Intelligent control method for wafer glass powder coating process

Abstract

The present application relates to the field of semiconductor advanced packaging technology, and specifically relates to an intelligent control method of wafer glass powder coating process, comprising the following steps: S1: in the glass powder coating process of the wafer, the multi-source electromechanical physical parameters of the coating execution mechanism at the current time and the micro-area environment temperature of the end of the coating execution mechanism are synchronously collected; the present application deeply fuses the fluid dynamics soft measurement model and the neural network with physical information constraint, not only accurately deduces and predicts the complex thixotropy and thermodynamic decay characteristics of non-Newtonian fluid, but also creatively introduces mechanical oscillation penalty term and advance feed delay compensation in reinforcement learning feedforward control; without increasing expensive entity micro rheological sensor, the control jump caused by high-frequency dithering of the equipment, coating ink breakage and spatial lag distortion are completely eliminated from the bottom physical logic.

Description

Technical Field

[0001] This invention relates to the field of advanced semiconductor packaging technology, and more specifically to an intelligent control method for wafer glass powder coating process. Background Technology

[0002] In advanced semiconductor packaging processes (such as wafer-level packaging and MEMS packaging), glass powder is often used for hermetic bonding between wafers. The coating morphology (uniformity of width and height) of the glass powder slurry directly determines the bonding yield and reliability of the final device. However, glass powder slurry is a high-concentration suspension composed of glass microparticles, organic carriers, and solvents. It is a typical non-Newtonian fluid with thixotropy and temperature sensitivity, and its rheological properties are extremely complex.

[0003] Currently, the industry's control methods for wafer glass powder coating mainly rely on post-processing visual inspection (AOI) and simple PID feedback control, or the preliminary introduction of conventional data-driven neural networks for parameter recommendation. However, under the actual mass production coating conditions of ultra-micro and nanometer scales and ultra-high speeds, existing technologies have revealed the following insurmountable engineering bottlenecks: The underlying physical data is "inaccurate" and "impossible to measure": the space at the coating valve port is extremely small and the slurry is highly abrasive, making it impossible to directly install a physical rheological sensor; at the same time, the mechanical background vibration of the high-speed operating equipment will severely mask the micro-pressure fluctuations inside the fluid, and the strong background thermal radiation from the bottom heating stage (Chuck) of the wafer will also interfere with the measurement of the true contact temperature of the slurry.

[0004] "Static" and "black box" rheological state assessment: Most existing control methods ignore the thixotropic characteristics of glass powder, which "becomes thinner the longer it is stirred and recovers slowly when left to stand." In addition, conventional AI models lack physical law constraints and cannot perceive the stress relaxation and thermodynamic viscosity-temperature decay law of polymer materials. They are very likely to output prediction results that violate the common sense of fluid mechanics, resulting in control lag.

[0005] Control commands conflict with electromechanical and physical limits: Existing intelligent feedforward algorithms often only focus on minimizing topographic errors as their sole optimization objective. When a sudden change in viscosity is predicted, they output a large jump command. This can easily cause high-frequency resonance in robotic arms and overload of servo systems in precision semiconductor equipment, directly leading to "fluid breakage (ink breakage)" of the glass powder at the moment of coating. In addition, conventional algorithms do not consider the mechanical response delay of the underlying hardware, resulting in misalignment and distortion of the compensation action in spatial coordinates.

[0006] In summary, overcoming the limitations of physical measurement and deeply decoupling and integrating complex non-Newtonian fluid dynamics with electromechanical execution logic to achieve intelligent feedforward control that accurately conforms to rheological laws while also ensuring mechanical stability is a technological barrier that urgently needs to be overcome in the field of advanced semiconductor packaging. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an intelligent control method for a wafer glass powder coating process, comprising the following steps: S1: During the wafer glass powder coating process, simultaneously collect the multi-source electromechanical physical parameters of the coating actuator at the current moment and the micro-area ambient temperature at the end of the coating actuator; wherein, the multi-source electromechanical physical parameters include the torque feedback current of the drive motor and the high-frequency dynamic back pressure of the fluid in the feeding chamber; S2: Input the multi-source electromechanical physical parameters and the micro-area ambient temperature into a preset fluid dynamics soft measurement mapping model, and calculate and obtain the apparent viscosity and instantaneous shear rate of the glass powder slurry at the current moment; S3: Extract the cumulative integral of shear force over time within a preset historical time window to characterize the thixotropic properties of the glass powder slurry as a non-Newtonian fluid; fuse the thixotropic properties, the apparent viscosity, and the instantaneous shear rate into a state feature vector, and input it into a rheological state prediction network containing a time memory mechanism to output the predicted rheological state of the glass powder slurry at the next target coating coordinate point; S4: Input the predicted rheological state into a pre-trained reinforcement learning feedforward decision network to generate a device action compensation command for the next target coating coordinate point; S5: Before the coating actuator reaches the next target coating coordinate point, the coating pressure and moving speed of the coating actuator are adjusted in advance according to the equipment action compensation command to offset the coating morphology deviation caused by the predicted rheological state change. The reward function used by the reinforcement learning feedforward decision network during the training phase is constructed by weighted summation of the actual coating width deviation value, the actual coating height deviation value, and the equipment over-adjustment penalty term obtained after coating and molding, so that the three-dimensional morphology of the glass powder output by the system under dynamic compensation converges to the target process morphology.

[0008] Further, in step S1, the synchronous acquisition of the multi-source electromechanical physical parameters of the coating actuator at the current moment and the micro-area ambient temperature at the end of the coating actuator specifically includes: Based on the industrial real-time Ethernet bus, the real-time torque feedback current of the drive motor is synchronously extracted at a preset high-frequency sampling rate, as well as the high-frequency dynamic back pressure of the fluid collected by the piezoelectric micro-pressure sensor embedded in the inner wall of the feeding chamber. Adaptive Kalman filtering is applied to the high-frequency dynamic back pressure of the fluid to isolate and filter out the mechanical background vibration noise generated by the coating actuator during its movement, and to extract the true micro-pressure fluctuation characteristics caused by the friction of particles inside the glass powder slurry.

[0009] Furthermore, in step S1, the synchronous acquisition of the multi-source electromechanical physical parameters of the coating actuator at the current moment and the micro-area ambient temperature at the end of the coating actuator also includes: By using an infrared micro-nano temperature measurement array located at the end of the coating actuator, the micro-area ambient temperature gradient distribution data of the interface between the glass powder slurry and the wafer at the moment of extrusion is obtained, so as to eliminate the background thermal radiation interference of the wafer bottom heating stage on the ambient temperature. The real-time torque feedback current, the filtered fluid high-frequency dynamic back pressure, and the micro-area ambient temperature gradient distribution data are given a unified microsecond-level timestamp at the hardware level to construct a multi-source spatiotemporal aligned dataset with no phase delay at the current moment.

[0010] Furthermore, the preset fluid dynamics soft-sensor mapping model incorporates an electromechanical-rheological conversion relationship based on the constitutive equation of non-Newtonian fluids in glass powder slurry; in step S2, the initial reference values ​​of instantaneous shear rate and apparent viscosity are obtained through deduction and calculation, specifically including: Extract the key geometric parameters of the coating actuator, establish a linear proportional mapping matrix between the real-time torque feedback current of the drive motor and the mechanical shear stress inside the slurry, and calculate the real-time shear stress of the glass powder slurry at the current moment. The real-time shear stress of the glass powder slurry at the current moment is calculated using the following mapping formula: In the formula, This represents the real-time shear stress at the current moment. For real-time torque feedback current, The mapping coefficient from electromagnetic torque to shear stress is denoted as . The loss constant of the mechanical transmission system; Based on the non-Newtonian fluid correction model of Poiseuille's flow law, the high-frequency dynamic back pressure of the fluid in the feeding chamber is coupled with the cross-sectional area parameter of the outlet to calculate the internal velocity distribution of the glass powder slurry in the current chamber, and the instantaneous shear rate is obtained by differentiating the velocity at the pipe wall. The calculation formula is: In the formula, Instantaneous shear rate, The instantaneous flow rate is calculated based on the cavity pressure. The nozzle's equivalent radius. It is an index representing the non-Newtonian fluid flow behavior. For high-frequency dynamic back pressure of fluid, The cross-sectional area of ​​the glue outlet. The fluid resistance structural coefficient; Calculate the ratio of the real-time shear stress to the instantaneous shear rate to obtain the initial reference value of the apparent viscosity without thermodynamic correction. ; Based on the non-Newtonian fluid correction model of Poiseuille's flow law, the high-frequency dynamic back pressure of the fluid in the feeding chamber is coupled with the cross-sectional area parameter of the outlet to calculate the internal velocity distribution of the glass powder slurry in the current chamber, and the instantaneous shear rate is obtained by differentiating the velocity at the pipe wall. Calculate the ratio of the real-time shear stress to the instantaneous shear rate to obtain the initial reference value of the apparent viscosity without thermodynamic correction.

[0011] Furthermore, in step S2, calculating the apparent viscosity of the glass powder slurry at the current moment also includes a local thermodynamic dynamic compensation step for the initial reference value of the apparent viscosity: Based on the collected micro-area environmental temperature gradient distribution data, the core contact temperature of the interface between the glass powder slurry body and the wafer is extracted. The core contact temperature is converted into a thermodynamic viscosity compensation coefficient characterizing the relaxation properties of the polymer network by invoking a pre-calibrated Arrhenius viscosity-temperature index decay function for glass powder. The decay function is as follows: In the formula, This is the thermodynamic viscosity compensation coefficient. The activation energy for the flow of glass powder slurry. Let be the ideal gas constant. Core contact temperature; The Herschel-Bulkley rheological empirical constant is introduced as the yield stress correction term. The thermodynamic viscosity compensation coefficient, the yield stress correction term, and the initial reference value of apparent viscosity are nonlinearly fused to output the final apparent viscosity at the current moment after dual correction for temperature and yield characteristics. The nonlinear fusion formula is as follows: In the formula, For the final apparent viscosity, This is a correction term for yield stress. This is the consistency coefficient. Instantaneous shear rate, This is a liquidity behavior index.

[0012] Further, in step S3, the cumulative integral of shear force over time of the glass powder slurry within a preset historical time window is extracted to characterize its thixotropic properties and fused into a state feature vector, specifically including: Establish a sliding time window integral function based on the disintegration dynamics of material microstructure; Multiple historical instantaneous shear rates sampled continuously within the preset historical time window are extracted. An exponential forgetting factor, characterizing the recombination rate of the glass powder polymer network structure, is introduced. The historical instantaneous shear rates are then integrated with time decay weights to calculate the thixotropic structure state at the current moment. The sliding time window integral formula is as follows: In the formula, This is the current state vector, and its values ​​range from [0,1]. To preset the historical time window length, For history Instantaneous shear rate, The exponential forgetting factor is used to characterize the recombination rate of the polymer network structure of glass powder. The structural disintegration rate constant; The thixotropic structural state variables, the apparent viscosity at the current moment, etc., are tensor-concatenated to construct a spatiotemporal state feature vector: .

[0013] Furthermore, the rheological state prediction network incorporating a time memory mechanism is a long short-term memory neural network with physical information constraints; in step S3, the output of the predicted rheological state specifically includes: The spatiotemporal state feature vector is input into the cell state unit of the neural network, and invalid historical shear features that exceed the inherent stress relaxation time of the glass powder slurry are automatically removed through the forget gate mechanism inside the network. By using the input gate mechanism within the network, the cumulative heat generation characteristics and dynamic viscosity degradation characteristics caused by the continuous high-speed movement of the coating equipment are updated and extracted. During the model training phase of the neural network, its loss function is configured as a composite loss function consisting of the prediction mean square error term and the residual term of the rheological dynamic equation, with physical conservation laws as soft constraints to limit the gradient update direction of the network parameters. After nonlinear mapping of the output gate of the neural network, the predicted apparent viscosity and predicted flow behavior index of the glass powder slurry when it reaches the next target coating coordinate point are output as the predicted rheological state.

[0014] Further, in step S4, generating device action compensation instructions for the next target coating coordinate point, and constructing the reward function of the reinforcement learning feedforward decision network, specifically includes: A Markov decision state space is constructed that maps to the physical coating environment. The state space includes the predicted apparent viscosity and predicted flow behavior index in the predicted rheological state, as well as the current absolute spatial coordinates and remaining coating trajectory length of the coating actuator. A continuous device action space is defined. The Actor network of the reinforcement learning feedforward decision network outputs a continuous action vector based on the current state space. The action vector includes: the downpressure compensation amount of the Z-axis piezoelectric fine-tuning valve, the feed speed compensation amount of the XY-axis servo motor, and the duty cycle compensation amount of the air pressure pulse in the feeding chamber.

[0015] Furthermore, step S4 also includes: Construct a reward function that includes multi-objective physical constraints. The mathematical expression of the reward function is as follows: in, and These represent the coating linewidth deviation and line height deviation measured by a high-precision 3D profilometer, respectively. and The morphological convergence weight coefficient is used. The weighting coefficient is the penalty coefficient for mechanical oscillation. This item is used to constrain the abrupt changes in action within consecutive adjacent control cycles, in order to prevent high-frequency shaking of the robotic arm and ink breakage of glass powder caused by excessive changes in the rate of change of the equipment action compensation command.

[0016] Furthermore, step S5 and the training phase of the reinforcement learning feedforward decision network specifically include: Equipment response delay compensation execution: Extract the inherent hardware communication and mechanical response delay time between the Z-axis piezoelectric fine-tuning valve and the XY-axis servo motor; before the coating actuator reaches the next target coating coordinate point, issue the equipment action compensation command one mechanical response delay time in advance to ensure that the physical effective time of the mechanical action is strictly aligned with the spatial coordinates of the predicted rheological state; Cross-batch morphology closed loop: After the glass powder coating process of a single wafer is completed, a high-precision 3D laser confocal sensor arranged in a coaxial or off-axis manner is controlled to scan the actual three-dimensional morphology of the coating and extract the linewidth and lineheight matrix of the entire trajectory. Policy network adaptive evolution: Calculate the true reward value corresponding to the actual shape matrix, evaluate the Q value of the action vector using the Critic network in the deep deterministic policy gradient algorithm or the proximal policy optimization algorithm, calculate the time difference error, and update the model weights of the reinforcement learning feedforward decision network through error backpropagation, so as to realize the policy adaptive iteration under the condition of coating reference drift caused by batch differences of glass powder and natural wear of equipment.

[0017] Beneficial effects This invention overcomes the engineering challenges of "inaccurate measurement, incorrect calculation, and unstable control" in traditional wafer glass powder coating. By deeply integrating a fluid dynamics soft measurement model with a neural network with physical information constraints (PI-LSTM), it not only accurately deduce and predicts the complex thixotropic and thermodynamic decay characteristics of non-Newtonian fluids, but also creatively introduces a mechanical oscillation penalty term and feedforward delay compensation in reinforcement learning feedforward control. Without adding expensive physical micro-rheological sensors, it completely eliminates high-frequency equipment jitter, coating ink breakage, and spatial lag distortion caused by control jumps from the underlying physical logic. Ultimately, it achieves a high-precision morphology closed-loop control with strong engineering robustness and cross-batch adaptive maintenance-free operation. Attached Figure Description

[0018] Figure 1 This is the overall process flow of the intelligent control method for wafer glass powder coating of the present invention; Figure 2 This is a schematic diagram of the hardware architecture and signal flow of the system of the present invention; Figure 3 This is a logic diagram of the rheological soft measurement and thixotropic state prediction algorithm of the present invention; Figure 4 This is a schematic diagram illustrating the reinforcement learning feedforward execution and closed-loop principle of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] The present invention will now be described in further detail with reference to the accompanying drawings: Example: like Figure 1 As shown, this invention provides an intelligent control method for a wafer glass powder coating process, comprising the following steps: S1: During the glass powder coating process of the wafer, the multi-source electromechanical physical parameters of the coating actuator and the micro-area ambient temperature at the end of the coating actuator are collected simultaneously at the current moment; wherein, the multi-source electromechanical physical parameters include the torque feedback current of the drive motor and the high-frequency dynamic back pressure of the fluid in the feeding chamber. S2: Input the multi-source electromechanical physical parameters and the micro-area ambient temperature into a preset fluid dynamics soft measurement mapping model, and calculate and obtain the apparent viscosity and instantaneous shear rate of the glass powder slurry at the current moment; S3: Extract the cumulative integral of shear force over time within a preset historical time window to characterize the thixotropic properties of the glass powder slurry as a non-Newtonian fluid; fuse the thixotropic properties, the apparent viscosity, and the instantaneous shear rate into a state feature vector, and input it into a rheological state prediction network containing a time memory mechanism to output the predicted rheological state of the glass powder slurry at the next target coating coordinate point; S4: Input the predicted rheological state into a pre-trained reinforcement learning feedforward decision network to generate a device action compensation command for the next target coating coordinate point; S5: Before the coating actuator reaches the next target coating coordinate point, the coating pressure and moving speed of the coating actuator are adjusted in advance according to the equipment action compensation command to offset the coating morphology deviation caused by the predicted rheological state change. The reward function used by the reinforcement learning feedforward decision network during the training phase is constructed by weighted summation of the actual coating width deviation value, the actual coating height deviation value, and the equipment over-adjustment penalty term obtained after coating and molding, so that the three-dimensional morphology of the glass powder output by the system under dynamic compensation converges to the target process morphology.

[0022] Furthermore, in step S1, to address the coating unevenness issue caused by the non-Newtonian fluid characteristics of glass powder slurry in high-precision wafer-level packaging processes, the system first needs to obtain extremely accurate and zero-delay underlying physical feedback. Specifically, the execution process of step S1, "synchronously acquiring the multi-source electromechanical physical parameters of the coating actuator and the micro-area ambient temperature at the current moment," is as follows: First, during the coating process, as the coating actuator (e.g., a dispensing valve or screen printing squeegee) moves at high speed along a pre-defined wafer substrate trajectory, the system's underlying control board, based on an industrial real-time Ethernet bus (such as EtherCAT or PROFINET), captures the underlying electromechanical data of the drive motor in real time at a preset high-frequency sampling rate (e.g., 10kHz). Specifically, the system bypasses conventional surface position feedback and directly extracts the real-time torque feedback current inside the drive motor. The minute fluctuations in this torque current directly reflect the instantaneous mechanical resistance overcome by the coating actuator when pushing the high-viscosity glass powder slurry.

[0023] While extracting the torque feedback current, the system simultaneously activates a piezoelectric micro-pressure sensor embedded in the inner wall of the feeding chamber of the coating actuator to collect the high-frequency dynamic back pressure of the fluid. However, in actual micro-nano-scale semiconductor coating environments, due to the high-speed start-stop of the servo motor and the interpolation motion of the robotic arm, a large amount of mechanical background vibration noise is inevitably mixed into the raw pressure signal collected by the piezoelectric sensor. This noise often masks the real micro-pressure fluctuations generated by the friction of particles inside the glass powder.

[0024] To eliminate the aforementioned mechanical noise, this embodiment introduces an adaptive Kalman filter mechanism. During execution, the system dynamically reads the encoder speed and acceleration of the servo motor at the current moment as feedforward reference variables, and adaptively adjusts the process noise covariance matrix and observation noise covariance matrix of the Kalman filter in real time. When the robotic arm is in a high acceleration / deceleration phase, the system automatically increases the filtering weight for high-frequency mechanical vibrations, thereby accurately separating and extracting the real high-frequency dynamic back pressure of the fluid caused purely by rheological properties from the complex original signal. This step completely solves the engineering pain point of "inaccurate measurement" of traditional pressure sensors in high-speed coating equipment.

[0025] Furthermore, due to the exponentially sensitive viscosity of glass powder to temperature changes (i.e., temperature dependence in polymer rheology), and the heating stage at the bottom of the wafer radiating a large amount of interfering heat, the system activates an infrared micro-nano temperature measurement array located at the end of the coating actuator during the same execution cycle of acquiring the aforementioned electromechanical parameters. This array does not blindly measure the macroscopic ambient temperature, but rather precisely focuses its scanning field of view on the "meniscus" region at the moment the glass powder slurry is extruded and contacts the wafer, acquiring micro-area ambient temperature gradient distribution data. During this process, the system's internal image processing module uses a differential blanking algorithm to subtract the thermal radiation infrared spectrum of the uncoated wafer surface from the background reference, thereby completely eliminating background thermal radiation interference from the bottom heating stage and extracting the most accurate core contact temperature of the slurry body at the moment of contact.

[0026] Finally, to ensure that the input to the subsequent artificial intelligence prediction model is free from timing errors, the system must address the phase delay issue caused by asynchronous sampling clocks among different sensors (motor driver, piezoelectric sensor, infrared temperature measurement array). In this embodiment, the system uses a field-programmable gate array (FPGA) as a unified hardware clock source. When the real-time torque feedback current, filtered fluid high-frequency dynamic back pressure, and micro-area environmental temperature gradient distribution data arrive, the FPGA will assign a unified microsecond-level timestamp to each frame of heterogeneous data at the nanosecond level. For data streams sampled at different frequencies, the system uses a spline interpolation algorithm to force them to align to the same absolute time node. Above. Through this extremely rigorous spatiotemporal synchronization process, the system finally constructs a "multi-source spatiotemporal aligned dataset" with no phase delay at the current moment, laying an extremely pure and rigorous physical data foundation for subsequent high-precision soft measurement of apparent viscosity and prediction of thixotropic state.

[0027] After obtaining a high-purity and spatiotemporally strictly aligned multi-source physical dataset through step S1, the system proceeds to step S2, the rheological soft measurement deduction stage. Those skilled in the art know that directly mounting a solid viscometer in the micro-nano slits of wafer-level packaging faces engineering challenges such as spatial interference, corrosion and wear, and disruption of the flow field. Therefore, this embodiment creatively introduces a fluid dynamics soft measurement mapping model based on non-Newtonian fluid constitutive equations, enabling cross-dimensional deduction from "surface electromechanical parameters" to "underlying rheological parameters."

[0028] Furthermore, in step S2, during the specific simulation, the system's main control unit first extracts the key geometric parameters of the current coating actuator (e.g., the pitch and stator inner diameter of the dispensing micro-screw, or the taper and capillary length of the piezoelectric nozzle). Subsequently, the system imports the real-time torque feedback current obtained in S1, without delay, into a preset linear proportional mapping matrix. This matrix pre-calibrates the coupling relationship between the motor's electromagnetic torque, the reducer's transmission losses, and the terminal mechanical thrust, thereby eliminating mechanical transmission losses and accurately calculating the actual mechanical shear stress applied by the actuator to the glass powder slurry at the current moment.

[0029] Building upon this foundation, the system does not simply apply Newtonian fluid dynamics formulas. Instead, considering the high filling ratio characteristics of glass powder, it employs a non-Newtonian fluid dynamics correction model based on Poiseuille's flow law. The system uses the filtered high-frequency dynamic back pressure of the fluid as a boundary driving force parameter, coupled with the microscopic cross-sectional area parameter of the dispensing port through differential equation calculations. Through this calculation, the system not only reconstructs the non-parabolic internal velocity distribution profile of the slurry within the feeding chamber and nozzle, but also accurately extracts the instantaneous shear rate of the glass powder slurry under intense extrusion by differentiating the gradient of this distribution profile at the pipe wall boundary. At this point, the system divides the actual mechanical shear stress by this instantaneous shear rate to quickly obtain the initial baseline value of the apparent viscosity based solely on pure mechanical stress analysis.

[0030] However, relying solely on soft measurements using mechanical parameters cannot solve the problem of ink breakage caused by glass powder during wafer coating. This is because glass powder slurry is essentially a complex suspension composed of glass powder, organic carrier, and solvent. Its viscous resistance stems not only from mechanical shear but is also extremely sensitive to the thermodynamic state of the contact interface. Therefore, this embodiment further implements in-depth local thermodynamic dynamic compensation after obtaining the initial reference values.

[0031] The system's core processor retrieves the micro-area ambient temperature gradient distribution data stripped of background radiation from S1, and uses an optimization algorithm to locate and extract the core contact temperature at which the slurry just contacts the wafer interface. Subsequently, the system substitutes this core contact temperature into a glass powder Arrhenius viscosity-temperature index decay function pre-calibrated in the laboratory. This function simulates the relaxation characteristics of the polymer network under different thermal activation conditions, thus outputting a dimensionless thermodynamic viscosity compensation coefficient.

[0032] In the final fusion output stage, to accurately reproduce the yield characteristic of glass powder—that it "does not flow without sufficient external force"—the system introduces the Herschel-Bulkley rheological empirical constant as a yield stress correction term. Finally, the soft-sensor mapping model performs a nonlinear product and bias fusion of the thermodynamic viscosity compensation coefficient, the yield stress correction term, and the previously calculated initial apparent viscosity reference value. After this series of rigorous fluid dynamics and thermodynamic corrections, the system finally outputs a highly physically faithful "final apparent viscosity" for the current moment. This soft-sensor result not only far surpasses traditional solid viscometers in response speed but also represents a generational leap in measurement accuracy under extreme process environments.

[0033] Specifically, the derivation and calculation process of the fluid dynamics soft-sensor mapping model is as follows: First, based on the electromagnetic torque principle and fluid back pressure characteristics, the real-time shear stress and instantaneous shear rate are calculated respectively, with the specific formulas as follows: Real-time shear stress: ; Instantaneous shear rate: ; in, For torque feedback current, For high-frequency dynamic back pressure of fluid, These are calibration coefficients.

[0034] Subsequently, to accurately extrapolate the apparent viscosity, this embodiment introduces Arrhenius viscosity-temperature compensation and the Herschel-Bulkley rheological equation to calculate the apparent viscosity at the current moment after dual corrections based on thermodynamics and yield characteristics: In the formula, For micro-area ambient temperature, For the yield stress of the slurry, This is the consistency coefficient. These are non-Newtonian exponents. Using the above formula, the system accurately reconstructs the physical state of the fluid's underlying layer without adding physical micro-sensors.

[0035] Furthermore, in step S3, after the system accurately obtains the final apparent viscosity and instantaneous shear rate at the current moment through step S2, the technical challenge in this field lies in the fact that glass powder slurry is not a static fluid, but a typical thixotropic non-Newtonian fluid. This means that its viscosity depends not only on the "current" stress state, but also on the profound influence of the "past stress history" (i.e., the physical characteristic of "becoming thinner the longer it is stirred, and slowly recovering its viscosity after standing"). If control is based solely on the current state, the system will always be in a state of lag. Therefore, this embodiment introduces a physical constraint prediction mechanism that deeply integrates materials rheology in step S3.

[0036] To quantitatively calculate the impact of the aforementioned thixotropy on the fluid state at the current moment, this embodiment uses a time integral decay formula with an exponential forgetting factor to calculate the thixotropic structural state quantity. : In the formula, The preset historical time window length, For a historic moment The instantaneous shear rate; It is an exponential forgetting factor, which represents the smaller the impact of shearing on the current polymer network structure as time goes by (i.e., the natural recovery and reorganization of the slurry structure). is the structural disintegration constant.

[0037] First, the system's main control unit initiates the feature extraction module to construct a sliding time window based on the material's microstructure disintegration dynamics. The system not only extracts the current state but also retrieves multiple historical instantaneous shear rates continuously sampled within this preset historical time window from the cache. To accurately quantify the degree of damage to the polymer network structure inside the slurry under continuous shear, the system does not use a simple arithmetic average but introduces an exponential forgetting factor characterizing the glass powder's structural recombination rate. By performing an integral operation with time decay weights on the historical instantaneous shear rates, the system calculates a crucial dimensionless physical quantity—the thixotropic structural state quantity. ,when When the value approaches 1, it indicates that the internal micro-network structure of the slurry is intact; when... When the value approaches 0, it indicates that the structure has been completely sheared and destroyed.

[0038] Subsequently, the system performs high-dimensional tensor splicing on the calculated thixotropic structural state variables, the apparent viscosity at the current moment, the instantaneous shear rate, and the spatial three-dimensional displacement vector of the coating actuator moving from the current absolute position to the next target coating coordinate point, thereby constructing a spatiotemporal state feature vector that includes both the multidimensional physical mapping relationship of the fluid and the spatial kinematics of the equipment.

[0039] After extracting and tensorizing the physical features, the system inputs the spatiotemporal state feature vector into the core prediction center of this invention—a long short-term memory neural network with physical information constraints (Physics-InformedLSTM, or PI-LSTM for short). In this embodiment, the PI-LSTM is not a traditional pure data-driven black-box model, but rather its internal gating units are deeply bound to rheological physical mechanisms: When feature vectors enter the cellular state units of the PI-LSTM, the network's forget gate mechanism is specifically configured to map the stress relaxation time of the polymer material. For historical shear features that occurred extremely long ago and exceed the material's own memory period, the forget gate adaptively clears their weights to zero, perfectly simulating the natural dissipation of internal stress in a fluid after it has settled. Meanwhile, the network's input gate mechanism is specifically used to extract and update the accumulated heat generation features and dynamic viscosity degradation features caused by the continuous high-speed frictional motion of the coating equipment.

[0040] More importantly, to ensure that the predictions output by the AI ​​network absolutely conform to the laws of reality, the system innovatively configures the loss function of the PI-LSTM model as a composite loss function during the offline training and online fine-tuning stages. This loss function not only includes the traditional "prediction mean square error term" used to measure prediction bias, but also rigidly incorporates the "rheological dynamic equation residual term" derived from the Navier-Stokes equations.

[0041] By using physical conservation laws as soft constraints for model gradient descent, the system completely restricts network parameters from updating in erroneous directions that violate energy conservation or fluid dissipation laws. Finally, through nonlinear mapping of the PI-LSTM network output gate, the system accurately outputs the predicted apparent viscosity and predicted flow behavior index of the glass powder slurry as it approaches the next target coating coordinate point, serving as the final predicted rheological state. This prediction result not only possesses extremely high data accuracy but also extremely rigorous physical interpretability, securing a valuable "time difference" for subsequent millisecond-level feedforward compensation.

[0042] Furthermore, the feedforward execution and adaptive closed-loop process in steps S4 and S5: After accurately obtaining the predicted apparent viscosity and predicted flow behavior index at the next target coating coordinate point through the PI-LSTM network in step S3, the system then enters the reinforcement learning feedforward decision-making stage in step S4 and the physical execution closed-loop stage in step S5. Those skilled in the art understand that in high-speed wafer coating processes at hundreds of millimeters per second, simple hysteresis feedback control inevitably leads to periodic fluctuations in the coating morphology. Therefore, this embodiment constructs a Markov decision process (MDP) that deeply maps the physical coating environment to achieve dynamic compensation through advanced feedforward.

[0043] Specifically, the system's main control unit assembles the predicted rheological state, the current absolute spatial coordinates of the coating actuator, and the remaining coating trajectory length into a high-dimensional state space. This state space is then input into a pre-trained reinforcement learning feedforward decision network (in this embodiment, a Deep Deterministic Policy Gradient (DDPG) algorithm Actor network suitable for continuous motion control is preferably used). Based on the current complex rheological state, the Actor network rapidly outputs a set of continuous action vectors. These action vectors are not abstract numbers but directly correspond to the control commands of the underlying hardware, specifically including: the downward pressure compensation amount of the Z-axis piezoelectric fine-tuning valve, the feed speed fine-tuning amount of the XY-axis servo motor, and the duty cycle compensation amount of the air pressure pulse in the feeding chamber.

[0044] In traditional AI control, control accuracy is often the sole optimization objective. This is extremely dangerous in real high-precision semiconductor equipment. In pursuit of extreme shape consistency, AI may frequently output wide-bandwidth, large-amplitude jump commands, directly causing high-frequency resonance in the robotic arm or even triggering "fluid breakage (ink breakage)" of the glass powder slurry at the nozzle. To completely overcome this engineering fatal flaw, this embodiment innovatively constructs a composite reward function with multi-objective physical constraints in the training and evaluation mechanism of the DDPG network. The basic tracking reward, composed of coating shape error, further incorporates physical penalties for the risks of high-frequency mechanical resonance and fluid breakage. The specific mathematical expression of this composite reward function is as follows: In the formula, Penalty for shape tracking error; The instantaneous jump rate of control commands (such as Z-axis fine-tuning, servo feed, and air pressure duty cycle) is characterized by weights. Limit AI from outputting high-frequency, large-amplitude commands; As a fluid fracture penalty term, when the derived instantaneous shear rate... When the critical fracture threshold of the glass powder material is exceeded, severe penalties will be imposed.

[0045] This reward function not only includes penalty terms for coating linewidth and line height deviations measured by a high-precision 3D profilometer, but also introduces a crucial "mechanical oscillation penalty weight term." By summing the squares of the rate of change of action commands (i.e., the first derivative of the action) within consecutive adjacent control cycles and imposing heavy penalties, the system rigidly constrains the abrupt changes in the AI's output actions, thus finding a perfect physical balance between "optimization of morphological accuracy" and "stability of the electromechanical system."

[0046] After generating the aforementioned compensation instructions for physically constrained device actions, the system enters the execution and closed-loop stage in step S5. To eliminate spatial coordinate misalignment caused by the "time difference" between instruction issuance and actual device action, the system's underlying driver extracts the piezoelectric ceramic voltage setup response time of the Z-axis piezoelectric fine-tuning valve and the mechanical response delay time of the XY-axis servo motor. Before the coating actuator reaches the next target coating coordinate point, the system issues the compensation instruction precisely in advance by a "comprehensive hardware response delay time." This "feedforward" action ensures that at the instant the mechanical mechanism undergoes physical deformation, the nozzle of the actuator precisely reaches the spatial coordinate point where the predicted rheological state abruptly changes, perfectly offsetting the morphological distortion caused by viscosity fluctuations.

[0047] Finally, to endow the system with the ability to self-evolve against natural wear and tear (such as nozzle micro-wear) and batch-to-batch material variations, this embodiment introduces a cross-batch morphology closed-loop mechanism. After the glass powder coating process on a single wafer or in a single batch is completed, the system controls a high-precision 3D laser confocal sensor arranged in a paraxial configuration to perform a full-trajectory scan of the actual three-dimensional morphology after coating and curing, extracting the true linewidth and lineheight matrices. The system substitutes this true morphology matrix into the aforementioned composite reward function to calculate the true reward value, and uses the Critic network in the DDPG algorithm to evaluate the Q-value of the previously issued action vector, calculating the temporal difference error. Through backpropagation of the error, the system silently updates the weight parameters of the reinforcement learning feedforward decision network in the background. Through this "batch-to-batch self-iteration" closed-loop mechanism, the entire coating system completely eliminates the extreme dependence of traditional processes on manual parameter tuning by senior engineers, achieving high-level intelligence of "becoming more accurate with use" and "maintenance-free" under complex and variable working conditions.

[0048] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart control method for a wafer glass powder coating process, characterized in that, Includes the following steps: S1: During the glass powder coating process of the wafer, the multi-source electromechanical physical parameters of the coating actuator and the micro-area ambient temperature at the end of the coating actuator are collected simultaneously at the current moment; wherein, the multi-source electromechanical physical parameters include the torque feedback current of the drive motor and the high-frequency dynamic back pressure of the fluid in the feeding chamber. S2: Input the multi-source electromechanical physical parameters and the micro-area ambient temperature into a preset fluid dynamics soft measurement mapping model, and calculate and obtain the apparent viscosity and instantaneous shear rate of the glass powder slurry at the current moment; S3: Extract the cumulative integral of shear force over time within a preset historical time window to characterize the thixotropic properties of the glass powder slurry as a non-Newtonian fluid; fuse the thixotropic properties, the apparent viscosity, and the instantaneous shear rate into a state feature vector, and input it into a rheological state prediction network containing a time memory mechanism to output the predicted rheological state of the glass powder slurry at the next target coating coordinate point; S4: Input the predicted rheological state into a pre-trained reinforcement learning feedforward decision network to generate a device action compensation command for the next target coating coordinate point; S5: Before the coating actuator reaches the next target coating coordinate point, the coating pressure and moving speed of the coating actuator are adjusted in advance according to the equipment action compensation command to offset the coating morphology deviation caused by the predicted rheological state change. The reward function used by the reinforcement learning feedforward decision network during the training phase is constructed by weighted summation of the actual coating width deviation value, the actual coating height deviation value, and the equipment over-adjustment penalty term obtained after coating and molding, so that the three-dimensional morphology of the glass powder output by the system under dynamic compensation converges to the target process morphology.

2. The intelligent control method for a wafer glass powder coating process according to claim 1, characterized in that, In step S1, the synchronous acquisition of the multi-source electromechanical physical parameters of the coating actuator at the current moment and the micro-area ambient temperature at the end of the coating actuator specifically includes: Based on the industrial real-time Ethernet bus, the real-time torque feedback current of the drive motor is synchronously extracted at a preset high-frequency sampling rate, as well as the high-frequency dynamic back pressure of the fluid collected by the piezoelectric micro-pressure sensor embedded in the inner wall of the feeding chamber. Adaptive Kalman filtering is applied to the high-frequency dynamic back pressure of the fluid to isolate and filter out the mechanical background vibration noise generated by the coating actuator during its movement, and to extract the true micro-pressure fluctuation characteristics caused by the friction of particles inside the glass powder slurry.

3. The intelligent control method for a wafer glass powder coating process according to claim 2, characterized in that, In step S1, the synchronous acquisition of the multi-source electromechanical physical parameters of the coating actuator at the current moment and the micro-area ambient temperature at the end of the coating actuator also includes: By using an infrared micro-nano temperature measurement array located at the end of the coating actuator, the micro-area ambient temperature gradient distribution data of the interface between the glass powder slurry and the wafer at the moment of extrusion is obtained, so as to eliminate the background thermal radiation interference of the wafer bottom heating stage on the ambient temperature. The real-time torque feedback current, the filtered fluid high-frequency dynamic back pressure, and the micro-area ambient temperature gradient distribution data are given a unified microsecond-level timestamp at the hardware level to construct a multi-source spatiotemporal aligned dataset with no phase delay at the current moment.

4. The intelligent control method for a wafer glass powder coating process according to claim 3, characterized in that, The preset fluid dynamics soft measurement mapping model has a built-in electromechanical-rheological conversion relationship based on the constitutive equation of non-Newtonian fluid of glass powder slurry; In step S2, the initial baseline values ​​for the instantaneous shear rate and apparent viscosity are obtained through deduction and calculation, specifically including: Extract the key geometric parameters of the coating actuator, establish a linear proportional mapping matrix between the real-time torque feedback current of the drive motor and the mechanical shear stress inside the slurry, and calculate the real-time shear stress of the glass powder slurry at the current moment. The real-time shear stress of the glass powder slurry at the current moment is calculated using the following mapping formula: In the formula, This represents the real-time shear stress at the current moment. For real-time torque feedback current, The mapping coefficient from electromagnetic torque to shear stress is denoted as . The loss constant of the mechanical transmission system; Based on the non-Newtonian fluid correction model of Poiseuille's flow law, the high-frequency dynamic back pressure of the fluid in the feeding chamber is coupled with the cross-sectional area parameter of the outlet to calculate the internal velocity distribution of the glass powder slurry in the current chamber, and the instantaneous shear rate is obtained by differentiating the velocity at the pipe wall. The calculation formula is: In the formula, Instantaneous shear rate, The instantaneous flow rate is calculated based on the cavity pressure. The nozzle's equivalent radius. It is an index representing the non-Newtonian fluid flow behavior. For high-frequency dynamic back pressure of fluid, The cross-sectional area of ​​the glue outlet. The fluid resistance structural coefficient; Calculate the ratio of the real-time shear stress to the instantaneous shear rate to obtain the initial reference value of the apparent viscosity without thermodynamic correction. ; Calculate the ratio of the real-time shear stress to the instantaneous shear rate to obtain the initial reference value of the apparent viscosity without thermodynamic correction.

5. The intelligent control method for a wafer glass powder coating process according to claim 4, characterized in that, Step S2, which calculates and obtains the apparent viscosity of the glass powder slurry at the current time, also includes a local thermodynamic dynamic compensation step for the initial reference value of the apparent viscosity: Based on the collected micro-area environmental temperature gradient distribution data, the core contact temperature of the interface between the glass powder slurry body and the wafer is extracted. The core contact temperature is converted into a thermodynamic viscosity compensation coefficient characterizing the relaxation properties of the polymer network by invoking a pre-calibrated Arrhenius viscosity-temperature index decay function for glass powder. The decay function is as follows: In the formula, This is the thermodynamic viscosity compensation coefficient. The activation energy for the flow of glass powder slurry. Let be the ideal gas constant. Core contact temperature; The Herschel-Bulkley rheological empirical constant is introduced as the yield stress correction term. The thermodynamic viscosity compensation coefficient, the yield stress correction term, and the initial reference value of apparent viscosity are nonlinearly fused to output the final apparent viscosity at the current moment after dual correction for temperature and yield characteristics. The nonlinear fusion formula is as follows: In the formula, For the final apparent viscosity, This is a correction term for yield stress. This is the consistency coefficient. Instantaneous shear rate, This is a liquidity behavior index.

6. The intelligent control method for a wafer glass powder coating process according to claim 5, characterized in that, In step S3, the cumulative integral of shear force over time of the glass powder slurry within a preset historical time window is extracted to characterize its thixotropic properties and fused into a state feature vector, specifically including: Establish a sliding time window integral function based on the disintegration dynamics of material microstructure; Multiple historical instantaneous shear rates sampled continuously within the preset historical time window are extracted. An exponential forgetting factor, characterizing the recombination rate of the glass powder polymer network structure, is introduced. The historical instantaneous shear rates are then integrated with time decay weights to calculate the thixotropic structure state at the current moment. The sliding time window integral formula is as follows: In the formula, This is the current state vector, and its values ​​range from [0,1]. To preset the historical time window length, For history Instantaneous shear rate, The exponential forgetting factor is used to characterize the recombination rate of the polymer network structure of glass powder. The structural disintegration rate constant; The thixotropic structural state variables, the apparent viscosity at the current moment, etc., are tensor-concatenated to construct a spatiotemporal state feature vector: .

7. The intelligent control method for a wafer glass powder coating process according to claim 6, characterized in that, The rheological state prediction network incorporating a time-memory mechanism is a long short-term memory neural network with physical information constraints; in step S3, the predicted rheological state is output, specifically including: The spatiotemporal state feature vector is input into the cell state unit of the neural network, and invalid historical shear features that exceed the inherent stress relaxation time of the glass powder slurry are automatically removed through the forget gate mechanism inside the network. By using the input gate mechanism within the network, the cumulative heat generation characteristics and dynamic viscosity degradation characteristics caused by the continuous high-speed movement of the coating equipment are updated and extracted. During the model training phase of the neural network, its loss function is configured as a composite loss function consisting of the prediction mean square error term and the residual term of the rheological dynamic equation, with physical conservation laws as soft constraints to limit the gradient update direction of the network parameters. After nonlinear mapping of the output gate of the neural network, the predicted apparent viscosity and predicted flow behavior index of the glass powder slurry when it reaches the next target coating coordinate point are output as the predicted rheological state.

8. The intelligent control method for a wafer glass powder coating process according to claim 7, characterized in that, In step S4, generating device action compensation instructions for the next target coating coordinate point and constructing the reward function of the reinforcement learning feedforward decision network specifically includes: A Markov decision state space is constructed that maps to the physical coating environment. The state space includes the predicted apparent viscosity and predicted flow behavior index in the predicted rheological state, as well as the current absolute spatial coordinates and remaining coating trajectory length of the coating actuator. A continuous device action space is defined. The Actor network of the reinforcement learning feedforward decision network outputs a continuous action vector based on the current state space. The action vector includes: the downpressure compensation amount of the Z-axis piezoelectric fine-tuning valve, the feed speed compensation amount of the XY-axis servo motor, and the duty cycle compensation amount of the air pressure pulse in the feeding chamber.

9. The intelligent control method for a wafer glass powder coating process according to claim 8, characterized in that, Step S4 also includes: Construct a reward function that includes multi-objective physical constraints. The mathematical expression of the reward function is as follows: in, and These represent the coating linewidth deviation and line height deviation measured by a high-precision 3D profilometer, respectively. and The morphological convergence weight coefficient is used. The weighting coefficient is the penalty coefficient for mechanical oscillation. This item is used to constrain the abrupt changes in action within consecutive adjacent control cycles, in order to prevent high-frequency shaking of the robotic arm and ink breakage of glass powder caused by excessive changes in the rate of change of the equipment action compensation command.

10. The intelligent control method for a wafer glass powder coating process according to claim 8, characterized in that, Step S5 and the training phase of the reinforcement learning feedforward decision network specifically include: Equipment response delay compensation execution: Extract the inherent hardware communication and mechanical response delay time between the Z-axis piezoelectric fine-tuning valve and the XY-axis servo motor; before the coating actuator reaches the next target coating coordinate point, issue the equipment action compensation command one mechanical response delay time in advance to ensure that the physical effective time of the mechanical action is strictly aligned with the spatial coordinates of the predicted rheological state; Cross-batch morphology closed loop: After the glass powder coating process of a single wafer is completed, a high-precision 3D laser confocal sensor arranged in a coaxial or off-axis manner is controlled to scan the actual three-dimensional morphology of the coating and extract the linewidth and lineheight matrix of the entire trajectory. Policy network adaptive evolution: Calculate the true reward value corresponding to the actual shape matrix, evaluate the Q value of the action vector using the Critic network in the deep deterministic policy gradient algorithm or the proximal policy optimization algorithm, calculate the time difference error, and update the model weights of the reinforcement learning feedforward decision network through error backpropagation, so as to realize the policy adaptive iteration under the condition of coating reference drift caused by batch differences of glass powder and natural wear of equipment.

Images