Wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm

Abstract

The invention discloses a wafer anti-sticking layer coating process optimization method based on fluid mechanics and a small sample prediction algorithm. Comprising the following steps: constructing a vacuum reaction cavity simulation geometric model including a vacuum cavity, a wafer carrier and a wafer; and according to the simulation geometric model, on the basis of a fluid mechanics simulation operation purging process, observing a discrete phase distribution condition, and selecting a model customization cavity with the best purging effect. A corresponding hardware system and a human-computer interaction platform are built, a test experiment is carried out, small sample learning is carried out according to obtained experiment data, and multi-mapping prediction is carried out through environmental data, sample data and other coating properties such as hydrophobicity and the like. According to the scheme, the wafer coating process with better purging and deposition effects and better effect in long-term use can be realized, and targeted process optimization is realized through a small sample prediction algorithm.

Description

Technical Field

[0001] The present invention belongs to the field of wafer anti-sticking layer coating, and specifically, designs a wafer anti-sticking layer optimization coating process based on fluid mechanics calculation simulation and small sample prediction algorithm. Background Art

[0002] The manufacturing process of semiconductor chips is mainly divided into eight steps: wafer processing, oxidation, photolithography, etching, thin film deposition, interconnection, testing, and packaging. Wafers are round thin slices formed by cutting single crystal cylinders made of silicon (Si) or gallium arsenide (GaAs). During the production and transportation process, the wafer substrate usually inevitably contains other chemical impurities, physical pollutants, adsorbed water vapor, etc. The presence of these substances will affect the photolithography process and the quality of the pattern. Therefore, after the wafer processing is completed, it is easier to complete the maximum degree of desorption and cleaning by coating it with an anti-sticking layer through thin film deposition technology.

[0003] The process of wafer anti-sticking coating is: chamber purging (multiple times), vacuuming, sample heating, sample deposition in a vacuum environment, chamber purging multiple times, and restoration of atmospheric pressure. The sample gas is deposited to form a single layer of film and covalently bonded to the processed surface of the wafer, thereby adding a single layer of anti-sticking coating to the wafer surface. However, the thin film deposition equipment currently used in China for wafer anti-sticking coating is basically imported, with high costs, and in the long-term use process, there will be incomplete initial purging and residual sample deposition, resulting in poor subsequent coating effects.

[0004] At the same time, the software supporting foreign equipment has relatively poor scalability and tends to simulate the deposition process; due to different structures, foreign simulation systems are not suitable for domestic equipment, and historical data are not universal. Summary of the invention

[0005] The purpose of the present invention is to design a thin film deposition device for wafer anti-sticking layer coating to achieve uniform deposition and complete purging coating process optimization.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A wafer anti-sticking layer coating purging simulation method based on fluid mechanics calculation, characterized by comprising:

[0008] Constructing a simulation geometric model of the vacuum reaction chamber, the geometric model including: a vacuum chamber, a wafer carrier and a wafer; and calculating a purge process based on a fluid mechanics simulation according to the simulation geometric model;

[0009] According to the simulation geometry model, the CFD calculation object is a low-speed incompressible two-phase flow. The Euler-Lagrange method is used to consider the mutual influence between the dispersed phase and the continuous phase. The continuous phase flow is first calculated to obtain a preliminary convergent flow field, and then the DPM model is turned on and iterated. After that, the dispersed phase trajectory and source terms are calculated alternately with the continuous phase flow field. Each time the continuous phase flow field is calculated, the momentum, mass and energy exchange between the two phases in the previous step are updated and included in the source term of the continuous phase.

[0010] Set the solver equations and calculate the simulation data;

[0011] Visualize the simulation data to obtain the discrete phase distribution.

[0012] Different materials and corresponding parameters are set for the modules of the simulation geometric model, four 6mm circular holes are drilled in the cavity wall, and the boundary conditions are set as follows: two holes are inlets and two holes are outlets; during the purge simulation, the initial velocity of one of the inlets is 0 (blocked), and the initial velocity of the other is set to 10L / s, one of the outlets is blocked, and the initial velocity of the other outlet is set to 0; particles with a molecular mass of 481.55g / mol and a relative density of 1.639 are uniformly distributed in the cavity, iterate 100 times, and start the calculation.

[0013] Fluid calculations use the SIMPLE algorithm based on pressure solutions, including:

[0014] The governing equation for the continuous phase is as follows:

[0015]

[0016]

[0017]

[0018]

[0019] Where: u, v, w are the velocities of the continuous phase in the x, y, and z directions, ρ is the density, f is the force on the continuous phase, μ is the viscosity, S m is the mass source term, V is the volume source term, t is the time, F is the additional force of the dispersed phase on the continuous phase, and x, y, and z represent the three directions respectively;

[0020] The fluid phase is regarded as a continuous medium and the dispersed phase is regarded as a discrete medium. The Euler-Lagrangian method is used.

[0021] The turbulent kinetic energy equation and diffusion equation of the model are expressed as follows:

[0022]

[0023]

[0024] Where: k is the turbulent kinetic energy, ε is the dissipation coefficient, G k represents the turbulent kinetic energy generated by the mean velocity gradient, G b represents the turbulent kinetic energy generated by the buoyancy effect, Y M represents the effect of compressible turbulent kinetic energy expansion on the total dissipation rate, C 1ε , C 2ε and C 3ε is a constant, σ k and σ ε is the turbulent Prandtl number.

[0025] A wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm, characterized by comprising:

[0026] Step 1, using the simulation method described in any one of claims 1 to 3 to optimize and improve the purging condition of the vacuum chamber (reactor) for coating;

[0027] Step 2: Customize the cavity according to the optimized model in step 1, punch two holes on the top and bottom of the cavity, punch one hole in the center of the cavity wall, the top hole is connected to the nitrogen cylinder and the sample gas path, and the bottom hole is connected to the exhaust gas box and the vacuum pump;

[0028] Step 3: Design the overall system of the equipment, including designing the gas circuit part of the lower computer electronic control system, designing the lower computer and the heating platform part of the electronic control system, and designing and programming the upper computer, building a complete electronic control system and connecting it with the upper computer to complete the serial port protocol configuration to realize human-computer interaction;

[0029] Step 4: Test the overall system obtained in step 3 and record the test data;

[0030] Step 5: Build a network model for small sample learning based on the experimental data obtained in step 4, and establish a neural network model for multi-mapping prediction.

[0031] In step three, the purge gas of the gas path of the lower computer is nitrogen. The nitrogen purge gas path, sample gas path and exhaust gas path are opened and closed by relay-controlled solenoid valves, and the vacuum extraction and decompression gas paths are completed by relay-controlled vacuum pumps. The precise temperature control of the heating table is achieved through PID temperature control. The upper computer receives and stores the vacuum degree, voltage, frequency, and temperature sensor data uploaded by the lower computer, and sends instructions to the lower computer to control different gas paths, set temperatures, and time parameters.

[0032] There are two solutions for the small sample model. One is to use genetic algorithm and particle swarm algorithm to optimize BP neural network and RBF neural network respectively; the other is to split the traditional BP network to simplify the network structure. In this process, single-step prediction is used instead of multi-step prediction to reduce the dimension of the input vector.

[0033] In step five, environmental data, sample data, etc. are used as input parameters, and hydrophobicity and thermal conductivity are used as output parameters.

[0034] Use the least squares algorithm of small sample regression to compare the models of the two solutions, and choose the model that best suits your data set by comparing the two solutions.

[0035] Beneficial effects of the present invention: The chamber design will achieve better purging effect and deposition results in both long-term and short-term use: The vacuum chamber designed according to the simulation results can achieve a more uniform deposition effect and higher deposition efficiency in the short-term use process; the purging before and after deposition is cleaner, avoiding the deposition residues left due to incomplete purging that affect the subsequent deposition performance; and solving the problems of unstable particle size concentration and pipeline blockage caused by the continuous accumulation of deposition residues during long-term use. The small sample prediction network will be able to predict the deposition performance under the current environment and parameters in the absence of historical data, and achieve targeted optimization of the current process. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Figure 2 shows two chamber models loaded with wafers and wafer carriers.

[0037] Figure 2 It is a two-way coupled system used in CFD simulation.

[0038] Figure 3 This is the SIMPLE algorithm flow chart.

[0039] Figure 4 Designed for the overall system of the equipment.

[0040] Figure 5 It is the PID temperature control curve.

[0041] Figure 6 It is the upper computer interface.

[0042] Figure 7 Prediction process for small samples.

[0043] Figure 8 is the OLS least squares regression line. DETAILED DESCRIPTION

[0044] The present invention will be further described below in conjunction with the accompanying drawings and specific implementations:

[0045] Step 1: Construct a vacuum reaction chamber simulation geometry model. The geometry model includes: vacuum chamber, wafer carrier and wafer; it is modeled using ANSYS FLUENT software. Figure 1 For the constructed model.

[0046] Step 2: For the cavity built in step 1, the cavity is a cylinder, and four 6mm circular holes are punched in the cavity wall. The boundary conditions are set as follows: two holes are inlets and two holes are outlets. During the purge simulation, the initial velocity of one inlet is 0 (blocked), and the initial velocity of the other is set to 10L / s. One outlet is blocked, and the initial velocity of the other outlet is set to 0. The particles with a molecular mass of 481.55g / mol and a relative density of 1.639 are uniformly distributed in the cavity. Iterate 100 times and start the calculation. Change the cavity size parameters and placement method to observe the discrete phase trajectory.

[0047] The CFD calculation object in this step is a low-speed incompressible two-phase flow. The Euler-Lagrange method is used. The bidirectional coupling system used in the CFD simulation calculation is Figure 2 Consider the mutual influence between the dispersed phase and the continuous phase. First, calculate the continuous phase flow to obtain a preliminary converged flow field, then start the DPM model and continue to iterate, after which the dispersed phase trajectory and source terms will be calculated alternately with the continuous phase flow field. Each time the continuous phase flow field is calculated, the momentum, mass, and energy exchange between the two phases in the previous step are updated and included in the source terms of the continuous phase.

[0048] Fluid calculations use the SIMPLE algorithm based on pressure solution. The specific process of the SIMPLE algorithm is as follows: Figure 3 The continuous phase is clean air (nitrogen), and the velocity distribution in the box is calculated to solve the low-speed steady incompressible Reynolds average (NS) equation. The continuous phase flow field solves the low-speed incompressible Reynolds average Navier-Stokes equation, using the Simple method based on pressure solution; the finite volume method of the grid center format is used for discretization, and the second-order headwind format is used for spatial discretization; the standard k-ε model is selected as the turbulence model. The control equation of the continuous phase is as follows:

[0049]

[0050]

[0051]

[0052]

[0053] Where: u, v, w are the velocities of the continuous phase in the x, y, and z directions, ρ is the density, f is the force on the continuous phase, μ is the viscosity, S mis the mass source term, V is the volume source term, t is the time, F is the additional force of the dispersed phase on the continuous phase, and x, y, and z represent the three directions respectively.

[0054] Low-speed incompressible flow does not require the solution of the energy field, so the equations do not contain energy equations. Compared with the single-phase flow control equation, the continuity equation of the two-phase flow control equation has an additional mass source term of the dispersed phase acting on the continuous phase, and the momentum equation has an additional force F of the dispersed phase on the continuous phase.

[0055] The fluid phase is regarded as a continuous medium and the dispersed phase is regarded as a discrete medium. Therefore, the Euler-Lagrange method is used. The particle motion equation is written according to Newton's laws of motion. The motion equation is as follows:

[0056]

[0057] Where: f D (uu p ) represents the resistance per unit particle mass; u is the continuous phase velocity, u p is the particle phase velocity, ρ is the density of the continuous phase; ρ p is the particle density; represents the combined force of buoyancy and gravity of a particle per unit mass; F is the additional acceleration term of a particle per unit mass, including the additional force generated by the pressure gradient, Saffman lift, "virtual mass" force, Basset force, electrostatic force, Brownian force, Magnus lift, etc. p >0.1, so the virtual mass force and pressure gradient force have a more obvious effect on the particle, while other forces have little effect on the particle and can be ignored in the calculation; in the calculation, it is assumed that the particle is spherical and will not deform. The virtual mass force and pressure gradient force are expressed as follows [20-21]:

[0058]

[0059]

[0060] The coefficient f in the resistance term D The calculation formula is as follows:

[0061]

[0062] Where: μ is the viscosity coefficient of the continuous phase; d p is the particle diameter; Re p is the Re number of the relative particle, which can be expressed as follows:

[0063]

[0064] Where: C Dis the drag coefficient, which can be expressed as follows for spherical particles:

[0065]

[0066] Where: a1, a2, a3 are constants, assuming the relative Reynolds number Re p <0.1, a1=0, a2=24, a3=0. The turbulent kinetic energy equation and diffusion equation of the standard k-ε model are expressed as follows

[22] :

[0067]

[0068]

[0069] Where: k is the turbulent kinetic energy, ε is the dissipation coefficient, G k represents the turbulent kinetic energy generated by the mean velocity gradient, G b represents the turbulent kinetic energy generated by the buoyancy effect, Y M represents the effect of compressible turbulent kinetic energy expansion on the total dissipation rate, C 1ε , C 2ε and C 3ε is a constant, σ k and σ ε is the turbulent Prandtl number.

[0070] Step 3: The cavity with the best purge effect calculated in step 2 is Figure 1 (a) Diameter 400mm, depth 300mm. Customize the cavity according to the model, punch two 6mm holes on the top and bottom, and punch one 6mm hole in the center of the cavity wall. The top hole is connected to the nitrogen cylinder and the sample gas line, and the bottom hole is connected to the exhaust box and vacuum pump.

[0071] Step 4: Design the overall system of the equipment, Figure 4 It is the overall electronic control system and human-computer interaction system of the equipment. The gas circuit of the lower computer electronic control system is designed. The purge gas is nitrogen. The nitrogen purge gas circuit, sample gas circuit and exhaust gas circuit are opened and closed by relay-controlled solenoid valves. The vacuum pump is controlled by relays to complete the vacuum extraction and decompression gas circuits. The heating table of the lower computer and electronic control system is designed. The precise temperature control of the heating table is achieved through PID temperature control to meet the temperature requirements of heating and gasification of different samples. Figure 5 The PID temperature curve for parameter adjustment test shows that temperature control can be achieved accurately and stably. The host computer is designed to receive and store sensor data such as vacuum degree, voltage, frequency, temperature, etc. uploaded by the slave computer. The host computer sends instructions to the slave computer to control different gas circuits, set temperature, time and other parameters. Figure 6 The host computer interface designed and written.

[0072] Step 5: Build a complete electronic control system and connect it to the host computer to complete the serial port protocol configuration and realize human-computer interaction.

[0073] Step 6: Use the thin film deposition system built in step 5 to conduct wafer coating experiments and record experimental data. The experimental data includes: environmental data such as vacuum degree, temperature, humidity, etc.; performance indicators of deposition results such as roughness, adhesion, and film thickness.

[0074] Step 7: Build a network model for small sample learning based on the experimental data obtained in step 6, and use environmental data, sample data, etc. as input parameters and coating properties such as hydrophobicity and thermal conductivity as output parameters to establish a neural network model for multi-mapping prediction. Compare the two solutions and select the model that best suits your data set. Figure 7 Prediction process for small samples.

[0075] In this step, two schemes are designed for the small sample prediction algorithm:

[0076] Solution 1: Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) are used to optimize BP neural network and RBF neural network respectively. The core of genetic algorithm to improve BP neural network is to optimize the weights and thresholds in the learning process of BP neural network. Each gene in the genetic algorithm represents a network weight or threshold. These genes constitute individuals, and several individuals constitute a population. The competitiveness of these individuals is determined according to the advantages and disadvantages of a certain set of weights and thresholds. Then, through the genetic algorithm, the individuals with the strongest competitiveness are obtained, and the weights and thresholds suitable for the network can be obtained. Particle swarm optimization has excellent global optimization ability. It can be used to improve BP and RBF neural network models, so as to achieve the purpose of escaping from the local minimum trap and alleviate the local minimum caused by the small number of samples.

[0077] Solution 2: Split the traditional BP network to simplify the network structure. In this process, single-step prediction is used instead of multi-step prediction to reduce the dimension of the input vector. The simpler the model, the better it performs in the case of small samples.

[0078] Finally, it is compared with the least squares algorithm model which is most commonly used for small sample regression. Figure 8 .

[0079] First, use the data to get a point estimate It is a linear combination of y, which has linearity. In addition, after imposing some assumptions, its conditional expectation is an unbiased estimate of the coefficients, that is,

[0080]

[0081] Then, a series of indicators such as goodness of fit R2 = SSR / SST can be calculated to illustrate how well the fit is.

[0082] Finally, after obtaining the interval estimate, we can perform hypothesis testing on the pre-made assumptions about the coefficients. 2 I)(where σ is unknown), then It also follows the normal distribution accordingly, so its interval estimate can be obtained. However, obtaining its interval estimate is not our ultimate goal. Our ultimate goal is to test whether the assumption such as Rβ=r (where R is a J×K matrix) is valid. Unbiased estimate σ 2 , and satisfies and From this, the F statistic can be constructed for testing:

[0083]

[0084] In this way, J linear hypotheses about the coefficients can be jointly tested by simply writing the hypothesis in the form of Rβ=r. If only one hypothesis needs to be tested, that is, J=1, then Therefore, the statistic can be transformed into a t distribution and the t test can be performed directly.

[0085] Note: The present invention is not limited to the field of wafer coating. The present invention can also be used for coating of other small parts based on thermal evaporation, and has applications in the semiconductor field, mechanical field, etc. Any changes that can be thought of by technicians in the corresponding fields should fall within the protection scope of the present invention.

Claims

1. A simulation method for purging a wafer anti-sticking layer based on fluid mechanics calculation, characterized in that include: Constructing a simulation geometric model of the vacuum reaction chamber, the geometric model including: a vacuum chamber, a wafer carrier and a wafer; and calculating a purge process based on a fluid mechanics simulation according to the simulation geometric model; According to the simulation geometry model, the CFD calculation object is a low-speed incompressible two-phase flow. The Euler-Lagrange method is used to consider the mutual influence between the dispersed phase and the continuous phase. The continuous phase flow is first calculated to obtain a preliminary convergent flow field, and then the DPM model is turned on and iterated. After that, the dispersed phase trajectory and source terms are calculated alternately with the continuous phase flow field. Each time the continuous phase flow field is calculated, the momentum, mass and energy exchange between the two phases in the previous step are updated and included in the source term of the continuous phase. Set the solver equations and calculate the simulation data; Visualize the simulation data to obtain the discrete phase distribution.

2. The wafer anti-sticking layer coating purging simulation method based on fluid mechanics calculation according to claim 1, characterized in that: Different materials and corresponding parameters are set for the modules of the simulation geometric model, four 6mm circular holes are drilled in the cavity wall, and the boundary conditions are set as follows: two holes are inlets and two holes are outlets; during the purge simulation, the initial velocity of one inlet is 0, and the initial velocity of the other is set to 10L / s, one of the outlets is blocked, and the initial velocity of the other outlet is set to 0; particles with a molecular mass of 481.55g / mol and a relative density of 1.639 are uniformly distributed in the cavity, iterate 100 times, and start the calculation.

3. The wafer anti-sticking layer coating purging simulation method based on fluid mechanics calculation according to claim 1 is characterized in that: The fluid calculation application is based on the SIMPLE algorithm for pressure solution, including: The governing equation for the continuous phase is as follows: Where: u, v, w are the velocities of the continuous phase in the x, y, and z directions, ρ is the density, f is the force on the continuous phase, μ is the viscosity, S m is the mass source term, V is the volume source term, t is the time, F is the additional force of the dispersed phase on the continuous phase, and x, y, and z represent the three directions respectively; The fluid phase is regarded as a continuous medium and the dispersed phase is regarded as a discrete medium. The Euler-Lagrangian method is used. The turbulent kinetic energy equation and diffusion equation of the model are expressed as follows: Where: k is the turbulent kinetic energy, ε is the dissipation coefficient, G k represents the turbulent kinetic energy generated by the mean velocity gradient, G b represents the turbulent kinetic energy generated by the buoyancy effect, Y M represents the effect of compressible turbulent kinetic energy expansion on the total dissipation rate, C 1ε , C 2ε and C 3ε is a constant, σ k and σ ε is the turbulent Prandtl number.

4. A wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm, comprising: Step 1, using the simulation method described in any one of claims 1 to 3 to optimize and improve the purging condition of the vacuum chamber for coating; Step 2: Customize the cavity according to the optimized model in step 1, punch two holes on the top and bottom of the cavity, punch one hole in the center of the cavity wall, the top hole is connected to the nitrogen cylinder and the sample gas path, and the bottom hole is connected to the exhaust gas box and the vacuum pump; Step 3: Design the overall system of the equipment, including designing the gas circuit part of the lower computer electronic control system, designing the lower computer and the heating platform part of the electronic control system, and designing and programming the upper computer, building a complete electronic control system and connecting it with the upper computer to complete the serial port protocol configuration to realize human-computer interaction; Step 4: Test the overall system obtained in step 3 and record the test data; Step 5: Build a network model for small sample learning based on the experimental data obtained in step 4, and establish a neural network model for multi-mapping prediction.

5. The wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm according to claim 4 is characterized in that: In step three, the purge gas of the gas path of the lower computer is nitrogen. The nitrogen purge gas path, sample gas path and exhaust gas path are opened and closed by relay-controlled solenoid valves, and the vacuum extraction and decompression gas paths are completed by relay-controlled vacuum pumps. The precise temperature control of the heating table is achieved through PID temperature control. The upper computer receives and stores the vacuum degree, voltage, frequency, and temperature sensor data uploaded by the lower computer, and sends instructions to the lower computer to control different gas paths, set temperatures, and time parameters.

6. The wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm according to claim 4 is characterized in that: There are two solutions for the small sample model. One is to use genetic algorithm and particle swarm algorithm to optimize BP neural network and RBF neural network respectively; the other is to split the traditional BP network to simplify the network structure. In this process, single-step prediction is used instead of multi-step prediction to reduce the dimension of the input vector.

7. The wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm according to claim 4 is characterized in that: In step five, environmental data, sample data, etc. are used as input parameters, and hydrophobicity and thermal conductivity are used as output parameters.

8. The wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm according to claim 6, characterized in that: The least squares algorithm of small sample regression is used to compare the models of the two schemes, and the most suitable model is selected by comparing the two schemes.

9. The wafer anti-sticking layer coating process optimization method based on fluid mechanics and small sample prediction algorithm according to claim 1, characterized in that: The chamber design will achieve better purging and deposition results in both long-term and short-term use; the small sample prediction network will be able to optimize the current process in a targeted manner in the absence of historical data.

Images