A semiconductor wet cleaning process control method and system based on multi-source time series data fusion and AI prediction

Abstract

The application discloses a semiconductor wet cleaning process control method, system, equipment and medium based on multi-source time sequence data fusion and AI prediction. The method comprises the following steps: synchronously collecting wafer rotation speed, cleaning liquid / gas flow and driving motor load current signal; performing multi-dimensional feature extraction on the load current and fusing other parameters to construct a feature vector, inputting the feature vector into an LSTM network model to output a cleaning effect prediction value in real time; based on the deviation between the prediction value and a target value, querying an expert rule library to generate a process parameter adjustment instruction; executing the instruction and realizing double control of quality and process state based on load current feedback. The application realizes online virtual measurement of cleaning effect and adaptive optimization of process, solves the problems of quality control lag and difficulty in real-time compensation of process disturbance, and improves cleaning uniformity and yield.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, specifically to a method, system, electronic device, and computer-readable storage medium for intelligent prediction and control of the process in semiconductor wafer wet cleaning processes. More particularly, it relates to an online virtual measurement and predictive control technology based on multi-source data fusion and artificial intelligence (AI) models for real-time optimization of the process in rotary spray wafer cleaning equipment. Background Technology

[0002] Wafer cleaning is a critical step in semiconductor manufacturing, and its effectiveness directly affects device performance and yield. As technology nodes continue to shrink, the requirements for the uniformity, precision, and low damage of the cleaning process are becoming increasingly stringent.

[0003] Traditional wet cleaning equipment, especially the mainstream rotary spray cleaning equipment, has the following limitations in its control method, which constitute the core technical problem that this invention aims to solve: First, existing control strategies lack the ability to adapt to the inherent dynamics of the process. While existing technologies (such as Chinese invention patent CN108172533A) can achieve closed-loop stable control of macroscopic environmental parameters such as liquid level and flow rate, they are essentially still based on feedback adjustment of preset formulas or set values. This mode cannot perceive and respond to inherent process disturbances introduced by differences in wafer state, decay of chemical activity, and microscopic changes in fluids, resulting in the uniformity and repeatability of cleaning effects being heavily dependent on the initial state of the equipment, making it difficult to achieve true self-adaptation.

[0004] Secondly, there is a lack of direct, online sensing and control methods for the ultimate quality objective of "cleaning effect." Existing monitoring and control targets are mostly intermediate process parameters, but these parameters have complex and non-linear mapping relationships with the final cleaning effect (such as particle residue and dryness). Current evaluation of cleaning effect heavily relies on offline, post-process detection; this lag prevents real-time correction of ongoing processes.

[0005] Furthermore, existing predictive control technologies have shortcomings when applied to real-time single-wafer cleaning. For example, the CN112433472A scheme uses historical data to build a predictive model for batch-to-batch control, but its model cannot provide real-time prediction and intervention for transient and microscopic dynamics during the single-wafer cleaning process at the millisecond / second level. In addition, the AI ​​models in such schemes are easily regarded as "algorithmic black boxes" detached from the specific physical process, which may face the risk of subject matter examination in patent applications.

[0006] Finally, the implicit electrical signals containing rich process information have not been explored and utilized. In rotary spray cleaning, the load current signal of the drive motor reflects the complex interaction process of "wafer-fluid-mechanical" in real time and at high resolution. However, in the prior art, such signals are only used for basic motor control, and the deep information they contain, which is directly related to the micro-state of cleaning, has not been extracted and used for process optimization.

[0007] Through extensive experimental observation, the applicant discovered a repeatable and quantifiable correlation between the spectral characteristics of the load current signal I(t) of the drive motor and specific process physical states during rotary spray cleaning. For example, in a stable laminar flow rinsing stage, its energy is mainly concentrated at low frequencies (<100Hz); when gas-liquid two-phase flow or turbulence occurs, the energy at mid-frequency frequencies (100-500Hz) increases significantly; and high-frequency resonance or loosening of the mechanical structure excites specific high-frequency components (>500Hz). Simultaneously, the harmonic amplitude of the rotating fundamental frequency is closely related to the dynamic balance state of the wafer clamping. However, existing technologies have never systematically established the mapping relationship between these electrical characteristics and cleaning quality results, let alone used them for control. This invention is based on this unexplored physical phenomenon, constructing a complete technological chain from electrical characteristics to quality prediction and then to process control.

[0008] In summary, a significant technological gap exists in the existing technology: how to achieve real-time, online, and accurate sensing of the intrinsic microscopic state of semiconductor wet cleaning processes without significantly increasing hardware costs, and how to use this sensing to perform proactive and adaptive predictive control of process parameters to directly optimize the final cleaning effect. Therefore, a new technical solution is urgently needed to overcome the aforementioned deficiencies in the existing technology. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a semiconductor wet cleaning process control method and system, electronic device, and computer-readable storage medium based on multi-source time-series data fusion and AI prediction, so as to solve the technical problems existing in the prior art, such as weak process adaptability, lack of real-time perception and control of final cleaning quality, and failure to fully utilize electrical signals containing rich process information.

[0010] In this invention, "online virtual measurement" or "virtual measurement" refers to: using multi-source time-series data collected in real time during the process (as input), and through intelligent algorithms for calculation and reasoning, outputting in real time estimated values ​​of key process quality indicators that cannot or are difficult to measure directly online. This technology enables real-time, indirect perception and evaluation of process effects without interrupting production or relying on offline testing equipment.

[0011] This invention provides a semiconductor wet cleaning process control method based on multi-source time-series data fusion and AI prediction, comprising the following steps: S10. Multi-source timing data acquisition step: Synchronously acquire multi-source timing data during the semiconductor wet cleaning process; the multi-source timing data includes motion parameters reflecting the mechanical motion state, fluid parameters reflecting the fluid supply state, and electrical parameters reflecting the process load; wherein, the motion parameters include the rotational speed ω(t) of the wafer carrier stage; the fluid parameters include the supply flow rates of cleaning liquid and drying gas to the wafer surface; the electrical parameters include the load current signal I(t) of the drive motor that drives the wafer carrier stage to rotate; wherein, t is a continuous time variable; S20. Feature Extraction and Online Virtual Measurement Step: Based on the multi-source time-series data collected in step S10, features are extracted and an instantaneous feature vector F(k) is constructed. The feature sequence S(k) containing the instantaneous feature vector F(k) is input into the trained neural network model, and the cleaning effect prediction vector Y is output in real time. pred (k); where k is the window index after sliding window processing of the signal; S30, Decision suggestion generation step: The cleaning effect prediction vector Y... pred (k) and the preset target effect vector Y target The comparison is performed, and based on the comparison results, a pre-stored expert rule base is queried to generate process parameter adjustment instructions; S40. Process parameter adjustment command execution and feedback control steps: Execute the process parameter adjustment command to adjust the rotation speed of the wafer carrier stage and / or the supply flow rate of the cleaning fluid and drying gas, and realize dual control based on the feedback of the load current signal I(t); the dual control includes quality control based on cleaning effect prediction and process state control based on load current signal.

[0012] As described above, in step S20, the extraction of features and construction of the instantaneous feature vector F(k) specifically includes: The multi-source time-series data is processed using a sliding window method, where for each window index k: For the load current signal I(t), extract its root mean square value I. rms (k) Spectral energy E in the low, medium, and high frequency bands used to characterize different process physical phenomena low (k), E mid (k), E high (k), and the harmonic amplitudes A1(k), A2(k), A3(k) associated with the current rotating fundamental frequency; For the aforementioned motion parameters, extract their angular acceleration α(k); For the fluid parameters, extract the average flow rate change rate; All extracted features are concatenated to form the instantaneous feature vector F(k).

[0013] As described above, in step S20, the neural network model is a long short-term memory network model; the feature sequence S(k) is composed of the instantaneous feature vectors of the current and the preceding M-1 windows, i.e., S(k) = [F(k-M+1), F(k-M+2), ..., F(k)]; where M is a preset sequence length, and M≥2; the cleaning effect prediction vector Y pred (k) includes at least the predicted number of particles per unit area P. pred (k) and predicted water contact angle θ pred (k).

[0014] As described above, in step S30, the expert rule base includes rules based on the predicted cleaning effect and rules based on process state characteristics; the rules based on process state characteristics use the root mean square value of the load current signal, the spectral energy in the low, medium and high frequency bands, and / or whether the harmonic amplitude related to the current rotating fundamental frequency exceeds their respective preset target ranges as one of the triggering conditions.

[0015] As described above, in step S30, generating the process parameter adjustment command includes generating the speed setpoint ω for the next control cycle. set (k+1) and / or fluid flow rate setpoint Q set (k+1).

[0016] In step S40 of the method described above, The quality control based on cleaning effect prediction includes: based on the cleaning effect prediction vector Y pred (k) and the target effect vector Y target To address deviations, iterative optimization of process parameters is performed. The process state control based on the load current signal includes: generating a compensatory adjustment command based on whether the load current signal I(t) deviates from the preset process target curve.

[0017] In the method described above, the low, mid, and high frequency bands are pre-divided continuous frequency intervals, wherein: The low-frequency band is the interval where the frequency f satisfies 10 Hz ≤ f < 100 Hz; The mid-frequency band is the interval where the frequency f satisfies 100 Hz ≤ f < 500 Hz; The high-frequency band is the interval where the frequency f satisfies 500 Hz ≤ f ≤ 1000 Hz.

[0018] The present invention also provides a semiconductor wet cleaning process control system based on multi-source time-series data fusion and AI prediction, for executing the method described in any of the preceding claims, the system comprising: The multi-source data acquisition module (100) is used to execute step S10; A feature extraction and online virtual measurement module (200) is used to perform step S20; The decision suggestion generation module (300) is used to perform step S30; The process execution and feedback control module (400) is used to execute step S40.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the method as described in any of the preceding claims.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor of an electronic device, causes the electronic device to perform the method described in any of the preceding claims.

[0021] Compared with the prior art, the present invention has the following beneficial effects: This invention enables real-time online virtual measurement of cleaning results: By integrating multi-source time-series data such as drive motor load current and using a trained AI model for prediction, the invention can infer the final cleaning quality (such as particle count and water contact angle) in real time during the process, completely changing the lagging mode that relies on offline and post-event detection, and providing a decision basis for real-time control.

[0022] An adaptive predictive control system with dual closed loops of quality objectives and process status was constructed: the system not only optimizes with the final cleaning quality prediction as the target, but also uses the load current signal to quickly compensate for the micro-dynamics of the process, forming a dual control loop, which significantly improves the process's adaptability, robustness and batch consistency.

[0023] The project uncovered and utilized implicit key electrical signals: innovatively elevating the load current signal of the drive motor from a basic equipment control signal into a core process state observation variable characterizing the complex interaction between "wafer-fluid-mechanical". By extracting deep features (time domain, frequency domain, harmonics) from this signal, it achieved indirect and precise perception of microscopic cleaning physical processes (such as liquid film uniformity and turbulence state) that are difficult to observe directly.

[0024] It provides a clear and implementable intelligent solution: This invention deeply integrates AI models, expert rule bases, and specific semiconductor wet cleaning processes and hardware (sensors, actuators). The solution is specific and complete, avoids the risk of "algorithm black box", and has high industrial feasibility and practical value. Attached Figure Description

[0025] Figure 1 This is a schematic flowchart of a semiconductor wet cleaning process control method provided in an embodiment of the present invention.

[0026] Figure 2 This is an embodiment of the present invention for performing... Figure 1 The block diagram of the control system of the method shown is shown.

[0027] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0028] Figure 4 This is a schematic diagram of a computer-readable storage medium provided in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0030] This embodiment provides a semiconductor wet cleaning process control method based on multi-source time-series data fusion and AI prediction. The process flow is as follows: Figure 1 As shown. An exemplary structure of a control system that performs this method is as follows. Figure 2 As shown. The following will combine... Figure 1 The steps of this method will be explained in detail.

[0031] like Figure 1 As shown, the method mainly includes the following steps: S10, Multi-source time-series data acquisition steps.

[0032] In this step, the system synchronously acquires multi-source timing data during the semiconductor wet cleaning process at a high sampling frequency (e.g., not less than 1 kHz). This data is not isolated, static parameters, but rather a strictly time-aligned set of signals that dynamically reflects the wafer-fluid-mechanical interaction process. The acquired data specifically includes the following categories: 1. Motion parameters reflecting the mechanical motion state: This parameter includes at least the real-time rotational speed ω(t) of the wafer carrier stage, where t represents a continuous, absolute time variable. This speed signal ω(t) is acquired in real time by a high-precision rotary encoder mounted on the drive shaft of the carrier stage. The dynamic changes of ω(t) (such as acceleration, deceleration, and constant speed maintenance) are one of the core control variables affecting the distribution of the cleaning liquid film, shear force, and drying effect.

[0033] 2. Fluid parameters reflecting the fluid supply status: These parameters include at least the real-time supply flow rates of the cleaning fluid and drying gas to the wafer surface.

[0034] For the cleaning fluid, based on the process formulation, the instantaneous flow rate of various process liquids is acquired in real time using a high-precision mass flow controller or its built-in sensor installed on the corresponding supply pipeline. For example, in the standard cleaning (RCA cleaning) process, steps SC1 and SC2 are performed sequentially, followed by deionized water (DIW) rinsing and isopropanol (IPA) drying. Correspondingly, the system collects: Instantaneous supply flow rate Q of the cleaning fluid used in step SC1 sc1 (t); Instantaneous supply flow rate Q of the cleaning fluid used in step SC2 sc2 (t); Instantaneous supply flow rate Q of deionized water (DIW) diw (t); Instantaneous supply flow rate Q of isopropanol (IPA) ipa (t).

[0035] For dry gases (usually high-purity nitrogen N2), the supply flow rate is also obtained in real time through a mass flow controller. Here, the supply flows Q mentioned above will be... sc1 (t), Q sc2 (t), Q diw (t), Q ipa (t), The combination of is denoted as Q(t).

[0036] These flow signals are the "input recipe" of the process, and their stability, switching timing and dynamic response directly affect the chemical reaction rate and physical cleaning effect.

[0037] 3. Electrical parameters reflecting the process load (key data source of this invention): These parameters include at least the load current or torque signal of the drive motor that drives the wafer carrier stage to rotate. This is the core feature that distinguishes this invention from schemes that only monitor conventional process parameters.

[0038] Specifically, the three-phase stator current I of the drive motor (usually a servo motor or a brushless DC motor) is acquired in real time at a high sampling rate (usually no less than the sampling rate of the speed loop) through the servo driver or dedicated current sensor of the drive motor. A (t), I B (t), I C These two signals, ω(t) and ω(t), together constitute the load current signal I(t) of the drive motor. This current signal directly reflects the real-time electromagnetic torque required by the motor to maintain the set speed ω(t).

[0039] Optionally, the output torque τ(t) at the motor shaft end can be estimated in real time in the processor based on the current signal and the motor model.

[0040] Technical Principles and Innovation Connections: During rotary spray cleaning, the motor load is not constant. When the wafer is immersed in the liquid film, the liquid film morphology changes (e.g., from a continuous film to discrete droplets), or it encounters scrubbing resistance, the hydrodynamic and mechanical friction loads that the motor needs to overcome undergo characteristic changes. These microscopic, difficult-to-observe physical processes are encoded in real-time in the amplitude, fluctuation frequency, and timing waveform of the load current signal I(t) or torque τ(t). For example, when the liquid film provides stable coverage, the load current exhibits a high steady-state value; at the end of the spin-drying process, the load current experiences a characteristic small step drop the instant the last droplet detaches from the wafer. Therefore, this electrical parameter is an information-rich "implicit" observation variable connecting macroscopic control commands (rotation speed) with the microscopic physical state of the process.

[0041] To ensure strict time alignment of the multi-source time-series data, this embodiment achieves synchronous acquisition through the following method: The signal outputs of all sensors (encoders, various mass flow controllers, and motor current sensors) are connected to a central data acquisition unit. This central data acquisition unit is driven by a high-precision clock source (such as a GPS phase-locked clock or a temperature-controlled crystal oscillator) to synchronously sample all input channels with a unified sampling clock, or to add a high-precision timestamp based on this clock source to the data of each channel. Thus, the system obtains a time-aligned, multi-source time-series dataset {ω(t), Q} that can be used for subsequent feature extraction and fusion analysis. sc1 (t), Q sc2 (t), Q diw (t), Q ipa (t), , I(t)}.

[0042] The technical effect achieved through this step is as follows: By synchronously acquiring motion, fluid, and key electrical timing data, and utilizing a high-precision synchronization mechanism to ensure strict time alignment, the system obtains a raw dataset that comprehensively, dynamically, and without delay characterizes the microscopic physical state (liquid-solid interaction, fluid load changes) of a single wafer wet cleaning process. This provides a unique and reliable data foundation for subsequent in-depth data fusion analysis and intelligent decision-making, fundamentally different from existing technologies that only monitor static environmental parameters or rely on batch statistical data. It is the primary prerequisite for realizing single-chip, real-time, process-embedded intelligent control.

[0043] S20, Feature Extraction and Online Virtual Measurement Steps.

[0044] This step is the core of the present invention to achieve intelligent predictive control. Its purpose is to overcome the technical limitations of only being able to detect the cleaning effect offline and after the fact. By deeply analyzing and integrating the real-time process data obtained in step S10, the online and real-time prediction of the cleaning effect of the current process can be achieved (i.e., "virtual measurement"), thereby providing a direct decision-making basis for subsequent real-time process control.

[0045] Specifically, this step receives the time-strictly aligned multi-source time-series dataset {ω(t), Q} output from step S10. sc1 (t), Q sc2 (t), Q diw (t), Q ipa (t), First, define a sliding time window with a window length of T. win The sliding step size is T step The window index is denoted as k. For each window index k, its corresponding absolute time interval is [t]. start (k), t end (k)]. For the load current signal I(t), take its value in the time interval [t]. start (k), t end The signal segment within [k] is denoted as I. k Similarly, the rotational velocity segment ω is obtained. k and each flow signal segment Q sc1,k Q sc2,k Q diw,k Q ipa,k as well as In subsequent steps, "within the k-th window" refers to the aforementioned signal segment I. k Q sc1,k Q sc2,k Q diw,k Q ipa,k as well as The T win The selected value should be sufficient to cover the key dynamic processes in the cleaning process (such as the liquid film oscillation period), for example, 50ms to 200ms (in this embodiment, T is used as the reference value). win = 100 ms as an example); T step The selection of T requires a trade-off between time resolution and computational complexity; for example, it can be taken as 5ms to 10ms (in this embodiment, T is used as T). step (For example, a sliding window length T = 10 ms). win The selection of the frequency range should meet the following engineering principles: 1) Frequency domain analysis requirements: The window should be long enough to include at least one complete cycle of the lowest frequency component to be analyzed (e.g., the lower limit of the low-frequency band of interest in this invention is 10Hz), thereby providing a stable frequency resolution for subsequent spectrum analysis (FFT). 10Hz corresponds to a period of 100ms, therefore T win 1) The time constant should not be less than 100ms. 2) Process dynamic capture requirements: The window should be able to cover the typical time constants of key dynamic processes of the target process (such as liquid film oscillation, droplet detachment during spin drying), which are typically in the tens to hundreds of milliseconds range. 3) Real-time requirements: The window should not be too long to avoid introducing excessive control delay. Considering all factors, T... win The value range is typically from 50ms to 500ms. In this embodiment, T is selected. win =100ms is considered a relatively optimal balance point, as it can effectively analyze frequency components above 10Hz while maintaining a 10Hz update rate (corresponding to T). step =10ms) for rapid response to process changes. Those skilled in the art can adjust T based on the specific equipment's dynamic characteristics and signal conditions, within this principle. step The specific value.

[0046] All subsequent features are calculated based on the data within the k-th window. The specific steps are as follows: S201. Temporal Feature Extraction: Its core objective is to deeply process the multi-source time-series dataset collected in step S10, extracting a set of information-dense features that can quantitatively characterize different aspects of the physical state of the cleaning process. The specific extraction process follows different technical paths depending on the parameters: 1. Multi-dimensional feature extraction for electrical parameters: The load current signal I(t) is the most direct dynamic observation variable reflecting the interaction of the "wafer-fluid-mechanical" system. To analyze the process information it contains from different dimensions, this embodiment performs a triple parallel analysis: Short-time average load feature extraction: To characterize the smooth changes in overall mechanical load during the cleaning process (such as load steps caused by liquid film formation and detachment), the I value within the k-th window is extracted. k Calculate its root mean square value, denoted as I.rms (k). I rms The (k) sequence quantifies the macroscopic average level of the motor output torque and is a key indicator for distinguishing different process stages such as "immersion," "rinsing," and "high-speed rotation drying." k is a positive integer index, representing the number of the k-th time window after the continuous time signal is divided into sliding windows.

[0047] Frequency domain energy distribution feature extraction: To capture high-frequency load fluctuation components caused by fluid turbulence, unstable flow, or mechanical vibration, the energy distribution feature within the window is extracted. k After the data is windowed (e.g., Hanning window), a Fast Fourier Transform is performed, and the power spectral density PSD(k, f) of the signal within the window is estimated based on the Periodogram method, where k: the index (integer) of the sliding window, used to identify the k-th time segment of the signal being analyzed; f: the time frequency (a continuous variable, unit: Hz), representing the rate of periodic change of the signal.

[0048] Subsequently, based on prior analysis of the typical dynamic characteristics of rotary cleaning equipment, three characteristic frequency bands are defined (the frequency bands are the range of values ​​for frequency f): Low frequency band (LF): f ∈ [10, 100] Hz; mid frequency band (MF): f ∈ [100, 500] Hz; high frequency band (HF): f ∈ [500, 1000] Hz. The low, mid, and high frequency bands correspond to the frequency bands in the load current signal that reflect rotating fundamental harmonics and macroscopic flow field disturbances, the frequency bands that reflect liquid-solid interactions and medium-scale turbulence, and the frequency bands that reflect high-frequency micro-vibrations of machinery, respectively.

[0049] The division of the low-frequency band (10-100 Hz), mid-frequency band (100-500 Hz), and high-frequency band (500-1000 Hz) is based on spectral analysis and process calibration results from a large amount of experimental data. Specifically: Low-frequency band (10-100 Hz): This band mainly covers the fundamental frequency of rotation and its directly related low-frequency disturbance energy. Within a typical rotational speed range (e.g., 100-1000 rpm, corresponding to a fundamental frequency of ≈1.7-16.7 Hz), the fundamental frequency and its 2nd-3rd order harmonics primarily fall within this band. The energy E in this band... low The stationarity of (k) (e.g., the coefficient of variation of its sliding window statistic remains below a threshold) is positively correlated with the stability of the macroscopic flow field (e.g., the overall coverage of the liquid film).

[0050] Mid-frequency band (100-500 Hz): This is a comprehensive frequency band sensitive to various dynamic processes. Experiments show that strong liquid-solid interactions (such as jet impact and unstable liquid film rupture) can induce broadband disturbances in this band. Simultaneously, this band often covers the characteristic frequency range of early failures in rotating machinery (such as bearings). Therefore, E mid (k) anomalies need to be combined with other characteristics for comprehensive diagnosis. In this invention, it is mainly used as a sensitive indicator of "the severity of the process or potential changes in the mechanical state".

[0051] High-frequency band (500-1000 Hz): This band is mainly used to monitor the resonant frequency components of mechanical structures and high-frequency electronic noise in drive systems. The upper limit of 1000 Hz is based on engineering considerations of the current loop bandwidth and sampling frequency of typical servo systems. The energy E in this band... high An abnormal increase in (k) (such as the appearance of discrete spectral peaks) may indicate resonance of mechanical components (such as loose components) or abnormal switching noise of power devices.

[0052] The aforementioned judgments of "stable" and "abnormal growth" are all based on comparisons with a statistical baseline established using historical normal production data. This baseline is determined through E... low (k), E mid (k), E high (k) is obtained through statistical analysis (e.g., calculating the mean and standard deviation).

[0053] The total energy of each frequency band is calculated by integrating the spectral energy. Taking the low-frequency band as an example, its energy E low (k) is calculated using the following formula: Where Δf is the frequency resolution after FFT transformation (Δf = sampling frequency fs / number of sampling points N within the window). Similarly, E can be calculated. mid (k) and E high (k).

[0054] Rotating harmonic component characteristics: based on the rotational speed ω at the current window center time. k Calculate the rotating fundamental frequency f rot (k) =ω k / (2π). Dividing by 2π here is the standard physical conversion that converts the angular velocity unit rad / s to the frequency unit Hz.

[0055] Extract the fundamental frequency f from PSD(k, f). rot (k), second harmonic 2*f rot (k), Third harmonic 3*f rotThe amplitude at (k) (or the corresponding energy within the narrow band) yields the harmonic amplitude characteristics A1(k), A2(k), and A3(k), which are mainly used to monitor the periodic symmetry of the rotating system. For example, when there is a slight imbalance in wafer clamping, the amplitude of the second harmonic A2(k) will increase significantly (up to a factor of two), which is strongly correlated with the results of offline dynamic balancing.

[0056] 2. Feature extraction for motion and fluid parameters: To enhance the system's early detection capability of changing process conditions, dynamic characteristic calculations are performed on motion and fluid parameters: With respect to rotational speed ω k The real-time angular acceleration α(k) is calculated as α(k) = [ω(k) - ω(k-1)] / Δt, where Δt is the sampling time interval. α(k) reflects the dynamic performance of the drive system's response to speed commands, and its anomalies can indirectly reflect sudden changes in load.

[0057] Supply flow rate Q for each fluid sc1,k Q sc2,k Q diw,k Q ipa,k as well as The average flow rate change rate is extracted as a feature, denoted as R. sc1,k , R sc2,k , R diw,k ,R ipa,k as well as The "average flow rate change rate" aims to quantify the supply trend of the fluid within the current time window. A basic calculation method is to take the start and end values ​​of the flow rate signal segment within the k-th time window, calculate the difference, and then divide it by the window length T. win , that is, R(k)=[Q(t end )-Q(t start )] / T win .

[0058] Furthermore, to ensure the robustness of this feature and suppress the interference of measurement noise, a better calculation method can be adopted. Before calculation, the original flow signal Q(t) can be preprocessed by low-pass filtering to retain process dynamics and filter out high-frequency noise. Then, linear least squares fitting is performed on the flow data points within the filtered window, and the slope of the resulting fitted line is used as the average flow rate change feature R(k) of the window. This method, by fitting the entire data rather than relying on the start and end points, can effectively resist data fluctuations and obtain more stable feature values ​​that better reflect the true trend of change. This feature R(k) has a clear process indication significance: during the transition phase of formula switching, its absolute value is large, indicating a rapid increase or decrease in flow; during the steady-state supply phase, its value should approach zero, indicating stable flow; unexpected continuous non-zero values ​​or abnormal fluctuations can indirectly indicate problems such as abnormal flow control loops, pipeline pulsation, or slight leakage in valves. Combining this feature with the spectral characteristics of the load current can provide the AI ​​model with richer information about the correlation between "input disturbance" and "system response".

[0059] S202, Multi-source feature fusion and instantaneous feature vector construction: After completing the feature extraction for various signals in step S201, this sub-step aims to integrate the scattered features with different physical meanings into a unified, high-dimensional digital representation to provide input for subsequent AI models.

[0060] The specific integration and construction process is as follows: 1. Time Base Alignment: All features fused in this step are calculated based on the same sliding window index k. This means that feature I rms (k), E low (k), α(k), R sc1,k etc., describes the same time period (i.e., the k-th sliding window [t]). start (k), t end (k) The states of different aspects of the process system. Through a unified window mechanism, strict synchronization of all features in the time dimension is ensured, avoiding feature misalignment caused by signal processing delays or asynchronous sampling times.

[0061] 2. Feature Scalar Normalization: To eliminate the potential impact of differences in the dimensions and orders of magnitude of different features on AI model training and prediction, feature scalar normalization can be performed on each feature sequence (such as the I values ​​of all windows) before constructing the feature vector. rms E for all windows lowEach feature (e.g., the mean and standard deviation of the training set) is standardized. For example, Z-score standardization is used to ensure that each feature follows a distribution with a mean of 0 and a standard deviation of 1 across the entire training set. In online applications, the mean and standard deviation determined from the training set are used to apply the same scaling transformation to the real-time features. This step is a standard data preprocessing technique aimed at improving the model's convergence speed and generalization performance.

[0062] Constructing the instantaneous feature vector F(k): For each window index k, all the feature scalars that have undergone the above extraction and normalization operations are arranged and concatenated in a preset, fixed order to form a high-dimensional column vector, which is defined as the instantaneous feature vector F(k).

[0063] Example of the specific structure of F(k): F(k) = [I rms (k), E low (k), E mid (k), E high (k), A1(k), A2(k), A3(k), ω k ,α(k),Q sc1,k , R sc1,k Q sc2,k , R sc2,k Q diw,k , R diw,k Q ipa,k , R ipa,k , , ] T Where T represents transpose; The first 7 dimensions of the vector (I rms (k) to A3(k) are derived from the deep analysis of the load current signal I(t), representing the average load, spectral energy distribution and rotating harmonics, respectively.

[0064] The subsequent two dimensions (ω(k), α(k)) are derived from the motion parameters, representing the rotational state and its dynamics.

[0065] The remaining dimensions appear in pairs (e.g., Q). sc1,k With R sc1,k Each pair corresponds to a fluid parameter, which respectively characterizes the instantaneous supply level of the fluid and its changing trend.

[0066] The total dimension D of the vector is equal to the number of all features. The value of D is not fixed and can be adjusted according to the complexity of the actual cleaning process (such as the types of chemicals used) and the number of features selected.

[0067] This sub-step achieves the following technical effect: Raw time-series signals from multiple sources such as motion, fluid, and electrical sensors are transformed into a structured digital feature vector F(k) that integrates time, frequency, and differential information within the same time slice. This vector F(k) is the direct and sole input for subsequent online virtual measurement by the AI ​​model. It is not simply a list of raw data, but rather, through targeted feature engineering, it purifies and encodes the microscopic dynamics of the process, enabling the AI ​​model to more efficiently and accurately establish the mapping relationship between observable process signals and indirect cleaning quality. This lays the core data foundation for the precise predictive control achieved in this invention.

[0068] S203. Online virtual measurement based on AI models: This sub-step follows step S202, and maps the feature vector F(k) representing the process dynamics into a quantitative prediction value of the final cleaning effect in real time through a specially trained neural network model that can capture time-series dependencies, thereby realizing online virtual measurement of process quality.

[0069] 1. Structure and Input / Output of Neural Network Models To process the feature vector sequence F(k) with strong temporal correlation and accurately model the dependency relationship between the current state and historical states in the cleaning process, this embodiment adopts a deep learning model architecture based on a Long Short-Term Memory (LSTM) network. The LSTM network is an improved version of a recurrent neural network, containing input gates, forget gates, and output gates, which can effectively learn long-term dependencies and is very suitable for mapping modeling from dynamic process sequences to final quality results in this scenario.

[0070] Model Input: The model input is a sequence of instantaneous feature vectors. Considering the continuity of the process state, the system combines the feature vectors of the most recent M consecutive time windows into a sequence. Specifically, the input feature sequence S(k) at time k is defined as: S(k) = [F(k-M+1), F(k-M+2), ..., F(k)]. Here, M is the preset sequence length (M≥2, e.g., M=10). S(k) is a sequence containing M elements, where the i-th element F(k-M+i) corresponds to the feature vector of the (k-M+i)-th time window. This sequence covers the continuous time period from the (k-M+1)-th window to the current k-th window.

[0071] Model architecture: Input layer: Receives a feature sequence S(k) with dimensions (M, D), where M is the sequence length and D is the dimension of the feature vector F(k).

[0072] Embedding layer: First, a fully connected layer is used to perform nonlinear transformation and dimension adjustment on F(k) at each time step, mapping the original features to a more expressive latent space. The output of this layer is still a sequence of (M, D'), where D' is the embedding dimension.

[0073] Temporal Modeling Layer: This layer consists of stacked LSTM layers that take the above sequence as input. Each LSTM unit processes the sequence S(k) sequentially step by step. The forget gate determines how much information to retain from the previous cell state, the input gate controls how much information from the current input needs to be updated to the cell state, and the output gate determines the output for that time step based on the current cell state. Through this gating mechanism, the LSTM can selectively remember and pass on important historical information. After processing by the last LSTM layer, the hidden state vector h(k) of the last time step (corresponding to window k) is taken as a summary representation of the entire input sequence, which encodes all relevant process dynamic information up to window k.

[0074] Regression Output Layer: The final hidden state h(k) output by the LSTM layer is input into one or more fully connected layers. The last fully connected layer has the same number of neurons as the number of cleaning effect metrics to be predicted, and uses a linear activation function to directly output the predicted cleaning effect value.

[0075] Model Output: The model output is the cleaning effect prediction vector Y corresponding to the current input sequence S(k) (i.e., the process dynamics up to window k). pred (k). For example, Y pred (k) = [P pred (k), θ pred [(k)], where P pred (k) represents the predicted number of particles per unit area, θ pred (k) represents the predicted water contact angle. These outputs are specific, quantifiable physical quantities. During online operation, the model uses forward propagation to calculate the input F(t) in real time for each moment, continuously outputting predicted cleaning effects of the current process, such as P. pred (t) and θ pred (t). This means that while the cleaning process is still in progress, the system can dynamically "predict" the final cleaning quality that the current process state will result in. This achieves true "online virtual measurement," completely solving the lag problem of relying on offline detection pointed out in the background technology.

[0076] 2. Model Working Principle and Technical Mapping The LSTM temporal modeling layer dynamically models the evolution of the process state (represented by the feature vector F(·)) over time by progressively processing the input feature sequence S(k). Its internal gating mechanisms (forget gate, input gate, output gate) enable the model to selectively memorize key state information (such as the initial contamination load characteristics of stage SC1) that has long-term relevance to the prediction of the current cleaning effect within the historical window, while filtering out irrelevant transient disturbances. Finally, the hidden state h(k) output at the last time step is essentially a high-dimensional condensed representation of the process state, integrating all dynamic process information from the (k-M+1)th window to the current kth window.

[0077] Subsequently, the fully connected regression layer maps this abstract, high-dimensional process state representation h(k) to a specific, physically measurable space of cleaning effect indicators. This mapping relationship is h(k)→Y. pred (k) is precisely the approximation of the complex nonlinear function between multi-source temporal process features and final cleaning quality, learned by the model during the training phase from massive "process-outcome" paired data. Therefore, Y pred (k) is not simply calculated, but rather is an inference made by the model about the cleaning effect in the future (when the process ends) based on an understanding of the entire process history up to the current moment.

[0078] 3. Model training process and sample composition The neural network model is trained in a supervised manner using massive amounts of historical production data during the offline phase.

[0079] Training sample composition: Each sample is a historical feature sequence S(i) with a label Y. true (i) is the offline inspection result of the corresponding wafer after cleaning (e.g., [P]). true (i), θ true (i)]).

[0080] Training process: minimizing the predicted value Y pred (i) and the true value Y true The model parameters are optimized using the backpropagation algorithm with the error (e.g., mean square error) between (i) as the target.

[0081] 4. Engineering deployment and real-time performance assurance of the model To achieve millisecond-level real-time prediction, this invention performs engineering optimization on the trained LSTM model. The model uses the following parameters to ensure a balance between performance and accuracy: sequence length M = 10 (covering 100ms of process dynamics), 2 LSTM layers, and 64 hidden units per layer. Testing shows that this model, on an embedded industrial PC equipped with an Intel Core i7-12700 or equivalent computing power, has a single forward propagation inference time of less than 2ms, significantly faster than the data acquisition window step size T. step =10ms, which fully meets the real-time requirements.

[0082] The training data requirements for the model are as follows: Complete cleaning process data and corresponding offline detection results for at least 5000 wafers under different process formulations, different initial contamination states, and different equipment health states must be collected. The data must cover all expected process fluctuation ranges. During training, five-fold cross-validation is used to ensure the model's generalization ability, and Dropout layers (dropout rate 0.2) and L2 regularization are used to prevent overfitting. The optimizer is Adam, with an initial learning rate of 0.001. Training is terminated early when the validation set loss no longer decreases for 10 consecutive epochs. The final model's prediction error (root mean square error, RMSE) on the independent test set (approximately 1000 wafers) must be less than 1.5 times the measurement error of the offline detection equipment itself to be considered acceptable and usable.

[0083] 5. Execution of online virtual measurement During online operation, the system sequentially caches the feature vectors F(k) generated in real time in step S202, forming the current input sequence S(k), which is then input into the trained LSTM model. The model outputs the predicted value Y in real time through forward propagation. pred (k).

[0084] S30. Steps for generating decision recommendations: This sub-step follows step S203, and its purpose is to: based on the online virtual measurement value (i.e. the predicted value of cleaning effect) output by step S203, and combined with the preset process quality target, generate accurate and executable process parameter adjustment instructions in real time through a predefined expert rule base, so as to drive the actuator to actively correct the process that deviates from the target.

[0085] The specific execution process is as follows: Input and Comparison: This sub-step receives the cleaning effect prediction vector Y, which is output in real time from step S203. pred (k) = [P pred (k), θ pred (k)]. The system will use Y pred (k) and the preset target effect vector Y, representing qualified cleaning quality.target = [P target ,θ target Real-time comparison is performed to calculate the prediction bias vector ΔY(k) = Y. pred (k) - Y target Furthermore, its trend of change can be calculated (e.g., the rate of change of ΔY(k) relative to the first few windows).

[0086] Construction and basis of the expert rule base: This invention adopts an architecture of "AI model prediction + expert rule base decision-making," based on a comprehensive consideration of the reliability, interpretability, and engineering implementation cost of the control system. AI models excel at handling high-dimensional, nonlinear mapping problems, but as "black boxes," directly outputting control commands presents engineering risks such as difficulty in reliability verification, unpredictable behavior in extreme cases, and potential abrupt changes in control commands due to slight overfitting. Therefore, this invention strictly limits the function of the AI ​​model to a "state observer" (i.e., virtual measurement), while entrusting the "controller" function to an expert system based on explicit rules. This decoupling of observation and control aligns with the industry's design principles of "simplicity and auditability" for high-reliability systems. The expert rule base is established based on in-depth analysis of a large amount of historical production data and a systematic summary of domain expert experience. Specifically: Data foundation: Collect multi-source time-series data from historical production, corresponding predicted (or actual) cleaning effect values, and records of process parameter adjustments taken by experienced engineers and proven effective under this process condition (e.g., "In SC1 stage, when the predicted particle number decrease slope is lower than X, Q..."). sc1 Increase Y%).

[0087] Rule extraction: By summarizing, analyzing and verifying the correlation between the above-mentioned "process state characteristics (especially the predicted deviation ΔY and its trend)" and "verified effective adjustment actions", a series of deterministic rules in the form of "IF-THEN" are formed.

[0088] Knowledge solidification: These rules are encoded to form a structured, queryable decision rule base. Each rule contains at least a "condition part" (matching specific prediction deviation patterns, process stages, etc.) and an "action part" (outputting specific process parameter adjustment instructions).

[0089] Rule-based decision generation: During online operation, the system will display the current prediction deviation ΔY(k), the deviation trend, and the optional current process stage identifier (which can be determined based on I). rms Features such as (k) and ω(k) are automatically determined, and the expert rule base is queried by taking inputs such as "SC1 stage", "DIW rinsing stage", and "IPA drying stage".

[0090] Rule matching: The system searches the rule base for the rule with the highest matching degree to the current input state. A rule may be formalized as: IF the absolute value of the current predicted particle number deviation |ΔP(k)| (where ΔP(k) = P pred (k) -P target If the deviation is greater than the threshold δ1, AND the rate of change of the deviation d(ΔP) / dt is less than zero (indicating that the deviation is decreasing), then take appropriate action.

[0091] Instruction Generation: Once a rule is matched, its "action part" is triggered, generating a specific, quantifiable process parameter adjustment instruction. This instruction directly targets the controllable parameters defined in step S10, for example: Adjusting motion parameters: ω set (k+1) =ω(k) +Δω (Increase the speed setting value of the next window by Δωrpm).

[0092] Adjusting fluid parameters: Q sc1_set (k+1) = Q sc1 (k) * (1 +η) (Adjust the SC1 flow setting to (1+η) times the current value, where η is the specific adjustment ratio given by the rule, such as 5%).

[0093] Example (in conjunction with the preceding text): Condition: IF Current process stage == "SC1 Chemical Cleaning" AND P pred (k) is higher than P target The amplitude exceeds the threshold δ p AND recent P pred The rate of decrease of (k) is lower than the empirical value v th This means: "If the system determines that it is currently in the SC1 chemical cleaning step, and the number of particles P on the wafer surface is predicted in real time by the AI ​​model..." pred (k) is significantly higher than (i.e., the cleaning effect is worse than) the preset qualified target value P. target Furthermore, the predicted particle number has decreased too slowly recently, failing to meet the acceptable standard within the specified timeframe. Here, "recently" refers to a configurable observation time window, such as the most recent N sliding windows (e.g., N=5, corresponding to 50ms); δ p The value should be determined based on process specifications and control accuracy requirements. For example, it can be set as the qualified specification P. target A certain percentage (e.g., δ) p = 0.2 *P target (or determined based on a controllable range of stable fluctuations in historical data). thIt is a lower bound of the expected rate of decline obtained based on historical high-quality production data.

[0094] Action: THEN generates the instruction: "Increase the supply flow rate of SC1 cleaning fluid Q". sc1 "Increase by 10% in the next control cycle", that is, Q sc1_set (k+1) = 1.10 * Q sc1 (k). This means that the system considers the current SC1 chemical cleaning intensity to be insufficient or inefficient, and needs to intervene immediately to enhance it, thereby automatically generating adjustment commands such as "increase the cleaning fluid flow rate".

[0095] The expert rule base not only includes the rules based on the predicted cleaning effect, but also a series of predefined rules based on process state characteristics, in order to achieve rapid and stable control of the micro-dynamics of the process.

[0096] The basis for constructing such "process state rules" stems from the load current signal I(t) or its characteristics (such as It). rms (k), E mid A deeper understanding of the relationship between (k) and specific process physical states. For example, historical data analysis can determine that during a stable liquid film flushing phase, I rms (k) should be maintained within a reasonable range [I] low , I high [Inside; or, when the intermediate frequency energy E of the load current] mid (k) Abnormal sudden increases often correspond to unstable turbulence or minor mechanical collisions in the fluid flow. Here, the above is only based on specific characteristics of the load current signal (I...). rms (k), E mid (k) Two examples of constructing process state rules. Within the framework of this invention, the "rules based on process state characteristics" can be constructed using any or more feature values ​​extracted from the load current signal I(t). These feature values ​​include, but are not limited to, its root mean value It. rms (k) Spectral energy E in the low, medium, and high frequency bands low (k), E mid (k), E high (k), and the amplitudes of each harmonic related to the current rotating fundamental frequency, such as A1(k), A2(k), A3(k).

[0097] The specific form of the rule is as follows: when a certain characteristic value (or a combination of several characteristic values) deviates from its preset target range (or target curve), a corresponding compensatory adjustment command is triggered. The "preset target range" is obtained through statistical analysis of corresponding characteristic values ​​in a large amount of historical normal process data (e.g., taking the mean ± 3 times the standard deviation as the control limit), or determined based on the process physical model and safety boundaries. For example, to prevent mechanical resonance, E can be set... high The absolute upper limit of (k); to control the stability of the liquid film, E can be set. mid The upper limit of the growth rate of (k) can be set; in order to monitor the rotational equilibrium state, the upper limit of the ratio of the second harmonic amplitude A2(k) to the fundamental amplitude A1(k) can be set.

[0098] A typical example of a process status rule is as follows: Condition: IF current process stage == "DIW rinsing" AND real-time load current RMS value I rms (k) Continuously exceeds the upper limit I of the preset target range high More than 3 consecutive windows.

[0099] Physical meaning: This condition indicates that the current rinsing process is under excessive load, which may be due to excessive DIW water flow leading to abnormally increased fluid resistance, or abnormal contaminants on the wafer surface increasing friction.

[0100] Action: THEN generates adjustment command: "Increase the supply flow rate of DIW rinsing solution Q..." diw (k) Reduce by 5% in the next control cycle, i.e., Q diw_set (k+1) = 0.95 * Q diw (k). Additionally, a monitoring command can be attached, such as if I... rms If (k) still does not fall back, an early warning will be triggered.

[0101] By combining such direct judgment rules based on process signals with forward-looking rules based on quality prediction, the expert rule base can simultaneously achieve closed-loop optimization of quality objectives and rapid and stable control of process states. The two work together to ensure the smooth and efficient operation of the process.

[0102] Output and anti-oscillation processing: The generated adjustment commands will be sent directly to the process execution unit. To ensure the stability of the control system, the rule base design or post-processing logic usually includes anti-oscillation mechanisms, such as setting a minimum time interval for command output and cumulatively limiting the continuous adjustment amplitude of the same parameter within a short period of time, to prevent over-adjustment of commands from causing system oscillation.

[0103] The technical effect achieved through step S30 is as follows: by using an expert rule base built upon a large amount of historical experience and data, the forward-looking "quality prediction" problem based on AI models is transformed into a real-time, executable "process control" problem based on explicit rules. This bridges the gap between "prediction" and "control," enabling the system to automatically generate and execute optimal process parameter adjustment strategies based on real-time predictions of the final cleaning effect. This method not only boasts fast response speed, low computational overhead, and low hardware implementation cost, but also offers a transparent and highly interpretable decision-making process, perfectly meeting the high requirements of industrial sites for reliability, stability, and maintainability. Ultimately, it achieves closed-loop predictive adaptive control of the cleaning process, significantly improving process consistency and product yield.

[0104] S40. Process parameter adjustment command execution and feedback control This step follows step S30 and is the final execution stage for realizing the "prediction-control" closed loop in this invention. Its core objective is to transform the process parameter adjustment instructions generated in step S30 into precise control signals for the physical actuators, and through a feedback adjustment mechanism, drive the dynamic optimization of the cleaning process system to achieve the predetermined process quality and process state targets.

[0105] Specifically, this step receives a process parameter adjustment command from step S30, such as the speed setpoint ω for the next control cycle. set (k+1), Flow rate setpoint Q for each cleaning fluid or gas set (k+1), etc. These instructions are specific and quantifiable.

[0106] 1. Issuance and execution of control commands The adjustment commands are sent to the corresponding process execution units in real time via the corresponding industrial bus (such as EtherCAT, PROFINET) or analog / digital output interface: Control of motion parameters: The speed adjustment command ω set (k+1) is sent to the servo driver or frequency converter that drives the wafer carrier stage to rotate. The driver uses a built-in speed loop / position loop control algorithm to drive the actual speed ω(t) of the motor to track the set value quickly and smoothly.

[0107] Control of fluid parameters: This involves adjusting various flow rate commands Q. sc1_set (k+1), Q diw_set (k+1), Q ipa_set (k+1) The data, etc., are sent to the corresponding mass flow controller (MFC). Based on the received setpoints, the MFC precisely controls the flow rate of the medium into the process chamber by adjusting the opening of its internal control valve, so that the actual flow rate approaches the setpoint.

[0108] 2. Dual control objectives and adaptive adjustment process The execution of this step is not a simple "open-loop execution of instructions," but rather an adaptive process guided by the goal of achieving dual control objectives: Primary objective: Quality control – to achieve the predicted cleaning effect value Y. pred (k) Approaching the target effect value Y target .

[0109] Control Logic: This is the highest-level control objective of this invention. The execution effect of step S40 will be continuously monitored through the "process-sensing" link (i.e., step S10). Specifically, after executing new process parameters, the state of the process system changes, which will be reflected in the multi-source time-series data collected in subsequent windows, the extracted features, and the predicted output Y of the AI ​​model. pred (k+1), Y pred (k+2)...up.

[0110] Adaptive behavior: The system continuously compares the latest predicted value Y. pred With the target value Y target If the deviation ΔY does not decrease or even increases, it means that the current adjustment strategy (generated by the S30 rule base) may be inapplicable or insufficient. This will trigger a new round of the "sensing-predicting-decision" cycle (S20 step), which may generate different and better adjustment instructions based on the new process state. Through this iterative "execution-observation-re-decision" mechanism, adaptive and optimal control with the final cleaning quality as the goal is achieved.

[0111] Secondary objective: Process state control -- to make the change curve of the load current I(t) or torque signal τ(t) of the drive motor approach the preset process target curve.

[0112] Control Logic: This is a parallel, faster-responding control loop. As mentioned earlier, the timing characteristics of I(t) are strongly correlated with the micro-process conditions. An ideal "process target curve" defines the timing characteristics of I(t) (or its characteristics such as It) during the optimal cleaning process. rms (t) should exhibit the expected trajectory of change over time (e.g., the root mean square value of the load current I during the liquid film formation stage). rms (t) should show a smooth upward trend and eventually stabilize near a characteristic plateau value corresponding to the current process parameters (such as rotational speed and flow rate). This plateau value characterizes the typical load level when the liquid film has uniform and stable coverage; while it should decrease in a stepwise manner during the drying stage.

[0113] Adaptive behavior is manifested in the system's real-time monitoring of the actual curve of I(t) while executing adjustment commands. If the actual curve deviates from the target curve (e.g., I...), the system will adjust accordingly. rms(The rise is too slow), even if the predicted value Y of the cleaning effect is at this time. pred Even before any significant deviation is observed, the system can generate fine-tuning instructions (such as a slight increase in speed or flow rate) in advance (using rules preset in the S30 rule base that are specific to the process state), allowing the process state to return to normal beforehand and thus preventing deviations in final quality. This embodies the control concept of "feedforward compensation based on process state".

[0114] The coordination and arbitration mechanism of dual control: The "quality target control" and "process status control" are not completely independent; their output instructions are coordinated through a preset priority and fusion logic. Safety First: Any measures from "process status control" designed to prevent equipment overload (such as I) rms (Exceeding limits) or process abnormalities (such as E) high The instruction for sudden increase in warning of mechanical failure has the highest priority and will be executed immediately, while potentially conflicting "quality target control" instructions will be temporarily suspended.

[0115] Steady-state optimization: During normal process stages without safety warnings, the "quality target control" directives take precedence. The role of "process state control" is to perform feedforward fine-tuning; for example, when I... rms (k) When the speed is slightly below the target curve, the speed setpoint can be slightly increased. This fine adjustment will be linearly superimposed on the speed adjustment given by "Quality Target Control". The fine adjustment coefficient is determined based on the "process characteristic-quality" sensitivity relationship calibrated in historical data.

[0116] Conflict prevention and smoothing: If the two loops give opposite adjustment directions, the system will prioritize the instruction with the smaller amplitude and enter a more frequent monitoring cycle (e.g., T). step (Temporarily halved) to quickly determine the trend of changes in process status, thereby making a more accurate arbitration.

[0117] Through the aforementioned collaborative mechanism, the system ensures continuous adaptive optimization toward the optimal quality goal under safe and stable conditions.

[0118] 3. Stability and security assurance mechanisms for online regulation To ensure the stability and reliability of online rotation speed adjustment and avoid introducing destructive disturbances, this invention employs the following collaborative mechanism: Decision constraints: The speed adjustment commands generated in the expert rule base are based on effective adjustment strategies verified in historical process data, and are associated with specific process stages. For example, in the sensitive stage of initial liquid film formation, the rule base may prohibit large speed increases or only allow very small adjustments.

[0119] Softening is performed: The speed setpoint ω is adjusted. setWhen (k+1) is sent to the servo drive, the target value is not given directly, but rather through a rate limiter. This limiter ensures that the actual rate of change of rotational speed (angular acceleration) does not exceed a preset safety threshold (e.g., the maximum acceleration / deceleration rate is limited to a smooth range that the mechanical structure and fluid dynamics of the equipment can withstand), thereby generating a smooth speed transition curve and avoiding the impact of abrupt changes on the flow field.

[0120] Real-time monitoring and safety arbitration: This is the core feedback loop to prevent instability. After speed adjustment is executed, the load current signal I(t) is continuously monitored at a high frequency. Real-time calculations are performed, such as the mid-frequency energy E. mid (k) and high-frequency band energy E high (k) and other key process characteristics. The process state control loop continuously compares these characteristic values ​​with safety thresholds characterizing a "stable process." If E is detected... mid (k) Abnormal surge (indicating intensified turbulence) or E high (k) If a resonance peak appears (indicating mechanical vibration), the loop will immediately generate a high-priority "stabilize" or "revert" command, temporarily overriding the optimization command of the quality target control loop until the process state returns to stability. This rapid feedback based on the most direct physical signal (current) constitutes the inherent stabilizer of the control system.

[0121] Through the aforementioned three-tiered mechanism of "intelligent decision-making, gentle execution, and rigorous monitoring," the online speed adjustment achieved by this invention is a controlled and adaptive process optimization method that can effectively improve cleaning quality while ensuring the overall stability and safety of the process.

[0122] The technical effect achieved through step S40 is that it transforms intelligent decision-making instructions into precise process control, and constructs an adaptive and highly robust predictive control system through a dual closed loop of "quality-oriented" and "process-oriented" approaches. Specifically, "quality-oriented" approaches are based on real-time prediction of the final cleaning effect for strategic optimization, ensuring quality standards are met; "process-oriented" approaches, based on directly and quickly measurable electrical signals I(t), provide immediate compensation for microscopic dynamic disturbances in the process, offering a stable and reliable process foundation for quality optimization. These two approaches constitute a control architecture that combines virtual and real-world operations and works collaboratively. Ultimately, through the closed-loop operation of steps S10, S20, S30, and S40, this invention upgrades traditional cleaning equipment into intelligent equipment with "real-time perception, intelligent prediction, and autonomous optimization" capabilities, fundamentally solving the problems of insufficient process adaptability and lagging quality control, and significantly improving the stability, uniformity, and product yield of the cleaning process.

[0123] Please see Figure 2This is a schematic diagram of the architecture of a semiconductor wet cleaning process control system based on multi-source time-series data fusion and AI prediction provided in an embodiment of the present invention.

[0124] Based on the method described in Embodiment 1 above, this embodiment provides a semiconductor wet cleaning process control system based on multi-source data fusion and AI prediction. For example... Figure 2 As shown, the system includes: a multi-source data acquisition module 100, a feature extraction and virtual measurement module 200, a decision suggestion generation module 300, and a process execution and feedback control module 400. These modules are connected sequentially and work collaboratively to completely and accurately execute the method flow described in Embodiment 1.

[0125] A multi-source data acquisition module 100 is used to execute step S10 of Embodiment 1, namely, the multi-source timing data acquisition step. It is configured to synchronously acquire multi-source timing data during the semiconductor wet cleaning process at a high sampling frequency, and ensure strict time alignment of the data. The module 100 includes at least: A high-precision rotary encoder is used to acquire the real-time rotational speed of the wafer carrier stage.

[0126] Multiple high-precision mass flow controllers or their built-in sensors are used to collect the real-time supply flow rates of cleaning fluids (such as SC1, SC2, deionized water DIW, and isopropanol IPA) and drying gases (such as N2) directed to the wafer surface.

[0127] A servo driver current sensor or a dedicated current probe is used to acquire the load current signal I(t) of the drive motor that drives the wafer carrier stage to rotate.

[0128] The central synchronous acquisition unit is driven by a high-precision time reference source to ensure that the timestamps of all sensor data are synchronized.

[0129] The output of module 100 is a time-strictly aligned multi-source time-series dataset, which serves as the input for all subsequent processing.

[0130] The feature extraction and online virtual measurement module 200 is used to execute step S20 of Embodiment 1, namely the feature extraction and online virtual measurement step. This module receives multi-source time-series data from module 100 and specifically includes two logical processing units: Temporal feature extraction and fusion unit: Performs sliding window processing on the input data. For each window index k, extracts the short-time root mean square value I from the load current signal I(t). rms (k), band energy E low (k), E mid (k), E highMultidimensional features including (k), harmonic amplitude A1(k), A2(k), A3(k); dynamic features such as angular acceleration α(k) and average flow rate change rate are extracted from motion and fluid parameters; and all features are fused to construct an instantaneous feature vector F(k).

[0131] Online virtual measurement unit: This unit stores a trained deep neural network model (such as an LSTM network). It takes the feature vector sequence S(k) = [F(k-M+1), ..., F(k)] as model input and outputs a cleaning effect prediction vector Y in real time through forward propagation. pred (k) = [P pred (k), θ pred (k)], to realize online virtual measurement of process quality.

[0132] The decision suggestion generation module 300 is used to execute step S30 of Embodiment 1, namely the decision suggestion generation step. This module receives the cleaning effect prediction vector Y from module 200. pred (k). It has an internally pre-stored expert rule base. Module 300 will Y pred (k) and the preset target effect vector Y target The comparison is performed, the prediction deviation ΔY(k) is calculated, and the current process stage identifier (which can be based on I) is considered. rms (k), ω(k), etc.) and process state characteristics (such as I) rms (k) Whether it exceeds the target range), query the expert rule base. Through rule matching, generate specific, quantifiable process parameter adjustment instructions, such as the speed setpoint ω for the next generation control cycle. set (k+1) or the setpoint Q for each fluid flow rate set (k+1).

[0133] The process execution and feedback control module 400 is used to execute step S40 of Embodiment 1, namely the process parameter adjustment command execution and feedback control step. This module receives the process parameter adjustment command from module 300, converts it into a drive signal, and sends it to the corresponding physical actuator, including: The speed adjustment command is sent to the servo driver of the wafer carrier stage to control the actual speed ω(t) to track the set value.

[0134] Each flow regulation command is sent to the corresponding mass flow controller to control the medium supply flow to track the set value.

[0135] Simultaneously, module 400 works in conjunction with modules 100, 200, and 300 to form a dual closed-loop control: based on the Y output of module 200... pred (k) and Y targetThe deviation is used to achieve quality-oriented closed-loop optimization; based on the deviation between the load current I(t) collected in real time by module 100 and the preset process target curve, combined with the process state rules in module 300, process-oriented rapid closed-loop stabilization is achieved. The two work together to drive the system to dynamically optimize.

[0136] The modules 100, 200, 300, and 400 mentioned above can interact with each other through industrial control networks or internal system buses, together forming a complete intelligent cleaning process control system from perception, prediction, decision-making to execution.

[0137] Example 3 Please see Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Figure 3 As shown, the electronic device includes a memory 310, a processor 320, and a computer program stored in the memory 310 and executable on the processor 320.

[0138] When the processor 320 executes the computer program, it implements all the steps of the semiconductor wet cleaning process control method based on multi-source timing data fusion and AI prediction as described in Embodiment 1. The method includes steps S1, S2, S3, and S4, wherein step S2 includes steps S201, S202, and S203.

[0139] The electronic device can be an industrial control computer or embedded controller integrated into a semiconductor wet cleaning device, or it can be a standalone server, workstation, or cloud computing platform used to execute the control method.

[0140] Example 4 Please see Figure 4 This is a schematic diagram of the structure of a computer-readable storage medium provided in an embodiment of the present invention. Figure 4 As shown, a computer-readable storage medium 500 stores a computer program 510 thereon.

[0141] When the computer program 510 is executed by the processor of the electronic device, the electronic device performs all the steps of the semiconductor wet cleaning process control method based on multi-source timing data fusion and AI prediction as described in Embodiment 1. The method includes steps S1, S2, S3, and S4, wherein step S2 includes steps S201, S202, and S203.

[0142] The computer-readable storage medium 500 can be any tangible medium that contains or stores a program, such as ROM, RAM, disk, optical disk, USB flash drive, solid-state drive, etc.

[0143] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A semiconductor wet cleaning process control method based on multi-source time-series data fusion and AI prediction, characterized in that, Includes the following steps: S10. Multi-source timing data acquisition step: Synchronously acquire multi-source timing data during the semiconductor wet cleaning process; the multi-source timing data includes motion parameters reflecting the mechanical motion state, fluid parameters reflecting the fluid supply state, and electrical parameters reflecting the process load; wherein, the motion parameters include the rotational speed ω(t) of the wafer carrier stage; the fluid parameters include the supply flow rates of cleaning liquid and drying gas to the wafer surface; the electrical parameters include the load current signal I(t) of the drive motor that drives the wafer carrier stage to rotate; wherein, t is a continuous time variable; S20. Feature Extraction and Online Virtual Measurement Step: Based on the multi-source time-series data collected in step S10, features are extracted and an instantaneous feature vector F(k) is constructed. The feature sequence S(k) containing the instantaneous feature vector F(k) is input into the trained neural network model, and the cleaning effect prediction vector Y is output in real time. pred (k); where k is the window index after sliding window processing of the signal; S30, Decision suggestion generation step: The cleaning effect prediction vector Y... pred (k) and the preset target effect vector Y target The comparison is performed, and based on the comparison results, a pre-stored expert rule base is queried to generate process parameter adjustment instructions; S40. Process parameter adjustment command execution and feedback control steps: Execute the process parameter adjustment command to adjust the rotation speed of the wafer carrier stage and / or the supply flow rate of the cleaning fluid and drying gas, and realize dual control based on the feedback of the load current signal I(t); the dual control includes quality control based on cleaning effect prediction and process state control based on load current signal.

2. The method according to claim 1, characterized in that, In step S20, the extraction of features and construction of the instantaneous feature vector F(k) specifically includes: The multi-source time-series data is processed using a sliding window method, where for each window index k: For the load current signal I(t), extract its root mean square value I. rms (k) Spectral energy E in the low, medium, and high frequency bands used to characterize different process physical phenomena low (k), E mid (k), E high (k), and the harmonic amplitudes A1(k), A2(k), A3(k) associated with the current rotating fundamental frequency; For the aforementioned motion parameters, extract their angular acceleration α(k); For the fluid parameters, extract the average flow rate change rate; All extracted features are concatenated to form the instantaneous feature vector F(k).

3. The method according to claim 2, characterized in that, In step S20, the neural network model is a long short-term memory network model; the feature sequence S(k) is composed of the instantaneous feature vectors of the current and the preceding M-1 windows, i.e., S(k) = [F(k-M+1), F(k-M+2), ..., F(k)]; where M is the preset sequence length, and M≥2; the cleaning effect prediction vector Y pred (k) includes at least the predicted number of particles per unit area P pred (k) and predicted water contact angle θ pred (k).

4. The method according to claim 2, characterized in that, In step S30, the expert rule base includes rules based on the predicted cleaning effect and rules based on process state characteristics; the rules based on process state characteristics use the root mean square value of the load current signal, the spectral energy in the low, medium and high frequency bands, and / or whether the harmonic amplitude related to the current rotating fundamental frequency exceeds their respective preset target range as one of the triggering conditions.

5. The method according to claim 4, characterized in that, In step S30, generating the process parameter adjustment command includes generating the speed setpoint ω for the next control cycle. set (k+1) and / or fluid flow rate setpoint Q set (k+1).

6. The method according to claim 1, characterized in that, In step S40, The quality control based on cleaning effect prediction includes: based on the cleaning effect prediction vector Y pred (k) and the target effect vector Y target To address deviations, iterative optimization of process parameters is performed. The process state control based on the load current signal includes: generating a compensatory adjustment command based on whether the load current signal I(t) deviates from the preset process target curve.

7. The method according to claim 2, characterized in that, The low, medium, and high frequency bands are pre-divided continuous frequency intervals, wherein: The low-frequency band is the interval where the frequency f satisfies 10 Hz ≤ f < 100 Hz; The mid-frequency band is the interval where the frequency f satisfies 100 Hz ≤ f < 500 Hz; The high-frequency band is the interval where the frequency f satisfies 500 Hz ≤ f ≤ 1000 Hz.

8. A semiconductor wet cleaning process control system based on multi-source time-series data fusion and AI prediction, characterized in that, The system for performing the method according to any one of claims 1 to 7 comprises: The multi-source data acquisition module (100) is used to execute step S10; A feature extraction and online virtual measurement module (200) is used to perform step S20; Decision suggestion generation module (300) is used to perform step S30; The process execution and feedback control module (400) is used to execute step S40.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor of the electronic device, it causes the electronic device to perform the method as described in any one of claims 1 to 7.

Images