNO3 of 2:1. According to the number of multiple parallel etching reaction chamber structures formed in the chip, the initial component ratios of the liquids with corresponding number of component ratios at the inlets of different reaction chambers are configured. Assuming that there are four parallel etching reaction chambers on the chip, four liquids with different component ratios can be configured respectively. For example, reaction chamber 1 is configured with a solution containing HF and HNO3 in a 2:1 volume ratio, reaction chamber 2 is configured with a 3:1 volume ratio, reaction chamber 3 is configured with a 1:1 volume ratio, and reaction chamber 4 is configured with a 4:1 volume ratio. Spectroscopy is used to detect the various solutions output from different reaction chambers at the chip outlet. Spectroscopy can determine the composition and concentration of the solutions by measuring the characteristic spectral peaks of the solutions. For example, a UV-visible spectrometer is used to detect the solutions at the reaction chamber outlets. If a characteristic HF absorption peak is detected at a specific wavelength in the solution in a particular reaction chamber, and the peak intensity is within a preset threshold range, the solution composition in that reaction chamber meets expectations. The position data of the etching reaction chamber corresponding to that solution is obtained on the chip. Assuming that the solution composition in reaction chamber 2 meets the preset threshold range for the characteristic spectral peak, the specific location information of reaction chamber 2 on the chip is recorded, such as the coordinates (2, 3). The solution concentration information and flow rate control information for the injection pump corresponding to the reaction chamber position data on the chip are obtained. Based on the recorded location information, the control log of the injection pump was consulted, revealing that the solution configuration for reaction chamber 2 was a 3:1 volume ratio of HF and HNO₃. The injection pump delivered HF at a flow rate of 10 ml / min for the first 6 minutes and HNO₃ at a flow rate of 5 ml / min for the last 4 minutes. Data on the composition of the solution at various concentrations was obtained and associated with the corresponding reaction chamber numbers on the chip. For example, the solution composition information for reaction chambers 1, 2, 3, and 4 was recorded as (2:1, 3:1, 1:1, 4:1), respectively, and associated with their respective numbers (1, 2, 3, 4). Based on the different composition ratios of the solution corresponding to reaction chambers at different locations on the same chip, data association was used to determine the performance data for the different etching rates corresponding to the solution ratios. For example, experimental measurements revealed an etching rate of 5 μm / min for reaction chamber 1, 7 μm / min for reaction chamber 2, 4 μm / min for reaction chamber 3, and 8 μm / min for reaction chamber 4. These etching rate data are associated with the corresponding chemical solution component ratios to form an etching performance data set.A random forest algorithm was used to analyze and process the acquired chemical composition data and etching performance data at different concentrations. The random forest algorithm is a powerful machine learning algorithm capable of processing data with nonlinear relationships. By inputting chemical composition ratios and etching rate data, the random forest algorithm can establish an etching performance prediction model. Chemical etching recipes are screened by optimizing the input variables. For example, chemical data with different composition ratios are input into the prediction model, and the model outputs corresponding etching rate predictions. If the predicted values ​​meet the set conditions, such as an etching rate greater than 6 μm / min with good uniformity, the chemical recipe is considered to have high etching selectivity. When the output variables obtained from the input recipe data set of the etching performance prediction model meet the set conditions, a chemical recipe with high etching selectivity is obtained, and the screening process ends. For example, after multiple simulations and optimizations, it was found that a chemical recipe with a volume ratio of HF to HNO₃ of 3:1 exhibited the best etching performance in the prediction model, meeting the set conditions and ultimately determining this recipe as having high etching selectivity. This method not only efficiently screens the optimal solution formulation, but also provides a deeper understanding of the impact of different component ratios on etching performance, guiding process optimization in actual production. The application of this method can significantly improve etching process efficiency and product quality, and has broad application prospects.

[0028] S104. Based on the microfluidic chip screening results, a spray head array and stirring device are designed to create a uniform spray and turbulent flow field for a solution with high etching selectivity. By optimizing the nozzle shape, number of nozzles, spray angle, and stirring blade parameters, the mass transfer process between the solution and the material surface is enhanced, improving etching uniformity. The solution recovery device uses dual temperature and conductivity control parameters to adjust the recovery time in real time according to changes in the solution properties, reducing solution waste.

[0029] S104 also includes: constructing a three-dimensional model of the channel according to the channel data of the microfluidic chip, and gridding the three-dimensional model; obtaining the conductivity at different positions inside the flow channel, if there are differences in the conductivity curves at different positions, calculating the standard deviation between the values ​​at different positions, if the standard deviation is less than a preset threshold, calculating the average value of the slope of the grayscale value change of each column of pixels, obtaining the impeller rotation angle control value, and adjusting the impeller angle according to the control value; obtaining the adjusted spray parameters of the spray array, if the value of the spray flow curve change range is lower than the preset value, obtaining the flow value, conductivity value and temperature value at this time, calculating and digitizing the turbulent field streamline curvature information, and judging whether the flow value, conductivity value, temperature value and turbulent field streamline curvature value meet the preset requirements.

[0030] For example, based on the channel data of a microfluidic chip, a 3D model of the corresponding channel dimensions is constructed using COMSOL Multiphysics software and meshed. For example, assuming the channel width of the microfluidic chip is 200 microns, the depth is 50 microns, and the length is 1 centimeter, a 3D geometric model is first created in COMSOL, and these dimensional parameters are precisely set. The model is then meshed, ensuring a moderate mesh density to ensure both computational accuracy and efficiency. The velocity of the fluid within the 3D model is obtained, and a 3D velocity field map is generated. The distance h between the showerhead outlet and the material surface and the characteristic length L of the material surface are calculated to obtain the Reynolds number Re. Assuming the fluid is water with a density ρ of 1000 kg / m³, a characteristic velocity v of 0.1 m / s, and a dynamic viscosity μ of 0.001 Pa·s, the Reynolds number Re = ρvL / μ = 1000 × 0.1 × 0.01 / 0.001 = 1000. Based on the calculated Reynolds number, the Reynolds number range is determined and an appropriate turbulence model is selected. If the Reynolds number is less than 2300, the laminar flow model is selected; if it is between 2300 and 4000, the transitional flow model is selected; if it is greater than 4000, the turbulent flow model is selected. The three-dimensional model is considered to meet the standard when the blade speed parameters of the stirring device in the COMSOL simulation match the actual blade speed parameters and the turbulent field distribution data meets expectations. For example, if the blade speed is set to 1000 rpm and the simulation results show a uniform turbulent field distribution and the flow velocity distribution meets expectations, the model is considered to meet the standard. The chip's physical material properties are then assigned to the three-dimensional model based on the channel material parameters of the microfluidic chip. Assuming the chip material is polydimethylsiloxane (PDMS), its physical properties, such as elastic modulus and Poisson's ratio, are input into the model. After simulation, flow rate, conductivity, and pressure values ​​at different locations are obtained, which serve as data labels for the deep learning algorithm. After the preferred solution is sprayed into the flow channel, images of the spraying process are captured at preset time intervals to generate spray images at different moments. For example, an image is collected every 1 second for 10 minutes to obtain 600 frames of images. The continuous frame images obtained are analyzed and processed, and each frame of the image is split into sub-images by column. Each column of pixels is defined as a sub-image. The grayscale value distribution information of these sub-images is used as data features and is input into the deep learning neural network for training. If the real-time data collected at multiple points inside the flow channel differs from the labeled data by more than 5% after being output by the algorithm, it means that the data does not meet the requirements. After failing to meet the requirements, the convolutional neural network is used to analyze the change in the average pixel value of each column of sub-images and compare it with the images already in the database. If the similarity is less than the set value of 65%, it is judged that a new working condition has occurred, the liquid composition ratio is adjusted, and a new liquid is re-optimized. At the same time, the optimization algorithm is used to change the parameters of the sprinkler array to generate spray images with different spray angles.For example, if the current solution consists of a 3:1 volume ratio of HF and HNO₃ and the similarity is only 60%, the ratio is adjusted to 4:1 and the experiment is repeated. Data is collected for the new operating conditions and entered into the database as a dataset to expand the dataset. If the similarity is greater than 65%, the sprinkler head's spray parameters, such as spray angle and spray pressure, are adjusted. Based on the real-time temperature measurements at different locations within the flow channel, if the variance between the two temperature measurements exceeds 20%, the sprinkler head's outlet aperture is changed. For example, if the first temperature measurement is 25°C and the second is 35°C, with a variance of 100, the outlet aperture is adjusted from 0.5 mm to 0.6 mm. Based on the real-time conductivity measurements at different locations within the flow channel, if there are differences in the conductivity curves at different locations, the standard deviation of the values ​​at different locations is calculated. If the calculated standard deviation is less than a preset threshold, the impeller rotation angle control value is calculated by averaging the slope of the grayscale value change of each column of pixels. This control value is then used to control the impeller angle. For example, if the standard deviation is 0.02, which is less than the preset threshold of 0.05, the average grayscale value slope is calculated to be 0.1, and the impeller rotation angle is controlled to 10 degrees, so that the streamline curvature of the turbulent field formed by the blades is consistent with the surface undulations of the material. The adjusted spray array spray parameters are obtained. When the spray flow curve variation range is less than 1%, the flow value, conductivity value, and temperature value at that time are obtained. At the same time, the turbulent field streamline curvature information obtained in real time is calculated and digitized to determine whether the flow rate, conductivity, temperature, and turbulent field streamline curvature values ​​meet the requirements. For example, a flow rate variation range of 0.5%, a conductivity of 1.2 Siemens / meter, a temperature of 30°C, and a streamline curvature of 0.15 all meet the preset requirements. The difference between the conductivity, temperature, and other parameter values ​​at the end and start of the spray is calculated, and the resulting data set is used to characterize the properties of the liquid at that time. Assuming that the conductivity is 1.0 Siemens / meter at the beginning of the spray and 1.2 Siemens / meter at the end, with a difference of 0.2, the solution information update value is obtained in combination with the temperature change, and the updated value is used to adjust the control parameters of the liquid medicine recovery device. According to the feedback information of the recovery effect of the liquid medicine recovery device after the parameters are changed, the relationship model of temperature, conductivity and liquid medicine property parameters is corrected to achieve the optimized operation of the liquid medicine recovery system. Through the above method, not only can the fluid behavior in the microfluidic chip be accurately controlled, but the spray parameters and liquid medicine composition can also be adjusted in real time to ensure the stability and efficiency of the etching process. This comprehensive use of simulation, image processing and machine learning techniques has significantly improved the application effect of microfluidic chips in the etching process and provided strong support for high-precision manufacturing.

[0031] S105. Install a multi-parameter online monitoring unit in the etching equipment. Use a fiber optic spectrometer to measure the concentration of the chemical solution in real time, and use a thermocouple and pH electrode to measure the temperature and pH of the chemical solution in real time. When the monitored parameters exceed the set range, the chemical solution ratio adjustment program is activated. By controlling the liquid inlet and outlet volumes, the chemical solution properties are dynamically adjusted to ensure the stability and consistency of the etching process.

[0032] S105 also includes: obtaining real-time spectral data in the liquid medicine pool, determining the liquid medicine concentration data based on a pre-established spectral database, and if the difference between the measured liquid medicine concentration and the target concentration is greater than a preset threshold, calculating the concentration difference, starting the liquid medicine concentration adjustment algorithm, and generating control data for controlling the liquid inlet and discharge; establishing a recurrent neural network model for different proportions of multi-component liquid medicines, inputting data including liquid medicine concentration, temperature, pH, liquid inlet volume and liquid discharge volume at historical moments, and a pre-established liquid medicine ratio concentration curve, predicting the process parameter values ​​at future moments, and obtaining the liquid inlet valve and liquid discharge valve opening adjustment strategy; if the error between the predicted result and the target set value is greater than the preset threshold, performing model adaptation through an artificial neural network, collecting real-time data to fit the historical scheme for training, obtaining model adjustment data, and determining new process parameter adjustment weights.

[0033] For example, a fiber optic spectrometer collects real-time spectral data from a drug solution pool to accurately monitor the real-time changes in the drug solution's composition. For example, assuming the primary components of the drug solution pool are sulfuric acid and hydrochloric acid, the fiber optic spectrometer can capture the characteristic spectra of these two acids in real time. Using a pre-established spectral database, these characteristic spectra can be compared with standard spectra. Using chemometric methods such as partial least squares (PLS), the drug solution concentration data for the corresponding components within the real-time spectral data can be determined. Assuming the target sulfuric acid concentration is 5000 mol / L and the measured concentration is 4999 mol / L, the difference is 1 mol / L, which is greater than the preset threshold of 0.001 mol / L. At this point, the system activates the drug solution concentration adjustment algorithm and generates control data for controlling the inlet and outlet flow. This control data is calculated from the concentration difference data and the real-time inlet flow rate, and the size of the inlet and outlet valves is then adjusted to ensure that the drug solution concentration quickly returns to the target value. The size adjustment of the inlet and outlet valves depends not only on the concentration difference but also on real-time drug solution temperature data and historical data. For example, when constructing the thermodynamic equation for real-time liquid inflow rate, a Fourier series model was used for fitting, with the number of series set to 50. Assuming a target temperature of 25°C, the fitting coefficients of the Fourier series equation will change at this temperature. By adjusting these coefficients, the inflow rate at that specific temperature can be precisely determined. This method effectively addresses the impact of temperature changes on the fluidity of the drug solution and ensures precise control of the inflow rate. For pH control, the target pH value and temperature data are key parameters. Assuming a target pH of 8.00, a measured pH of 7.00, a temperature of 20°C, and a concentration of 5000 mol / L, the pH difference is 1.00. Using the integral control method for step-by-step inflow, the inflow time can be precisely determined, achieving an inflow control accuracy of 0.001 mL. This method gradually approaches the target pH value through small step adjustments, avoiding fluctuations caused by large adjustments and improving control stability. The proportioning model uses a support vector machine regression algorithm, with inputs of historical concentration, temperature, and pH data. Assuming the historical database contains 1,000 sets of experimental measurement data, the preset regression model generated from this data can output the inlet and outlet volumes of different chemical solutions. For example, based on the current state of the chemical solution pool, the model predicts that 100 mL of sulfuric acid and 50 mL of hydrochloric acid need to be added to ensure that the chemical solution composition reaches the preset ratio. This method utilizes a large amount of historical data, improves the accuracy of the prediction, and reduces the trial and error costs in the experiment. The use of recurrent neural networks to model process parameter time series data is to predict future trends in process parameter changes. Assuming the input includes chemical solution concentration, temperature, pH, inlet volume, and outlet volume data from the past 100 moments, the pre-established chemical solution ratio concentration curve can help predict the process parameter values ​​for the next 30 moments.For example, based on current data, the model predicts that the concentration of the drug solution will increase by 0.5% within the next 10 minutes. The openings of the inlet and outlet valves need to be adjusted accordingly to maintain stable drug solution composition. This approach uses time series analysis to predict process parameter changes in advance, achieving proactive control. If the recurrent neural network's predictions show that any of the three parameters, drug solution concentration, temperature, and pH, differ from the target setpoint by more than 5%, the system will use an artificial neural network to learn and adapt the model. For example, the system collects the previous 10,000 sets of real-time data on inlet speed, inlet volume, and pH values ​​from the production line, and trains the model using historical inlet plans to generate model adjustment data. Suppose the training results indicate that the inlet speed needs to be increased by 10% and the pH value needs to be adjusted by 0.2 units. These adjustment weights are then input into the inlet controller to ensure the new process parameters are more accurate. Based on the adjusted weights, the weight coefficients within the control logic judgment algorithm are updated. If the new weight coefficients differ from the original weight coefficients by more than 1, the system will conduct real-time valve pressure monitoring to ensure stable valve pressure. For example, pressure sensors monitor the pressure of the inlet and outlet valves in real time. If pressure fluctuations exceed a preset range, the system immediately adjusts the valve opening to ensure stable flow of the liquid medicine. This approach further improves control reliability and stability through real-time monitoring and adjustment. This series of measures not only achieves precise control of the liquid medicine composition but also improves the stability and efficiency of the entire process. Real-time monitoring by a fiber optic spectrometer, precise chemometric analysis, fitting of thermodynamic equations, application of a support vector machine regression algorithm, prediction using a recurrent neural network, and adaptive adjustment using an artificial neural network together form a multi-level, multi-dimensional control system that ensures the stability of the liquid medicine composition and the accuracy of process parameters, providing strong support for high-precision manufacturing.

[0034] S106. Characterize and evaluate the etching results after dynamic control. Using etching depth, surface roughness, and etching residue as evaluation indicators, optical microscopy and atomic force microscopy are used to characterize the etching results. By calculating the uniformity index and selectivity index, the etching performance of different material systems and solution formulations is quantitatively evaluated. The uniformity index calculation formula is the ratio of the standard deviation of the etching depth to the average value, and the selectivity index calculation formula is the ratio of the etching rates of the two materials. For the performance evaluation results, gray correlation analysis is used to determine the key process parameters to guide the learning and iterative update of the intelligent optimization system.

[0035] S106 also includes: obtaining three-dimensional topographic data of the sample after etching, scanning the sample surface using an optical microscope and an atomic force microscope respectively to obtain a height data matrix. Based on the height data matrix, the etching depth, surface roughness, and etching residue evaluation indicators are calculated. The etching depth is determined by the height difference between the reference plane and the etched surface, the surface roughness is characterized by the arithmetic average roughness (Ra), and the etching residue is obtained by setting a threshold to determine the degree of dispersion of the height data. The etching rate of samples of different material systems is measured. The solution formula is used as input and the etching rate as output, and a mapping relationship between the solution formula and the etching rate is established. When the solution formula changes, the corresponding etching rate can be quickly predicted. The uniformity index and selectivity index are calculated. The uniformity index is obtained by the ratio of the standard deviation of the etching depth to the mean value, reflecting the uniformity of the etching depth. The selectivity index is obtained by the ratio of the etching rates of two materials, reflecting the difference in the etching ability of the solution formula for different materials. Grey correlation analysis is used to determine the key process parameters affecting etching performance. The uniformity index and selectivity index are used as parent factors, and the solution formula and process parameters are used as sub-factors. The grey correlation between each sub-factor and the parent factor is calculated. The greater the correlation, the greater the impact of the factor on etching performance. The results of the grey correlation analysis are used to guide the learning of the intelligent optimization system. With the uniformity index and selectivity index as the optimization targets, and the solution formula and key process parameters as the optimization variables, a genetic algorithm is used to search for the optimal solution. After the intelligent optimization system completes an iteration, the experimental results are introduced into the sample set and the knowledge base of the intelligent optimization system is updated. The above steps are repeated continuously to achieve autonomous learning and iterative updates of the intelligent optimization system. As the optimization iterations proceed, the intelligent optimization system is able to summarize patterns from a large amount of experimental data, continuously improving prediction accuracy and optimization efficiency, and ultimately obtaining the optimal solution formula and process parameter combination that takes into account both uniformity and selectivity.

[0036] For example, obtaining three-dimensional morphological data of the sample after etching is an important part of the etching process evaluation. First, the sample surface is scanned separately using an optical microscope and an atomic force microscope. The optical microscope can quickly obtain the surface morphology of a large range through the principle of optical imaging, while the atomic force microscope provides more accurate three-dimensional height data through the interaction between the probe and the sample surface. For example, the height data matrix obtained from a certain scan shows that the highest point of a certain area on the sample surface is 5 microns and the lowest point is 1 micron, forming a three-dimensional morphology map with a height difference of 4 microns. Based on the height data matrix, the etching depth, surface roughness and etching residue evaluation indicators can be calculated. The etching depth is determined by the height difference between the reference plane and the etching surface. Assuming that the height of the reference plane is 2 microns and the height of the lowest point of the etching surface is 1 micron, the etching depth is 1 micron. Surface roughness is characterized by the arithmetic mean roughness (Ra). Assuming the heights of each point in the height data matrix are 2.1, 2.2, 2.0, and 2.3 microns, respectively, Ra is calculated as (2.1 + 2.2 + 2.0 + 2.3) / 4 = 2.15 microns. Etching residue is determined by setting a threshold to determine the degree of dispersion in the height data. For example, if the threshold is set to 0.5 microns, if more than 10% of the points in the height data differ from the average height by more than 0.5 microns, etching residue is determined to be present. The etching rates of samples from different material systems are measured. For example, a certain chemical formula has an etching rate of 10 microns / minute for silicon and 5 microns / minute for silicon oxide. Using the chemical formula as input and the etching rate as output, a mapping relationship between the chemical formula and the etching rate is established. When the chemical formula changes, such as adjusting the concentration of a certain chemical component, the corresponding etching rate can be quickly predicted using this mapping relationship. Calculating the uniformity index and selectivity index is an important indicator for evaluating etching performance. The uniformity index, calculated as the ratio of the standard deviation to the mean of the etching depth, reflects the uniformity of the etching depth. Assuming that in an experiment, the mean etching depth is 3 microns and the standard deviation is 0.5 microns, the uniformity index is 0.5 / 3, which is approximately 0.167. The selectivity index, calculated as the ratio of the etching rates of two materials, reflects the difference in the etching performance of the chemical solution for different materials. For example, if the etching rates of silicon and silicon oxide are 10 microns / minute and 5 microns / minute, respectively, the selectivity index is 10 / 5 = 2. Grey correlation analysis is used to identify the key process parameters affecting etching performance. The uniformity index and selectivity index are used as parent factors, and the concentrations of the ingredients in the chemical solution formula, temperature, and pH value are used as subfactors. The grey correlation between each subfactor and the parent factor is calculated. For example, the grey correlation index of hydrofluoric acid concentration is 0.8, significantly higher than that of other factors, indicating that hydrofluoric acid concentration has a significant impact on etching performance. Using the uniformity index and selectivity index as optimization targets, a genetic algorithm is used for intelligent optimization. The initial population contains a variety of drug solution formulas and process parameter combinations, which are continuously iterated and optimized through operations such as crossover and mutation.For example, after one iteration, an optimal parameter combination was obtained: 10% hydrofluoric acid concentration, 30°C temperature, and pH 4, corresponding to a uniformity index of 0.1 and a selectivity index of 3. The experimental results were introduced into the sample set to update the intelligent optimization system's knowledge base. For example, the newly introduced experimental data showed that at a hydrofluoric acid concentration of 12%, the uniformity index significantly improved to 0.08. The system adjusted the optimization strategy accordingly, increasing the weight of the high-concentration hydrofluoric acid formula. As the optimization iterations progressed, the intelligent optimization system summarized patterns from the large amount of experimental data, continuously improving prediction accuracy and optimization efficiency. For example, after 100 iterations, the error between the system's predicted etching rate and the actual value decreased from 5% to 1%, significantly improving the optimization effect of process parameters. Through the above steps, the intelligent optimization system can continuously learn and optimize, ultimately obtaining the optimal solution formula and process parameter combination that balances uniformity and selectivity, significantly improving the stability and efficiency of the etching process.

[0037] S107. Based on process optimization experimental data, a nonlinear mapping model is constructed between the chemical solution formula, process parameters, and etching results. This model uses a multi-layer neural network structure, with the chemical solution formula and process parameters as input and the etching depth, uniformity, and selectivity indicators as output. The obtained etching patterns and control strategies are transformed into dynamically updated knowledge-based reasoning rules. An intelligent decision-making system based on case-based reasoning and fuzzy logic is developed to achieve automation and intelligentization of chemical solution formula design, process parameter selection, and etching defect diagnosis.

[0038] The S107 further includes: constructing a nonlinear mapping model between the chemical solution formula, process parameters, and etching results, wherein the chemical solution formula and process parameters are used as input layers, and the etching depth, uniformity, and selectivity indicators are used as output layers. For new chemical solution formula and process parameter combinations, the most similar formula design and parameter selection scheme are retrieved from historical optimization cases using case-based reasoning technology. An intelligent decision-making system is developed to achieve automation and intelligence in chemical solution formula design, process parameter selection, and etching defect diagnosis by combining case-based reasoning and fuzzy logic.

[0039] For example, based on process optimization experimental data, a multi-layer neural network was used to construct a nonlinear mapping model between the chemical solution formula, process parameters, and etching results. Specifically, the input layer includes the concentrations of each component in the chemical solution formula (such as hydrofluoric acid and nitric acid) and process parameters (such as temperature, pH value, etching time, etc.), while the output layer represents key indicators such as etching depth, uniformity index, and selectivity index. For example, in one experiment, the input layer data included a hydrofluoric acid concentration of 10%, a nitric acid concentration of 5%, a temperature of 30°C, a pH value of 4, and an etching time of 30 minutes. The output layer results were an etching depth of 3 microns, a uniformity index of 0.1, and a selectivity index of 2. By training and optimizing the constructed nonlinear mapping model, the etching patterns and control strategies between the chemical solution formula, process parameters, and etching results were obtained. During the training process, a large amount of historical experimental data was used, and the network weights were continuously adjusted through a backpropagation algorithm to ensure that the model's predicted results were as close as possible to the actual experimental data. For example, after 1,000 training iterations, the error between the model's predicted etching depth and the actual value was reduced from 10% to 2%, significantly improving prediction accuracy. The trained model is converted into dynamically updated knowledge inference rules. For example, the model discovered a positive correlation between hydrofluoric acid concentration and etching depth, while increasing temperature leads to a decrease in the uniformity index. These patterns are converted into rules, such as "For every 1% increase in hydrofluoric acid concentration, the etching depth increases by 0.2 microns" and "For every 5°C increase in temperature, the uniformity index decreases by 0.05." Based on the acquired knowledge inference rules, case-based reasoning techniques are used to perform similarity matching and search for new chemical solution formulas and process parameter combinations based on historical optimization cases. For example, a newly proposed chemical solution formula with a hydrofluoric acid concentration of 12% and a nitric acid concentration of 6% is highly similar to a successful case, where the process parameters of that case, temperature of 28°C and pH of 3.5, ultimately produced excellent etching results. Based on this, the system recommends similar parameter combinations, predicting excellent etching results. To address the uncertainty and ambiguity in chemical solution formula design and process parameter selection, fuzzy logic theory is introduced to establish a fuzzy rule base. For example, the fuzzy rule base contains the rule "If the hydrofluoric acid concentration is high and the temperature is moderate, the etching depth will be greater and the uniformity will be better." During specific operations, parameters such as the hydrofluoric acid concentration and temperature are fuzzy processed. For example, "high hydrofluoric acid concentration" is defined as a concentration greater than 10%, and "moderate temperature" is defined as 25°C to 35°C. Through fuzzy reasoning, the system can give reasonable formula design and parameter selection suggestions when the parameters are not completely determined. Combining case reasoning and fuzzy logic, an intelligent decision-making system is developed to automate and intelligently process the design of chemical solution formulas, the selection of process parameters, and the diagnosis of etching defects. For example, after receiving a new etching task, the system first finds similar historical cases through case reasoning, and then uses fuzzy logic to fine-tune the parameters in the case to generate the optimal combination of chemical solution formulas and process parameters.The system also predicts potential etching defects, such as uneven etching and excessive residue, based on historical data and provides corresponding solutions. A human-computer interface within the intelligent decision-making system allows process experts to evaluate and modify the system-generated recipe design and parameter selection schemes. For example, an expert discovered that the system-recommended hydrofluoric acid concentration was too high and could cause equipment corrosion. They adjusted the concentration to 9% and provided feedback to the system. The system then updated its knowledge base and prioritized the adjusted concentration in similar situations. Through continuous iterative optimization, the intelligent decision-making system automatically generates the optimal solution formula and process parameter combination based on process requirements and target performance. For example, after one iteration, the system generated a recipe with a hydrofluoric acid concentration of 11%, a nitric acid concentration of 5.5%, a temperature of 29°C, and a pH of 3.8, achieving an etching depth of 3.5 microns, a uniformity index of 0.08, and a selectivity index of 2.5, significantly improving the stability and efficiency of the etching process. The application of this intelligent decision-making system not only improves decision-making efficiency and accuracy but also reduces the number of experiments and costs. Through the optimization solutions generated by the system, process engineers can quickly find the optimal process parameters, avoiding extensive trial-and-error experiments. Furthermore, the system's continuous learning and updating make the process optimization process more intelligent and efficient, providing strong support for the continuous improvement of etching processes.

[0040] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent wafer etching process, characterized in that: The process includes the following steps: S101, obtaining the etching rates of different materials under various chemical solution formulations, building an etching rate database covering multiple material systems, and screening out material combinations with high etching selectivity by comparing the etching rate differences in the database to determine the targeted optimization direction of the chemical solution formulation; S102: For the selected material combinations, first-principles calculations are used to study the interaction mechanism between each component in the solution and the material surface, obtain the kinetic parameters of the surface reaction, and perform molecular dynamics simulations to reveal the microscopic mechanism of the solution diffusion and mass transfer process. A multi-scale coupled etching process model is established, including a molecular-scale surface reaction model, a mesoscopic-scale diffusion and mass transfer model, and a macroscopic-scale etching morphology evolution model. Through the coupling of these models, the rate-controlling steps and chemical reaction pathways of material removal are clarified, guiding the optimization of the solution formulation. S103 uses microfluidic chip technology to design and integrate multiple parallel etching reaction chambers to achieve high-throughput screening of drug solution formulas. By controlling the flow rate of the sampling pump and the drug solution ratio, a concentration gradient is formed within the chip. Spectroscopy is used to monitor the drug solution composition at the chip outlet in real time to obtain etching rate data for different formulations. Machine learning algorithms are used to establish a quantitative relationship between formulation and etching performance, resulting in a drug solution formula with high etching selectivity. S104. Based on the microfluidic chip screening results, a spray head array and stirring device are designed for a solution formulation with high etching selectivity to form a uniform solution spray and turbulent flow field. By optimizing the nozzle shape, number of nozzles, spray angle, and stirring blade parameters, the mass transfer process between the solution and the material surface is enhanced, improving etching uniformity. The solution recovery device uses dual parameter control of temperature and conductivity to adjust the recovery time in real time according to changes in the solution properties, reducing solution waste. S105: Install a multi-parameter online monitoring unit in the etching equipment. Use a fiber optic spectrometer to measure the concentration of the chemical solution in real time. Use a thermocouple and pH electrode to measure the temperature and pH of the chemical solution in real time. When the monitored parameters exceed the set range, the chemical solution ratio adjustment program is activated. By controlling the liquid inlet and outlet volumes, the chemical solution properties are dynamically regulated to ensure the stability and consistency of the etching process. S106: Characterize and evaluate the etching results after dynamic control, using etching depth, surface roughness, and etching residue as evaluation indicators. Optical microscopy and atomic force microscopy are used to characterize the etching results. The etching performance of different material systems and solution formulations is quantitatively evaluated by calculating the uniformity index and selectivity index. The uniformity index is calculated as the ratio of the standard deviation of the etching depth to the average value, and the selectivity index is calculated as the ratio of the etching rates of the two materials. For the performance evaluation results, gray correlation analysis is used to determine key process parameters to guide the learning and iterative update of the intelligent optimization system. S107, based on process optimization experimental data, constructs a nonlinear mapping model between solution formula, process parameters and etching results. The model adopts a multi-layer neural network structure, with the input layer being the solution formula and process parameters, and the output layer being the etching depth, uniformity and selectivity indicators. The obtained etching laws and control strategies are transformed into dynamically updated knowledge reasoning rules, and an intelligent decision-making system based on case reasoning and fuzzy logic is developed to realize the automation and intelligence of solution formula design, process parameter selection and etching defect diagnosis.

2. The intelligent wafer etching process according to claim 1, characterized in that: The S101 further includes: Acquire etching rate data covering a preset number of types of material systems and chemical solution formulations, and establish the etching rate database; Using a data analysis algorithm to analyze the etching rate data in the etching rate database, and calculating the etching rate differences of different material combinations under the same solution formula; Determining the selectivity of the material combination according to the etching rate difference, and screening out a material combination with high etching selectivity; For the material combination with high etching selectivity, a machine learning algorithm is used to analyze the relationship between the material combination and the chemical solution formulation parameters to obtain the key chemical solution formulation parameters that affect the etching selectivity; Determining a direction for optimizing a drug solution formulation based on the key drug solution formulation parameters, and optimizing the drug solution formulation by adjusting the value ranges of the key drug solution formulation parameters; The optimized solution formula is applied to the material combination with high etching selectivity, and the optimization effect is verified through etching experiments. If the preset optimization target is not achieved, the process returns to the solution formula optimization step and continues to optimize until the optimal solution formula is obtained to achieve high selective etching of the material.

3. The intelligent wafer etching process according to claim 1, characterized in that: The S103 further includes: Obtaining the number of multiple parallel etching reaction chamber structures formed in the chip, and configuring the initial component ratios of the corresponding number of chemical solutions at the entrances of different reaction chambers according to the number; Spectroscopy is used to detect a preset number of different types of liquid medicine outputted from different reaction chambers at the chip outlet, and it is determined whether the liquid medicine composition of a certain reaction chamber among the detected liquid medicines meets the preset spectral characteristic peak threshold range. If so, the position data of the etching reaction chamber corresponding to the liquid medicine on the chip is obtained; According to the position data, obtaining the drug solution concentration information and flow rate control information of the corresponding injection pump, obtaining a preset number of drug solution composition information data of different types and concentrations, and associating the drug solution composition information data with the corresponding numbers of the multiple reaction chambers on the chip; According to the chemical solutions with different component ratios corresponding to the reaction chambers at different positions on the same chip, the performance data of different etching rates corresponding to the chemical solutions with different component ratios are determined by the association to obtain an etching performance data set; A random forest algorithm is used to analyze and process the obtained chemical composition data of different concentrations and the etching performance data to establish an etching performance prediction model; The chemical solution etching recipe is screened by optimizing the input variables to determine whether the output variables obtained from the input recipe data set of the etching performance prediction model meet the preset conditions. If so, the corresponding chemical solution recipe is determined to be a chemical solution recipe with high etching selectivity.

4. The intelligent wafer etching process according to claim 1, characterized in that: The S104 further includes: constructing a three-dimensional model of the channel according to the channel data of the microfluidic chip, and performing grid processing on the three-dimensional model; Obtaining the conductivity at different locations within the flow channel. If the conductivity curves at different locations differ, calculating the standard deviation between the values ​​at different locations. If the standard deviation is less than a preset threshold, calculating the average slope of the grayscale value change of each column of pixels to obtain the impeller rotation angle control value, and adjusting the impeller angle based on the control value. Obtain the adjusted spray parameters of the spray array. If the spray flow curve variation range is lower than the preset value, obtain the flow value, conductivity value, and temperature value at this time, calculate and digitize the turbulent field streamline curvature information, and determine whether the flow value, conductivity value, temperature value, and turbulent field streamline curvature value meet the preset requirements.

5. The intelligent wafer etching process according to any one of claims 1 to 4, characterized in that: The S105 further includes: Acquire real-time spectral data from the drug solution pool and determine drug solution concentration data based on a pre-established spectral database. If the difference between the measured drug solution concentration and the target concentration is greater than a preset threshold, calculate the concentration difference, activate the drug solution concentration adjustment algorithm, and generate control data for controlling liquid inlet and outlet; A recurrent neural network model was established for different ratios of multi-component liquids. The input data included historical liquid concentration, temperature, pH, inlet and outlet volumes, as well as a pre-established liquid ratio concentration curve. The model predicted the process parameter values ​​at future times and derived the inlet and outlet valve opening adjustment strategy. If the error between the predicted result and the target set value is greater than the preset threshold, the model is adapted through the artificial neural network, real-time data is collected to fit the historical plan for training, model adjustment data is obtained, and new process parameter adjustment weights are determined.

6. The intelligent wafer etching process according to any one of claims 1 to 4, characterized in that: The S106 further includes: Obtain the three-dimensional morphology data of the sample after etching, and use optical microscope and atomic force microscope to scan the sample surface respectively to obtain the height data matrix; Calculate the etching depth, surface roughness and etching residue evaluation indicators based on the height data matrix; The etching depth is determined by the height difference between the reference surface and the etched surface, the surface roughness is characterized by the arithmetic average roughness Ra, and the etching residue is obtained by setting a threshold to judge the discrete degree of the height data; Measure the etching rate of samples of different material systems; Taking the chemical solution formula as input and the etching rate as output, a mapping relationship between the chemical solution formula and the etching rate is established; When the solution formula changes, the corresponding etching rate can be quickly predicted; Calculate uniformity index and selectivity index; The uniformity index is obtained by the ratio of the standard deviation of the etching depth to the average value, which reflects the uniformity of the etching depth; The selectivity index is obtained by the ratio of the etching rates of two materials, reflecting the difference in the etching ability of the solution formula on different materials; Use grey correlation analysis to determine the key process parameters that affect etching performance; The uniformity index and selectivity index are taken as parent factors, and the solution formula and process parameters are taken as sub-factors. The grey correlation between each sub-factor and the parent factor is calculated. The greater the correlation, the greater the influence of the factor on the etching performance. The results of grey relational analysis are used to guide the learning of intelligent optimization system; With uniformity index and selectivity index as optimization targets, and liquid formulation and key process parameters as optimization variables, a genetic algorithm was used to search for the optimal solution. When the intelligent optimization system completes an iteration, the experimental results are introduced into the sample set to update the knowledge base of the intelligent optimization system; Repeat the above steps continuously to achieve autonomous learning and iterative updates of the intelligent optimization system; As the optimization iterations proceed, the intelligent optimization system can summarize patterns from a large amount of experimental data, continuously improve prediction accuracy and optimization efficiency, and ultimately obtain the optimal solution formula and process parameter combination that takes into account both uniformity and selectivity.

Images