Method for adaptive matching of semiconductor material cleaning time and particle size

Abstract

The application belongs to the technical field of semiconductor cleaning and specifically relates to a method for adaptively matching the cleaning time of a semiconductor material with particle size. The method comprises: using a wideband acoustic sensor array to capture acoustic emission signals in real time during the cleaning process; decoupling the mixed signals by wavelet transform and a pattern recognition algorithm to establish a mapping relationship between particle size and acoustic characteristics; dividing the cleaning process into continuous stages corresponding to large, medium and fine particles; dynamically adjusting the megasonic energy output according to the acoustic fingerprint intensity, and determining the cleaning endpoint when the fine particle signal disappears. The application realizes precise matching of energy output and particle size through closed-loop control driven by particle detachment events, protects fine pattern structures, improves production efficiency and process yield, and reduces the consumption of chemical liquids, deionized water and electric energy.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor cleaning technology, specifically relating to an adaptive matching method for cleaning time and particle size of semiconductor materials. Background Technology

[0002] With the continuous evolution of semiconductor manufacturing technology, wet cleaning processes play a crucial role in ensuring wafer surface cleanliness and device reliability. As one of the most frequently repeated processes in integrated circuit manufacturing, efficient and precise cleaning is not only related to the quality of subsequent film deposition and etching, but also a core element in improving the overall yield of semiconductor products. Especially at advanced process nodes, the removal of contaminants from wafer surfaces has shifted from simple physical rinsing to in-depth and refined control of micro-particle contamination at different scales.

[0003] Parameter matching and endpoint control in semiconductor material cleaning processes are key technologies for improving process window stability. This field primarily focuses on how to scientifically configure cleaning energy output and process duration based on the particle size distribution and adhesion characteristics of contaminants. An ideal cleaning process requires the system to provide differentiated physical driving forces for particulate contaminants of different sizes, and to optimize processing efficiency while ensuring surface morphology integrity.

[0004] Traditional semiconductor cleaning processes typically employ fixed-duration and constant-power modes based on empirical settings, which struggle to cope with the varied and complex particulate contamination distributions encountered in actual production. This simplistic control strategy results in a lack of targeted application of cleaning energy, failing to adaptively adjust to dynamic changes in contaminant particle size. Especially when processing wafers containing high aspect ratio trenches or intricate fin structures, prolonged high-energy input can easily trigger cavitation damage, leading to mechanical failure or performance degradation of the delicate patterned structures. Furthermore, existing systems lack real-time sensing and online decoupling capabilities for contaminant detachment, making it impossible to accurately determine the cleaning endpoint. This forces process development to seek inefficient compromises between insufficient cleaning and excessive damage. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive matching method for cleaning time and particle size of semiconductor materials, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: an adaptive matching method for semiconductor material cleaning time and particle size, comprising the following specific steps: Step 1: Construct a multimodal acoustic fingerprint monitoring environment by integrating a high-sensitivity broadband acoustic sensor array in the cleaning tank. The broadband acoustic sensor array is used to capture in real time the acoustic emission signals excited by the semiconductor material due to the detachment of particles from the surface during the cleaning process. Step 2: Perform real-time decoupling analysis on the captured acoustic signals, and use wavelet transform and pattern recognition algorithms to separate the mixed acoustic signals into multiple frequency band signals corresponding to different particle size ranges, and establish the mapping relationship between particle size and acoustic features. Step 3: Perform a multi-stage particle size cleaning process. Based on the intensity of the separated frequency band signals, the cleaning process is divided into multiple consecutive cleaning stages corresponding to large-sized particles, medium-sized particles, and fine particles. Step 4: Implement dynamic adaptive adjustment of cleaning energy output. When the acoustic fingerprint intensity corresponding to a specific particle size range is continuously lower than the preset threshold within a preset time period, the control system automatically generates a power adjustment command to adjust the megason energy output parameters of the current stage. Step 5: Determine the end point of the cleaning process and execute the shutdown logic. The complete disappearance of the acoustic fingerprint signal corresponding to the last particle size range, i.e. the fine particles, is used as the trigger event for the end of the entire cleaning process, so as to realize the real-time adaptive matching between cleaning time and particle size.

[0007] Preferably, in step 1, the high-sensitivity broadband acoustic sensor array consists of multiple piezoelectric ceramic sensors with high signal-to-noise ratio, which are symmetrically arranged around the inner wall of the cleaning tank. The piezoelectric ceramic sensors are configured to sense acoustic disturbances covering a frequency range from low-frequency mechanical vibrations to high-frequency ultrasonic waves, ensuring complete capture of transient stress waves generated when particles of different kinetic energies detach from the surface.

[0008] Preferably, in step 1, the acoustic sensor array is encapsulated with acoustic impedance matching material during installation to reduce sound wave reflection loss at different medium interfaces and improve signal acquisition sensitivity. The sensor array is connected to the back-end signal processor via a high-speed data acquisition card, and the sampling frequency is set to be more than twice the highest monitoring frequency to ensure that the waveform of the original acoustic emission signal is not distorted.

[0009] Preferably, in step 2, the process of real-time decoupling analysis of the captured acoustic signal includes signal preprocessing, feature extraction, and frequency band separation. The signal preprocessing step uses an adaptive filter to filter out low-frequency background noise generated by the mechanical transmission of the cleaning machine and fixed-frequency interference signals directly generated by the megohmmeter generator, and extracts non-stationary random signal components that are highly correlated with the particulate matter detachment event.

[0010] Preferably, in step 2, the wavelet transform algorithm employs multi-scale decomposition technology to decompose the complex original acoustic signal into different frequency scale spaces. The wavelet basis functions are selected as basis functions that are highly correlated with the transient pulse shape generated by particle detachment. By calculating the waveform coefficients at different scales, specific energy peaks corresponding to the detachment of particles of different physical sizes are identified.

[0011] Preferably, in step 2, the pattern recognition algorithm performs real-time comparison based on a pre-constructed acoustic fingerprint database. The acoustic fingerprint database contains detachment feature models of various common semiconductor contaminant particles under different cleaning fluid environments. The pattern recognition algorithm decouples the mixed signal into signals with particle sizes greater than 1 micrometer, 0.5 to 1 micrometer, and less than 0.5 micrometers by calculating the cosine similarity or Euclidean distance between the current real-time signal and the standard model in the database.

[0012] Preferably, in step 3, the cleaning process is divided into multiple consecutive cleaning stages as follows: the first stage is set as the large particle size removal stage, focusing on monitoring acoustic fingerprints in the particle size range greater than 1 micrometer; the second stage is set as the medium particle size removal stage, focusing on monitoring acoustic fingerprints in the particle size range of 0.5 to 1 micrometer; and the third stage is set as the fine residue treatment stage, focusing on monitoring acoustic fingerprints in the particle size range of less than 0.5 micrometer.

[0013] Preferably, in step 3, the initial energy parameters for each cleaning stage are preset based on the surface structure characteristics of the semiconductor material to be cleaned. For wafers with high aspect ratio trench structures, the initial energy of the first stage is limited to a preset safety threshold to prevent the cavitation force generated by megasonic waves from directly breaking the delicate fin structure.

[0014] Preferably, in step 4, the dynamic adaptive adjustment of the cleaning energy output is specifically manifested as a feedback-based closed-loop control logic. When the acoustic fingerprint intensity corresponding to the large particles monitored in the first stage decays to below a preset threshold, the controller immediately reduces the output power of the megohmmeter generator and switches the generator's operating frequency to the mid-frequency band, smoothly transitioning to the second stage.

[0015] Preferably, in step 4, during the transition from stage 2 to stage 3, the system not only adjusts the frequency and power of the megasonic wave, but also simultaneously regulates the flow rate and circulation pressure of the cleaning fluid. The system increases the dynamic pressure of the fluid to assist in removing medium-sized particles that have loosened but not yet completely detached from the surface, while simultaneously reducing the instantaneous energy density of the megasonic wave, thereby protecting the delicate patterned structure.

[0016] Preferably, in step 5, the logic for determining the end point of the cleaning process includes a time confirmation window. When the acoustic fingerprint signal intensity corresponding to microparticles smaller than 0.5 micrometers drops to near zero and remains there for a predetermined period, the system determines that the current surface has reached the ideal cleanliness state. At this time, the system triggers a shutdown command, stops energy output, and starts the pure water replacement process.

[0017] Preferably, the semiconductor material cleaning time and particle size adaptive matching method further includes a real-time health monitoring mechanism for online assessment of the sensitivity drift of the acoustic sensors in the cleaning tank. When the system detects a significant shift in the background static noise benchmark, it automatically activates a compensation algorithm to linearly correct the judgment thresholds for each frequency band, ensuring that the accuracy of particle size matching is not affected by hardware aging.

[0018] Preferably, the pattern recognition algorithm involved in step 2 further employs a support vector machine for classification optimization. The support vector machine utilizes a radial basis function kernel to map low-dimensional nonlinear acoustic features to a high-dimensional feature space, achieving high-precision separation of particle detachment signals and cavitation noise signals in high-noise environments.

[0019] Preferably, in step 4, the adjustment of the megaacoustic energy output also involves the optimization of phase modulation. By dynamically adjusting the phase difference between multiple megaacoustic transducers, a moving acoustic field standing wave is formed in the cleaning tank, eliminating cleaning dead zones and ensuring that particles of different size ranges can obtain a uniform detachment driving force across the entire wafer surface.

[0020] Preferably, during the execution of the semiconductor material cleaning time and particle size adaptive matching method, the system records the duration of each particle size cleaning stage in real time and compares it with stored historical data. If the time of a specific stage is found to be significantly longer than the preset reference value, the system will automatically trigger an alarm, prompting the operator to check the chemical composition concentration of the cleaning solution or the status of the circulation filtration system.

[0021] Preferably, the acoustic emission signal acquired in step 1 passes through a gain-adjustable preamplifier circuit before being transmitted to the processor. The preamplifier automatically adjusts the amplification factor according to the signal strength to prevent waveform clipping distortion when large particles detach and generate strong signals, while ensuring sufficient recognition accuracy when fine particles detach and generate weak signals.

[0022] Preferably, the wavelet transform algorithm in step 2 divides the signal into different frequency band components through a four-layer decomposition structure. The high-frequency detail components of the first and second layers are mainly used to characterize the detachment state of nanoscale particles, while the low-frequency approximation components of the third and fourth layers correspond to the detachment process of micron-sized large particles.

[0023] Preferably, the switching criteria for each stage in step 3 also incorporate real-time data from the particle counter in the cleaning fluid. Stage switching can only proceed after confirmation by both criteria when the acoustic fingerprint indicates particle detachment has ceased and the online particle counter detects a decrease in the number of corresponding particles of that size in the discharged fluid below a specific concentration.

[0024] Preferably, the method achieves synchronization of each module through a central control unit. The central control unit receives digital signals from the acoustic sensor array and completes algorithm calculations and distributes execution instructions within a microsecond delay, ensuring that the adjustment of energy output and the particulate matter detachment event are highly aligned in the time dimension.

[0025] Preferably, when processing semiconductor materials with complex fin field-effect transistor structures, the method, in the third stage, namely the micro-residue treatment stage, reduces the megasonic power to a specific percentage below the rated power and uses only the local fluid microfluidic beam effect caused by the megasonic wave for cleaning, so as to ensure that the physical impact on the nanoscale fine structure is at a safe level.

[0026] Preferably, the method further includes logic for monitoring and compensating for the temperature of the cleaning fluid. Since the velocity of sound changes with temperature in a liquid, the system automatically corrects the capture range of the acoustic fingerprint feature frequency based on real-time feedback from the temperature sensor, eliminating the impact of temperature drift on particle size recognition accuracy.

[0027] Preferably, the shutdown logic in step 5 also involves the coordinated control of the wafer rotation speed. At the instant the cleaning endpoint is determined, the rotating stage is controlled to increase its rotation speed, using centrifugal force to quickly remove residual particulate-containing cleaning fluid and prevent secondary contamination.

[0028] Preferably, the method is integrated into a fully automated wet cleaning system with multi-tank linkage capability. Each cleaning tank is equipped with an independent acoustic monitoring and adaptive control module to achieve parallel and precise control of the particulate contaminant removal status in different process steps.

[0029] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieving a high degree of balance between cleaning precision and yield assurance. This invention establishes a real-time mapping relationship between acoustic fingerprint spectrum and particle size, abandoning the traditional fixed-duration cleaning mode. By matching the cleaning energy with the specific particle size to be removed at the current moment, it ensures powerful removal of large-sized contaminants and flexible protection of microstructures, solving the problem of semiconductor fine pattern damage caused by over-cleaning and improving the yield of devices in advanced processes.

[0030] 2. Improved production efficiency and equipment capacity. Since the cleaning endpoint is dynamically determined by the actual particle detachment state, the system can perform on-demand cleaning based on the actual contamination level of different wafers. For wafers with light contamination or containing only easily removable particles, the system can accurately identify and terminate inefficient cleaning steps early, shortening the total cleaning time per wafer and optimizing the production line cycle time.

[0031] 3. Excellent process adaptability and robustness. This invention utilizes acoustic emission signal decoupling analysis to perceive subtle physical changes during the cleaning process in real time, unaffected by environmental factors such as the transparency of the cleaning solution or slight fluctuations in chemical composition. This closed-loop control strategy based on physical events enables the process to automatically compensate for deviations introduced by hardware aging or different batches of raw materials, ensuring consistent cleaning quality.

[0032] 4. This invention pioneers a new paradigm for precision control in semiconductor manufacturing processes. For the first time, it elevates broadband acoustic fingerprint analysis from simple condition monitoring to a core decision-driven process, providing a technical path for semiconductor cleaning processes to transition from a black-box model to a white-box model. This design approach based on real-time sensing and dynamic matching has significant reference value and technology reuse value for other semiconductor processes with extremely high requirements for endpoint detection, such as chemical mechanical polishing and plasma etching.

[0033] 5. Reduced production costs and resource consumption. By precisely controlling cleaning time and energy output, this invention effectively reduces the consumption of high-purity cleaning fluid, deionized water, and electricity. Because unnecessary redundant cleaning time is avoided, the service life of key components in the cleaning tank, such as the megasonic transducer, is also extended, reducing equipment maintenance costs. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of acoustic fingerprint signal decoupling and particle size mapping in this invention; Figure 3 This is a flowchart illustrating the logical flow of multimodal acoustic fingerprint capture and real-time signal preprocessing in this invention. Figure 4 This is a logical flow diagram of the process stage switching and dynamic energy matching based on particle size interval division in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the acoustic sensing array, the central control unit, and the power actuator in this invention; Figure 6 This is a flowchart of the cleaning process endpoint trigger determination and shutdown protection logic in this invention. Detailed Implementation

[0035] Example 1: Please refer to the appendix Figure 1 To be continued Figure 6 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0036] In this embodiment, the adaptive matching method for semiconductor material cleaning time and particle size is implemented by constructing a highly integrated perception-decision-execution closed-loop system. This system is deployed in a fully automated wet cleaning equipment with multi-tank linkage capability. Each cleaning tank is equipped with an independent acoustic monitoring and adaptive control module, enabling parallel and precise control of the particulate contaminant removal status in different process steps.

[0037] Step 1: Construct a multimodal acoustic fingerprint monitoring environment.

[0038] A highly sensitive, broadband acoustic sensor array is integrated within the cleaning tank. This array consists of multiple piezoelectric ceramic sensors with high signal-to-noise ratios, symmetrically arranged around the inner wall of the cleaning tank. The piezoelectric ceramic sensors are configured to sense acoustic disturbances covering a frequency range from low-frequency mechanical vibrations to high-frequency ultrasonic waves, specifically from 10 kHz to 5 MHz. This broadband design ensures the system can fully capture transient stress waves generated when particles of varying kinetic energies detach from the surface.

[0039] During installation, the acoustic sensor array is encapsulated using acoustic impedance matching material. This material has acoustic impedance characteristics similar to those of cleaning fluids (such as deionized water or diluting chemical solutions), significantly reducing sound wave reflection loss at interfaces between different media and improving signal acquisition sensitivity. The sensor array is connected to the back-end signal processor via a high-speed data acquisition card. The sampling frequency of the high-speed data acquisition card is set to be more than twice the highest monitoring frequency, specifically a sampling rate of 12.5 MHz, ensuring that the waveform of the original acoustic emission signal remains undistorted.

[0040] Before being transmitted to the processor, the acquired acoustic emission signal first passes through a gain-adjustable preamplifier circuit. This preamplifier automatically adjusts its amplification factor according to the signal strength to prevent waveform clipping distortion when large particles detach and generate strong signals, while ensuring sufficient recognition accuracy when small particles detach and generate weak signals. This preamplifier features ultra-low noise floor, with its equivalent input noise controlled in the microvolt range.

[0041] Step 2: Perform real-time decoupling analysis on the captured acoustic signal.

[0042] Wavelet transform and pattern recognition algorithms are used to separate mixed acoustic signals into multiple frequency bands corresponding to different particulate matter size ranges, establishing a mapping relationship between particulate matter size and acoustic features. This process specifically includes signal preprocessing, feature extraction, and frequency band separation.

[0043] As a preferred embodiment, the wavelet transform algorithm employs discrete wavelet transform and selects the Daubechies 4th order wavelet basis function. The compact support and orthogonality of this basis function enable it to effectively capture the transient impact pulses generated when particles detach.

[0044] Specifically, the original acoustic emission signal After four levels of discrete wavelet decomposition, the decomposition process can be expressed as follows: ; in, For the first The detail coefficients (high-frequency components) obtained from layer decomposition correspond to transient pulse signals; The approximate coefficients (low-frequency components) obtained from the fourth level of decomposition correspond to the background trend signal. The wavelet transform algorithm decomposes the signal into different frequency scale spaces through multi-resolution analysis.

[0045] Furthermore, the frequency band signal energy $E_j$ corresponding to particles of different sizes is obtained by calculating the detail coefficient energy of the corresponding level: ; in, In this invention, and The energy integral value is used to characterize the detachment state of nanoscale particles (less than 0.5 micrometers), while and The energy integral value corresponds to the detachment process of micron-sized particles (greater than 1 micrometer). This is achieved by real-time monitoring and calculation of different levels of energy. The system establishes a quantitative mapping relationship between particle size and acoustic characteristics.

[0046] The signal preprocessing step utilizes an adaptive filter to remove low-frequency background noise generated by the mechanical transmission of the cleaning machine and fixed-frequency interference signals directly generated by the megasonic generator. The adaptive filter continuously minimizes noise components in the output signal by updating its internal coefficients in real time, extracting non-stationary random signal components highly correlated with particulate matter detachment events.

[0047] The wavelet transform algorithm employs multi-scale decomposition technology to break down the complex original acoustic signal into different frequency scales. In this embodiment, the wavelet transform algorithm divides the signal into different frequency band components through a four-layer decomposition structure. The high-frequency detail components of layers 1 and 2 are mainly used to characterize the detachment state of nanoscale particles, while the low-frequency approximation components of layers 3 and 4 correspond to the detachment process of micrometer-sized large particles. The wavelet basis functions are selected to be highly correlated with the transient pulse shape generated by particle detachment, such as multi-scale wavelet bases with compact support and orthogonality. By calculating the waveform coefficients at different scales, specific energy peaks corresponding to the detachment of particles of different physical sizes are identified.

[0048] The pattern recognition algorithm performs real-time comparisons based on a pre-built acoustic fingerprint database.

[0049] Specifically, the construction process of the acoustic fingerprint database is as follows: In an offline state, standard particles of known diameter (e.g., polystyrene microspheres of 1.5 μm, 0.7 μm, and 0.3 μm) are dispersed in a cleaning solution and detached from the surface of a semiconductor material. Their acoustic emission signals are captured using the broadband acoustic sensor array. After performing wavelet transform on the captured signals, multidimensional feature vectors of detail coefficients at each layer are extracted. Its composition is as follows: in, The energy for the detail coefficients of the aforementioned layers; kurtosis factor , used to measure the impulsiveness of a signal; For waveform factor, This is used to describe the overall shape of the signal. The feature vector corresponding to each standard particle... The granularity labels and their constituent elements constitute a standard model, which is then stored in a database.

[0050] During the real-time cleaning process, the system performs the same feature extraction on the captured signals to obtain real-time feature vectors. The pattern recognition algorithm calculates... Compared with all standard model feature vectors in the database Matching using Euclidean distance: ; in, The dimension of the feature vector (in this embodiment) The system identifies and matches... The standard model with the smallest Euclidean distance corresponds to the granularity range represented by the current signal, thus achieving decoupling of mixed signals.

[0051] The acoustic fingerprint database contains detachment feature models of various common semiconductor contaminant particles (such as silica dust, photoresist residue, and metal ion clusters) under different cleaning solution environments. The pattern recognition algorithm decouples the mixed signal into signals with particle sizes greater than 1 micrometer, 0.5 to 1 micrometer, and less than 0.5 micrometer by calculating the cosine similarity or Euclidean distance between the current real-time signal and the standard model in the database.

[0052] To further improve recognition accuracy, the pattern recognition algorithm also employs support vector machines for classification optimization. The support vector machine utilizes radial basis function kernels to map low-dimensional nonlinear acoustic features to a high-dimensional feature space, achieving high-precision separation of particle detachment signals and cavitation noise signals even in high-noise environments. The system synchronizes all modules through a central control unit, which receives digital signals from the acoustic sensor array and completes algorithm computation and instruction distribution within microsecond delays.

[0053] Step 3: Perform a multi-stage particle size cleaning process.

[0054] Based on the intensity of the separated frequency band signals, the cleaning process is divided into multiple consecutive cleaning stages corresponding to large-sized particles, medium-sized particles, and micro-particles. Specifically, the first stage is set as the large-particle removal stage, focusing on monitoring the acoustic fingerprint of particles larger than 1 micrometer; the second stage is set as the medium-particle removal stage, focusing on monitoring the acoustic fingerprint of particles between 0.5 and 1 micrometer; and the third stage is set as the micro-residue treatment stage, focusing on monitoring the acoustic fingerprint of particles smaller than 0.5 micrometers.

[0055] The initial energy parameters for each cleaning stage are preset based on the surface structure characteristics of the semiconductor material to be cleaned. For wafers with high aspect ratio trench structures, the initial energy of the first stage is limited to a preset safety threshold to prevent the cavitation force generated by megasonic waves from directly breaking the delicate fin structure. For example, when processing semiconductor materials with complex fin field-effect transistor structures, the method reduces the megasonic wave power in the third stage, i.e., the micro-residue treatment stage, and only utilizes the local fluid microfluidic beam effect induced by the megasonic waves for cleaning to ensure that the physical impact force on the nanoscale delicate structure is at a safe level.

[0056] The switching criteria for each stage also incorporate real-time data from a particle counter in the cleaning fluid. Stage switching is only executed after confirmation by both criteria: the acoustic fingerprint indicates particle detachment has ceased, and the online particle counter detects a decrease in the corresponding particle size in the discharged fluid below a specific concentration. Furthermore, the system includes logic for monitoring and compensating for the cleaning fluid temperature. Since the velocity of sound in a liquid varies with temperature, the system automatically corrects the capture range of the acoustic fingerprint feature frequency based on real-time feedback from the temperature sensor, eliminating the impact of temperature drift on particle size recognition accuracy.

[0057] Step 4: Implement dynamic adaptive adjustment of cleaning energy output.

[0058] When the acoustic fingerprint intensity corresponding to a specific granularity range is detected to be continuously lower than a preset threshold within a preset time period, the control system automatically generates a power adjustment command to adjust the megasonic energy output parameters for the current stage. This adjustment manifests as a feedback-based closed-loop control logic.

[0059] When the acoustic fingerprint intensity corresponding to the large particles monitored in Stage 1 decays below a preset threshold, the controller immediately reduces the output power of the megasonic generator and switches the generator's operating frequency to the mid-frequency range (e.g., 900 kHz to 1.2 MHz), smoothly transitioning to Stage 2. During the transition from Stage 2 to Stage 3, the system not only adjusts the frequency and power of the megasonic waves but also simultaneously regulates the flow rate and circulation pressure of the cleaning fluid. By increasing the dynamic pressure of the fluid, the system assists in removing medium-sized particles that have loosened but not yet completely detached from the surface, while simultaneously reducing the instantaneous energy density of the megasonic waves.

[0060] Adjusting the megaacoustic energy output also involves optimizing phase modulation. By dynamically adjusting the phase difference between multiple megaacoustic transducers, a moving acoustic standing wave is formed within the cleaning tank, eliminating cleaning dead zones and ensuring that particles of different sizes receive uniform detachment driving force across the entire wafer surface. The system records the duration of each particle size cleaning stage in real time and compares it with stored historical data. If the duration of a specific stage is found to be significantly longer than the preset reference value, the system will automatically trigger an alarm, prompting the operator to check the chemical composition concentration of the cleaning solution or the status of the circulation filtration system.

[0061] In addition, the system has a real-time health monitoring mechanism to assess the sensitivity drift of the acoustic sensors in the cleaning tank online. When the system detects a significant shift in the background static noise benchmark (e.g., the shift exceeds 10% of the preset benchmark value), it automatically activates a compensation algorithm to linearly correct the judgment thresholds for each frequency band, ensuring that the accuracy of granular matching is not affected by hardware aging.

[0062] Step 5: Determine the end point of the cleaning process and execute the shutdown logic.

[0063] The complete disappearance of the acoustic fingerprint signal corresponding to the last particle size range, i.e., the fine particles, is used as the trigger event for the end of the entire cleaning process. The logic for determining the end of the cleaning process includes a time confirmation window. When the intensity of the acoustic fingerprint signal corresponding to fine particles smaller than 0.5 micrometers drops to near zero and remains there for a predetermined period (e.g., 5 to 10 seconds), the system determines that the current surface has reached the ideal cleanliness state.

[0064] At this point, the system triggers a shutdown command, stopping energy output and initiating the pure water replacement process. The shutdown logic also involves coordinated control of the wafer rotation speed. Upon determining the end of the cleaning process, the rotating stage's speed is increased (e.g., from 500 rpm to 2000 rpm), using centrifugal force to quickly remove residual particulate-containing cleaning fluid and prevent secondary contamination. The entire process achieves real-time adaptive matching between cleaning time and particle size, ensuring that the actual total cleaning time is precisely compressed to the minimum necessary time to remove all particulate contaminants.

[0065] Example 2: Based on Example 1, this example further refines the adaptive filtering and frequency band separation logic in acoustic fingerprint signal processing.

[0066] Specifically, in step 2, the implementation of the adaptive filter does not depend on a preset cutoff frequency, but rather dynamically adjusts its weighting coefficients by monitoring the second-order statistical characteristics of the signal in real time.

[0067] In a preferred embodiment, the adaptive filter employs the least mean square algorithm. This algorithm iteratively updates the filter weight coefficients. So that the output signal Expected response The mean square error between them is minimized. The update formula is as follows: ; in, As the input signal vector, in this embodiment, its initial value is set to the static background noise signal collected by the system when there is no wafer and only circulating cleaning fluid; For error signals, , here The original acoustic emission signal contains noise. Real-time estimation of background noise by the filter This is the step size parameter, and its value is set to... This ensures the convergence speed and stability of the algorithm. Through this process, the error signal... This refers to the non-stationary random signal components that are highly correlated with particulate matter detachment events after filtering out low-frequency mechanical vibrations and fixed-frequency interference.

[0068] During the initial cleaning phase, the system acquires an acoustic signal under pure fluid circulation conditions as a reference noise mode, which includes the resonant frequency component of the megaacoustic transducer itself. Once the wafer enters the cleaning area, the sensor captures a mixed signal consisting of the reference noise mode, mechanical interference noise, and impact pulse signals caused by particle detachment.

[0069] The adaptive filter uses a reference noise mode as the prediction target and extracts the particle detachment signal by calculating the residual signal. The residual signal is the target non-stationary random signal component after filtering out fixed-frequency interference. This approach ensures that even in environments with frequent switching of mega-sound power, the system can still stably lock onto the weak acoustic characteristics of particulate detachment.

[0070] For the wavelet transform algorithm, this embodiment uses discrete wavelet transform to perform multi-level decomposition of the signal. Each decomposition process involves cascaded high-pass and low-pass filtering operations. The first-level decomposition divides the original bandwidth signal into high-frequency and low-frequency bands, with the high-frequency band directly fed into the pattern recognition module to identify detachment events of nanoscale particles (particle size less than 100 nanometers). The second and third-level decompositions further refine the low-frequency band, separating characteristic frequency bands corresponding to submicron (0.1 micrometer to 0.5 micrometer) and micrometer (greater than 0.5 micrometer) particles.

[0071] During the feature extraction stage, the system not only extracts the energy peak value but also calculates the kurtosis factor and waveform factor of the signal. The kurtosis factor is highly sensitive to pulsed acoustic emission signals; when particles undergo brittle fracture or instantaneous detachment from the surface, the kurtosis factor exhibits abrupt changes. By constructing a multidimensional feature vector from the energy peak value, kurtosis factor, and the variance of wavelet coefficients, the system establishes a more robust particle size mapping relationship.

[0072] In pattern recognition algorithms, specific weight allocation schemes are applied to different types of semiconductor materials (such as polysilicon, silicon dioxide, and silicon nitride). Because the adhesion of particles to different material surfaces varies, their energy conversion efficiency upon detachment differs. The system compensates for the intensity of the identified signal using a preset material correction coefficient. For example, when processing silicon nitride surfaces, due to their lower surface energy, the acoustic signal amplitude generated by particle detachment is relatively weak. The system automatically increases the gain of the corresponding frequency band to ensure recognition accuracy.

[0073] In this embodiment, the support vector machine is configured as a multi-class classification model. The system is pre-trained offline with a large number of typical granular detached samples.

[0074] Furthermore, to achieve high-precision separation of particle detachment signals and cavitation noise signals in high-noise environments, the pattern recognition algorithm also employs a support vector machine (SVM) for classification optimization. The SVM's classification decision function is: ; in, The input is a real-time multidimensional feature vector; For the training sample set, This indicates whether the sample is a particle detachment signal or a cavitation noise signal; The number of support vectors; and For the Lagrange multipliers and bias terms obtained during training; The radial basis function kernel has the following form:

[0075] in, The kernel function parameter is pre-optimized to 0.5 using a grid search method. When When the signal is clear, the system will identify it as a particle detachment signal and proceed to the subsequent particle size matching process; otherwise, it will identify it as a cavitation noise signal and ignore it.

[0076] During online execution, after receiving multidimensional feature vectors, the support vector machine calculates the decision function of the hyperplane and classifies the signals into the corresponding granularity intervals in real time. This classification method effectively shields the continuous random noise generated by cavitation bubble collapse, because cavitation noise is continuous in the time domain, while particle detachment signals are abrupt in the time domain.

[0077] The central control unit utilizes a double-buffering mechanism to process high-frequency data streams when executing the algorithm. The data acquisition card continuously writes digitized acoustic data into the first buffer. When the first buffer is full, the DMA controller automatically switches to the second buffer and simultaneously triggers an interrupt request from the processor to perform parallel wavelet operations and pattern recognition on the data in the first buffer. This architecture ensures zero packet loss during processing and enables control over the delay in issuing energy regulation commands.

[0078] Example 3: This example focuses on describing the specific execution details and hardware coordination logic of the dynamic adaptive energy adjustment in step 4 under a specific semiconductor process scenario.

[0079] Specifically, in step 4, the control system performs multi-parameter coupled control on the megasonic generator based on the feedback acoustic fingerprint intensity. This control is not limited to simple power adjustment, but involves the comprehensive adjustment of voltage amplitude, operating frequency, duty cycle, and phase delay.

[0080] During the first stage, the large particle removal stage, the megasonic generator is configured in high-power output mode while employing a lower operating frequency (e.g., 800 kHz) to enhance the cavitation effect depth in the cleaning fluid. At this time, the central control unit monitors the acoustic energy integral value in the frequency band greater than 1 micrometer in real time. As the number of large particles decreases, this energy integral value decays exponentially. When the energy integral value drops to 15% of its initial value and remains so for 3 seconds, the system determines that the removal of large particles is essentially complete.

[0081] Subsequently, the system automatically switches to stage 2. The central control unit sends a frequency switching command to the megasonic generator, adjusting its resonant point to 1.5 MHz. High-frequency megasonic waves have a smaller acoustic flow boundary layer thickness, enabling them to penetrate submicron-level surface structures more effectively. In this stage, the system simultaneously activates cleaning fluid flow rate compensation. By adjusting the inverter frequency of the circulating pump, the linear flow rate of the cleaning fluid is increased. The increased flow rate generates additional tangential shear force, assisting the 0.5-micron to 1-micron particles, loosened by the acoustic waves, to detach from the surface.

[0082] Phase modulation plays a crucial role in this embodiment. The transducer array at the bottom of the cleaning tank is divided into multiple independently controlled quadrants. The central control unit applies a specific phase delay to the drive circuits in different quadrants, causing the synthesized sound field to drift slowly horizontally within the cleaning tank. This moving standing wave technology solves the problem of uneven cleaning caused by the fixed positions of antinodes and nodes in traditional fixed-phase sound fields. The acoustic sensor array provides real-time feedback on the particle detachment intensity in each quadrant region. If the acoustic fingerprint intensity of a specific quadrant is consistently higher than that of other quadrants, the system automatically increases the phase weight corresponding to that quadrant, extending its residence time at high power density.

[0083] In the third stage, the micro-residue treatment stage, the system employs an extremely flexible cleaning strategy. At this point, the megasonic generator switches to pulse mode with a duty cycle set to 20%. In this mode, acoustic energy is transmitted intermittently, maintaining the microfluidic effect to remove extremely fine particles while avoiding fatigue damage to the delicate fin structure caused by continuous sound pressure. The system then shifts its monitoring focus to the frequency band smaller than 0.5 micrometers.

[0084] To ensure that particulate matter is completely discharged from the cleaning tank and does not undergo secondary adsorption on the surface, the central control unit also controls the cleaning fluid recovery and discharge valve in real time. At the moment of stage switching, the system performs a rapid overflow operation, which increases the discharge volume in a short period of time to quickly remove a large number of large particles suspended in the liquid.

[0085] In addition, the system also involves a fault diagnosis logic during execution. If the frequency distribution of the signal captured by the acoustic sensor exceeds the preset physical boundary (for example, an abnormal high-frequency oscillation signal appears), the central control unit will immediately identify that the transducer is unloaded or the cleaning fluid level is insufficient, triggering emergency energy degradation protection.

[0086] Example 4: This example describes an implementation scheme that includes real-time health monitoring and closed-loop optimization of historical data.

[0087] In steps 1 and 5, the system integrates a sensor health assessment mechanism based on benchmark feature extraction. During the self-check phase before each cleaning task begins, a standard excitation source (a miniature broadband piezoelectric element) inside the cleaning tank emits a set of pulse test signals of preset intensity. An array of acoustic sensors distributed along the tank walls receives this signal and transmits it to the central control unit.

[0088] The central control unit performs a cross-correlation calculation between the currently received pulse response waveform and the original factory reference waveform stored in the system's non-volatile memory. If the cross-correlation coefficient is lower than 0.92, or if the center frequency of the signal shifts by more than 50 kHz, the system determines that the sensor has experienced a decrease in sensitivity or that the encapsulation layer is aging. In this case, the system will not shut down immediately, but will instead initiate an automatic compensation program.

[0089] The automatic compensation program calculates a sensitivity attenuation factor and linearly amplifies and corrects the signal strength of each frequency band extracted in subsequent step 2. Simultaneously, the system automatically increases the preset threshold in step 4. This compensation mechanism significantly extends the effective lifespan of the sensor and ensures accurate granularity identification and stage switching even in the early stages of sensor performance degradation.

[0090] In the decision logic of step 5, the system introduces a dynamic confirmation time window. The length of this window is not fixed, but is predicted based on the cumulative duration of each stage in steps 3 and 4. If historical data shows that the removal of fine particles from a certain type of wafer typically takes 30 seconds under current process conditions, the confirmation window will be automatically set to 6 seconds. If no pulse burst occurs again within 6 seconds after the signal strength drops to zero, the final confirmation endpoint is reached.

[0091] The system records the duration of each particle size stage in each cleaning batch in real time and constructs it into a process time fingerprint vector stored on the local server. For the same type of semiconductor material, if the second stage time of multiple consecutive batches is longer than the historical average, the central control unit will automatically analyze the cause. By correlating and analyzing the cleaning solution temperature, pH value, and differential pressure data of the filtration system, the system can accurately identify whether the decrease in cleaning solution cleanliness is due to filter cartridge clogging, or whether the weakened wetting and detachment ability of the chemical solution is due to ineffective chemical concentration.

[0092] After determining the end point of the cleaning process and executing the shutdown logic, the system will automatically generate a cleaning quality digital report. This report details the removal kinetics curves for large, medium, and fine particles, showing the variation of signal intensity in each frequency band over time. Process engineers can analyze the slopes of these curves to optimize the initial energy parameter presets for subsequent batches.

[0093] For the pure water replacement process mentioned in step 5, the system achieves a seamless switch from chemical cleaning to deionized water rinsing by controlling a multi-way reversing valve. During the switchover, the rotation speed of the wafer stage is synchronized with the water injection flow rate. Based on a fluid calculation model, the system maintains a lower rotation speed in the initial stage of water injection to prevent a large number of air bubbles from interfering with the acoustic signal, and increases the rotation speed in the later stage of water injection to utilize centrifugal force to complete the drying process.

[0094] The entire system runs on an industrial-grade real-time operating system, with strict prioritization of tasks. Acoustic signal acquisition and decoupling tasks are given the highest priority to ensure deterministic signal processing; energy regulation and actuator control tasks are secondary; while interface display and data storage tasks are executed asynchronously in the background. This software architecture ensures that even in complex cleaning scenarios with extremely high particle loads, the system can still achieve nanosecond-level sampling alignment and millisecond-level feedback control.

[0095] Example 5: This example focuses on the application of this method in processing semiconductor wafers with complex three-dimensional structures (such as GAAFET nanosheets) under advanced processes.

[0096] In this application scenario, the multi-stage particle size cleaning process in step 3 is refined into a denser energy ladder. Because the nanosheet structure is extremely sensitive to sound pressure, excessive cavitation pressure can cause the nanosheets to stick together or break. Therefore, the system pre-loads a damage sensitivity matrix into the central control unit based on the wafer layout information.

[0097] In step 4, the dynamic adaptive adjustment of energy output is subject to hard constraints of the damage sensitivity matrix.

[0098] Specifically, the feedback-based closed-loop control logic employs a proportional-integral-derivative controller to precisely adjust the mega-sound energy output. The controller adjusts the acoustic fingerprint intensity based on the current particle size range. Compared with the preset target threshold Deviation between Calculate the control quantity That is, the power adjustment command of the megasonic generator: ; in, ; This is the proportionality coefficient, and its function is to respond quickly based on the current deviation. These are the integral coefficients used to eliminate steady-state errors; These are differential coefficients used to predict the trend of deviation changes and suppress system oscillations. In this embodiment, they are applied to the large particle removal stage (first stage). 5, , To achieve rapid removal; and in the micro-residue treatment stage (third stage), the system automatically... The value was lowered to 0.1 to enhance the flexibility and stability of the control and avoid damage to the delicate structure.

[0099] When the controller according to After adjusting the power, the acoustic sensor array will monitor the new [power] in real time. This forms a closed loop, ensuring that the energy output and the particle detachment event are highly aligned in the time dimension.

[0100] When removing large particles in the first stage, even if the particle detachment speed is slow according to acoustic fingerprint feedback, the control system does not allow for unlimited power increases. Instead, it increases the sound field coverage within the deep and narrow trench by adjusting the frequency modulation bandwidth of the megasonic waves and using frequency sweeping technology. The frequency sweep range is set within ±50 kHz of the resonant frequency, with a variation period of 10 milliseconds. This method can improve particle removal efficiency without increasing the instantaneous sound pressure amplitude.

[0101] In the signal decoupling analysis of step 2, a dedicated structural damage feature monitoring frequency band was added to the pattern recognition algorithm to address the unique characteristics of the nanosheet structure. Experimental studies show that when nanosheets undergo mechanical resonance or fracture, they excite high-frequency acoustic signals within a specific frequency range (typically above 2 MHz). While monitoring particle detachment from the acoustic fingerprint, the system simultaneously monitors this damage feature frequency band in real time. If the signal energy in this frequency band exceeds a preset safety alarm limit, the system will execute an emergency cutoff within 10 microseconds, forcibly shutting down the megasonic generator and switching to a pure fluid circulation mode for flexible repair.

[0102] For the cleaning endpoint determination in step 5, considering the greater difficulty in removing particles inside three-dimensional structures, a diffusion equilibrium stage was added to the shutdown logic. After the acoustic fingerprint disappears, the system does not immediately stop the process, but maintains a very low fluid circulation rate and activates the heating system to raise the cleaning fluid temperature by 5 degrees Celsius. Thermal motion is used to increase the diffusion rate of the tiny particles remaining deep in the trenches. During this period, the acoustic sensor array continuously monitors. If the acoustic fingerprint of detached microparticles is detected again during the diffusion equilibrium stage, a short energy compensation pulse is automatically triggered.

[0103] During the determination of the endpoint and the execution of the shutdown logic, the acceleration curve of the rotating stage was smoothed. The system uses an S-shaped acceleration / deceleration algorithm to control the motor, avoiding secondary physical damage to the fragile nanosheet structure caused by shear forces resulting from sudden changes in rotation speed. In the centrifugal drying stage, nitrogen-assisted drying is introduced into the system. The nitrogen nozzle position is linked to the water injection position to ensure that the liquid film on the wafer surface spreads evenly from the center to the edge and disappears quickly, preventing the formation of drying marks.

[0104] The central control unit in this embodiment also possesses self-learning capabilities. By mining cleaning data from thousands of wafers, the system can automatically identify the optimal energy allocation scheme under different contaminant loads. For example, for wafers produced by a specific type of mask, the system, through historical data analysis, discovers that shortening the large particle removal stage by 10 seconds and increasing the phase modulation depth of the medium particle stage can improve yield by 5%. In this case, the system automatically modifies the initial parameters in the process formulation, achieving closed-loop self-evolution of the process.

[0105] Example 6: This example details the implementation process and data performance of the method of the present invention on a 12-inch silicon wafer production line through a specific operation example.

[0106] Operating environment: A wet cleaning system with four independent cleaning tanks is used to clean silicon wafers after chemical mechanical polishing (CMP). The main contaminants on the silicon wafer surface are cerium oxide particles and a small amount of organic debris generated by the wear of the polishing pads.

[0107] Step 1: The silicon wafer is loaded into the cleaning tank by a robotic arm, and the stage begins to rotate at a speed of 300 revolutions per minute. Twelve broadband piezoelectric ceramic sensors on both sides of the cleaning tank enter monitoring mode. At this time, the background noise of the acoustic monitoring environment stabilizes at -80 decibels.

[0108] Step 2: Activate the mega-acoustic wave with an initial power of 200 watts. The sensor array immediately captures complex acoustic waveforms. The wavelet transform module of the central control unit decomposes the signal. A large number of high-energy pulses appear in the frequency band from 500 kHz to 1.5 MHz, corresponding to the detachment of numerous cerium oxide particles larger than 1 micrometer from the silicon wafer surface. The pattern recognition algorithm classifies these signals as large-particle frequency band signals.

[0109] The closed-loop interaction between steps 3 and 4 is executed: At the 15th second of the cleaning process, the energy integral value of the large-particle frequency band signal begins to drop sharply. At this point, the central control unit determines that the large particles have been largely removed. The system issues a command to reduce the megahertz power to 120 watts and switch the frequency from 950 kHz to 1.1 MHz.

[0110] Phase 2: The system focuses on monitoring the 0.5-micron to 1-micron frequency band. Because particles of this size adhere more evenly to the silicon wafer surface, the acoustic signal appears as a high-frequency, medium-amplitude pulse sequence. At this stage, the system detects slight harmonic interference in the background noise. The adaptive filter updates its weighting coefficients and filters out the interference within 50 milliseconds, maintaining the signal-to-noise ratio.

[0111] At the 40th second, the acoustic fingerprint intensity in the 0.5-micron to 1-micron frequency band decreased to the preset threshold. The system automatically adjusted the cleaning fluid circulation pump, increasing the injection pressure from 0.2 MPa to 0.25 MPa, and switched the megasonic frequency to low-power mode (50 watts), entering stage 3.

[0112] Entering Phase 3: In this phase, the acoustic pulses captured by the system are extremely weak and sparse, corresponding to the last remaining nanoscale particles. The central control unit activates a high-sensitivity detection mode, capturing these weak signals by increasing the sampling depth.

[0113] Step 5: At approximately 55 seconds, the acoustic fingerprint signals for all preset particle size ranges completely disappear. The system enters a 10-second confirmation window. During this window, the online particle counter displays that the particle concentration in the drained solution is below 3 particles per milliliter (for particles larger than 0.1 micrometers). After the confirmation window ends, the system immediately shuts off the megasonic wave at 65 seconds and initiates rapid pure water replacement.

[0114] Analysis of Results: Compared to the previous fixed 90-second cleaning mode used on this production line, the method of this invention significantly shortens the cleaning time for this batch of wafers, resulting in a substantial increase in efficiency. Furthermore, subsequent defect detection results showed that the residual rate of large particles on the silicon wafer surface was zero, and the number of residual fine particles was superior to the original process. More importantly, scanning electron microscopy observation of the sensitive test patterns revealed no pattern breakage or edge chipping phenomena commonly found in traditional high-power cleaning.

[0115] In this embodiment, these circuits trigger the corresponding control logic by subtracting the real-time calculated value from the reference value in the memory and determining whether the difference is positive or negative. The whole process does not involve complex algebraic calculations, but is based on pure technical state determination.

[0116] Example 7: This example expands upon the multimodal acoustic fingerprint monitoring environment and its interaction with the backend hardware in the method of the present invention.

[0117] In step 1, the piezoelectric ceramic sensor array is arranged using spatial diversity technology. Specifically, the sensors are distributed according to the golden ratio along the major and minor axes of the cleaning tank to minimize the blind spots caused by acoustic standing waves in signal acquisition. Each sensor is connected to the front-end signal conditioning module via an independent shielded cable.

[0118] The front-end signal conditioning module includes a high-input-impedance charge amplifier with an input impedance greater than 10^12 ohms, perfectly matching the high source impedance characteristics of the piezoelectric ceramic sensor. The charge amplifier converts the weak charge signal output by the sensor into a voltage signal, which is then passed through a 5 MHz bandpass filter to eliminate extremely low-frequency power grid interference and extremely high-frequency spatial electromagnetic radiation noise.

[0119] The filtered signal enters a gain-adjustable preamplifier. The amplifier's gain is controlled by a 4-bit digital signal provided by a central control unit, offering 16 gain levels adjustable from 0 dB to 60 dB. This automatic gain control algorithm is based on peak signal detection logic: when the amplitude of three consecutive sampling points exceeds 90% of full scale, the system automatically reduces the gain level; when the amplitude of 1000 consecutive sampling points is below 10% of full scale, the system automatically increases the gain level. This dynamic range extension technology ensures the system can hear the whisper of tiny particles detaching as well as withstand the roar of large particles impacting.

[0120] The high-speed data acquisition card employs a multi-channel synchronous sampling architecture, ensuring that the acoustic signals from the 12 sensors are perfectly aligned on the time axis. This synchronization is crucial for the spatial positioning analysis in subsequent step 2. Although the primary objective of this invention is granularity matching, by performing cross-correlation calculations on the time differences of arrival of the 12 channel signals (which can also be translated into a logical process described in words, i.e., by comparing the order of appearance of wave peaks in different channels and their time differences), the central control unit can roughly determine the region where particles are most active and guide the megasonic generator to adjust its phase distribution accordingly, focusing the acoustic energy onto that region.

[0121] In step 2, the acoustic fingerprint database employs a high-speed index architecture based on feature hashing for its storage structure. Each acoustic feature model is compressed into a fixed-length fingerprint feature vector. During real-time matching, the system calculates the hash value of the current signal feature vector and directly searches for similar models in a pre-defined hash table. This algorithm avoids the enormous computational burden of traversing the database one by one, ensuring real-time performance when processing high-sampling-rate data.

[0122] For support vector machine classifiers, the training process is completed offline on the server side, but the generated classification model parameters (i.e., the normal vector and intercept values ​​of the hyperplane) are embedded in the field-programmable gate array (FPGA) of the central control unit. Utilizing its parallel computing advantage, the FPGA can output granular classification results within nanoseconds through a series of adders and multipliers (which are represented as linear weighted combinations of signals at the logic layer).

[0123] The central control unit mentioned in this invention adopts a heterogeneous architecture that combines a multi-core processor and an FPGA. The FPGA is responsible for computationally intensive tasks such as high-speed data acquisition, adaptive filtering, and wavelet decomposition at the lower level; the multi-core processor is responsible for higher-level logic judgments, state machine switching at each cleaning stage, and communication with the host computer. This architecture ensures the robustness of the system. Even if the host computer software experiences temporary lag, the underlying closed-loop cleaning control logic can still operate independently and safely according to the acoustic signals until completion.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive matching of cleaning time and particle size for semiconductor materials, characterized in that, Includes the following steps: A multimodal acoustic fingerprint monitoring environment is constructed by arranging a high-sensitivity broadband acoustic sensor array in the cleaning tank. The broadband acoustic sensor array is used to capture in real time the acoustic emission signals excited by the semiconductor material due to the detachment of particles of different sizes from the surface during the cleaning process. The captured acoustic emission signals are decoupled and analyzed in real time. Wavelet transform and pattern recognition algorithms are used to separate the mixed acoustic signals into multiple frequency band signals corresponding to different particle size ranges, and a mapping relationship between particle size and acoustic characteristics is established. A multi-stage particle size cleaning process is implemented, and the cleaning process is divided into multiple continuous cleaning stages corresponding to large-sized particles, medium-sized particles, and fine particles based on the intensity of the separated frequency band signals. The system implements dynamic adaptive adjustment of cleaning energy output. When the acoustic fingerprint intensity corresponding to the particle size range is continuously lower than the preset threshold within a preset time, the control system automatically generates a power adjustment command to adjust the megason energy output parameters of the current stage. The system determines the end point of the cleaning process and executes the shutdown logic. The complete disappearance of the acoustic fingerprint signal corresponding to the last particle size range, i.e. the fine particles, is used as the trigger event for the end of the entire cleaning process, thereby achieving real-time adaptive matching between cleaning time and particle size.

2. The semiconductor material cleaning time and particle size adaptive matching method according to claim 1, characterized in that, In the step of constructing a multimodal acoustic fingerprint monitoring environment, the high-sensitivity broadband acoustic sensor array is composed of multiple piezoelectric ceramic sensors with high signal-to-noise ratio. The piezoelectric ceramic sensors are symmetrically arranged around the inner wall of the cleaning tank, and the piezoelectric ceramic sensors are configured to sense acoustic disturbances ranging from low-frequency mechanical vibration bands to high-frequency ultrasonic bands, so as to fully capture transient stress waves generated when particles of different kinetic energies detach from the surface. The broadband acoustic sensor array is encapsulated with acoustic impedance matching material during installation. The acoustic impedance of the acoustic impedance matching material is matched with the acoustic impedance of the cleaning fluid to reduce the reflection loss of sound waves at the interface of different media. The wideband acoustic sensor array is connected to the back-end signal processor via a high-speed data acquisition card. The sampling frequency of the high-speed data acquisition card is set to be more than twice the highest monitoring frequency to ensure that the waveform of the original acoustic emission signal is not distorted.

3. The semiconductor material cleaning time and particle size adaptive matching method according to claim 2, characterized in that, In the step of constructing a multimodal acoustic fingerprint monitoring environment, the collected acoustic emission signal passes through a gain-adjustable preamplifier circuit before being transmitted to the processor. The preamplifier circuit automatically adjusts the amplification factor according to the amplitude of the received signal. When a large particle is detected to detach and generate a strong signal, the gain is reduced to prevent waveform clipping distortion. When a fine particle is detected to detach and generate a weak signal, the gain is increased to ensure recognition accuracy. The equivalent input noise of the preamplifier circuit is controlled in the microvolt range; The multimodal acoustic fingerprint monitoring environment also includes monitoring and compensation logic for the cleaning fluid temperature. Based on the real-time feedback from the temperature sensor, the system automatically corrects the capture range of the acoustic fingerprint feature frequency, eliminating the impact of temperature drift caused by changes in sound velocity with temperature on particle size recognition accuracy.

4. The semiconductor material cleaning time and particle size adaptive matching method according to claim 3, characterized in that, The steps for real-time decoupling analysis of the captured acoustic emission signal include signal preprocessing, feature extraction, and frequency band separation; The signal preprocessing step uses an adaptive filter to remove low-frequency background noise generated by the mechanical transmission of the cleaning machine and fixed-frequency interference signals directly generated by the megohmmeter generator. The adaptive filter minimizes the noise component in the output signal by updating the internal weight coefficients in real time, and extracts the non-stationary random signal component related to the particulate matter detachment event. The wavelet transform algorithm employs multi-scale decomposition technology, which decomposes the complex original acoustic signal into different frequency scale spaces through a four-layer decomposition structure. The high-frequency detail components of the first and second layers are used to characterize the detachment state of nanoscale particles, while the low-frequency approximation components of the third and fourth layers correspond to the detachment process of micron-sized large particles.

5. The semiconductor material cleaning time and particle size adaptive matching method according to claim 4, characterized in that, The frequency band separation process selects a basis function that is related to the shape of the transient pulse generated by particle detachment as a wavelet basis function, and identifies the energy peak generated when particles of different physical sizes detach by calculating the waveform coefficients at different scales. The pattern recognition algorithm performs real-time comparison based on a pre-built acoustic fingerprint database, which contains detachment feature models of various semiconductor contaminant particles under different cleaning fluid environments. The pattern recognition algorithm decouples the mixed signal into a particle size range signal larger than the first preset particle size, a particle size range signal between the first preset particle size and the second preset particle size, and a particle size range signal smaller than the second preset particle size by calculating the cosine similarity or Euclidean distance between the current real-time signal and the standard model in the database. The pattern recognition algorithm also employs support vector machines for classification optimization, and uses radial basis kernel functions to map low-dimensional nonlinear acoustic features to a high-dimensional feature space, thereby achieving separation of particle detachment signals and cavitation noise signals in high-noise environments.

6. The semiconductor material cleaning time and particle size adaptive matching method according to claim 5, characterized in that, In the step of executing the multi-stage particle size cleaning process, the specific way to divide the cleaning process into multiple continuous cleaning stages is as follows: the first stage is set as the large particle size removal stage, and the acoustic fingerprints larger than the first preset particle size range are monitored. The second stage is set as the medium-particle size removal stage, focusing on monitoring the acoustic fingerprint in the particle size range between the first preset particle size and the second preset particle size. The third stage is set as the micro-residue treatment stage, focusing on monitoring acoustic fingerprints smaller than the second preset particle size range. The initial energy parameters for each cleaning stage are preset based on the surface structure characteristics of the semiconductor material to be cleaned; For semiconductor materials with high aspect ratio trench structures, the initial energy of the first stage is limited to a safe threshold to prevent the cavitation force generated by megasonic waves from breaking the delicate structure. During the third stage, the megasonic power is reduced to a preset percentage of the rated power, and cleaning is performed using the local fluid microfluidic beam effect induced by the megasonic waves.

7. The semiconductor material cleaning time and particle size adaptive matching method according to claim 6, characterized in that, The step of implementing dynamic adaptive adjustment of cleaning energy output is specifically a feedback-based closed-loop control logic; When the acoustic fingerprint intensity corresponding to the large particles monitored in the first stage decays to below a preset threshold, the controller reduces the output power of the megasonic generator and switches the operating frequency of the megasonic generator to the mid-frequency band to transition to the second stage. During the switching between different cleaning stages, the control system synchronously adjusts the flow rate and circulation pressure of the cleaning fluid. By increasing the dynamic pressure of the fluid, it assists in removing medium-sized particles that have been loosened but have not yet completely detached from the surface, while reducing the instantaneous energy density of megasonic waves to protect the fine pattern structure. The switching criteria for each stage combine real-time data from the particle counter in the cleaning fluid. When the acoustic fingerprint shows that the particle detachment activity has stopped and the number of corresponding particles in the discharge fluid detected by the online particle counter drops below the concentration, the stage switching is executed.

8. The semiconductor material cleaning time and particle size adaptive matching method according to claim 7, characterized in that, The step of implementing dynamic adaptive adjustment of cleaning energy output also involves optimization of phase modulation; The control system dynamically adjusts the phase difference between multiple megaacoustic transducers distributed at different positions in the cleaning tank to form a horizontally moving acoustic standing wave in the cleaning tank, eliminating cleaning dead angles and ensuring that particles of different size ranges receive a uniform detachment driving force across the entire semiconductor material surface. The control system records the duration of each particle size cleaning stage in real time and compares it with the stored historical process time data. If the actual duration of a certain stage is longer than the preset reference value, an early warning will be automatically triggered. In addition, the method also includes a real-time health monitoring mechanism, which automatically activates a compensation algorithm to linearly correct the judgment threshold of each frequency band when the background static noise reference shifts by evaluating the sensitivity drift of the acoustic sensor in the cleaning tank online.

9. The semiconductor material cleaning time and particle size adaptive matching method according to claim 8, characterized in that, In the step of determining the end point of the cleaning process and executing the shutdown logic, the determination logic includes a time confirmation window; When the acoustic fingerprint signal intensity corresponding to microparticles smaller than the preset particle size drops to near zero and continues for a predetermined period, the system determines that the surface of the semiconductor material has reached a cleanliness state. The length of the time confirmation window is dynamically adjusted based on the cumulative duration of each of the aforementioned cleaning stages; After confirming that the endpoint has been reached, the system triggers a shutdown command to stop the megason energy output and start the pure water replacement process. The shutdown logic also involves the coordinated control of the rotation speed of the semiconductor material rotating stage. When the cleaning endpoint is determined, the rotation speed of the rotating stage is increased to quickly remove the residual cleaning liquid containing particles using centrifugal force, so as to prevent secondary adsorption and contamination of particles on the surface of the semiconductor material.

10. The semiconductor material cleaning time and particle size adaptive matching method according to claim 9, characterized in that, In the step of executing the shutdown logic, for semiconductor materials with complex three-dimensional structures, a diffusion equilibration stage is added to the shutdown logic; After the acoustic fingerprint disappears, the system maintains a low-flow-rate fluid circulation and activates the heating system to increase the temperature of the cleaning fluid, thereby increasing the diffusion rate of the tiny particles remaining deep in the three-dimensional structure. During this period, if the broadband acoustic sensor array detects the acoustic fingerprint of the detached microparticles again, a short energy compensation pulse is automatically triggered. During the centrifugal drying stage, the system uses a smooth acceleration and deceleration algorithm to control the motor of the rotating platform, avoiding physical damage to the fragile structure caused by the shear force generated by sudden speed changes. Nitrogen gas is introduced to assist drying, ensuring that the position of the nitrogen nozzle is linked with the water injection position, so that the liquid film spreads evenly from the center to the edge.

Images