Dynamic monitoring system for concentration of wafer cleaning solution

Abstract

The invention discloses a wafer cleaning liquid concentration dynamic monitoring system, and belongs to the technical field of semiconductor manufacturing. Aiming at the problems of large single parameter measurement error, poor equipment adaptability and frequent maintenance in the prior art, the invention provides a multi-physics field fusion detection scheme: acquiring ion concentration characteristics through a multi-frequency conductivity detection module (0.1 kHz to 10MHz), eliminating environmental interference in combination with a temperature-viscosity joint compensation module, and calculating the concentration of the ion in the multi-frequency conductivity detection module. An optical refractive index auxiliary verification module is used to realize cross verification, and a CNN-LSTM fusion model optimized by transfer learning is used to improve cross-device prediction precision. According to the innovativeness of the system, an anti-corrosion annular electrode array (titanium alloy iridium-plated) supports dynamic distance adjustment and pulse self-cleaning, an ultrasonic defoaming unit (1-3 MHz frequency self-adaption) restrains bubble interference, and a self-calibration module triggers online maintenance based on the conditions of the electrode impedance change rate (larger than or equal to 5%), temperature fluctuation and the like. Experiments show that the concentration measurement precision is improved to + / -0.5%, the HF / O3 rate ratio is controlled to be (1: 2.5 + / -0.3), the number of residual particles on the surface of the wafer is smaller than or equal to 5 per square inch, the service life of the electrode is prolonged to 8 months, and the yield is improved by 0.6%. The method is suitable for a wafer production line of 12 inches and above, and has the advantages of high precision, strong anti-interference and low maintenance cost.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and particularly relates to a dynamic monitoring system and method for the concentration of a wafer cleaning solution, which is particularly suitable for the precise control and process optimization of the concentration of a mixed cleaning solution of hydrofluoric acid (HF) and ozone (O3). Background Art

[0002] In semiconductor manufacturing, the concentration control of the wafer cleaning solution directly affects the surface quality of the wafer and the device performance. The traditional technology has the following problems:

[0003] Single parameter dependence: The existing technology (such as the patent with publication number CN202023069353.7) calculates the concentration through a single parameter such as conductivity or pH value. However, the wafer cleaning solution (such as the HF / O3 mixed solution) has strong corrosiveness and multi-component dynamic reaction characteristics, and a single parameter is easily interfered by factors such as temperature, bubbles, and electrode contamination, resulting in a measurement error exceeding ±10%.

[0004] Poor equipment adaptability: The volume and fluid disturbance of the cleaning tanks in different production lines vary significantly, and the traditional calibration model cannot be self-adaptive (such as the concentration drift problem mentioned in the literature "Semiconductor Cleaning Technology").

[0005] High maintenance cost: The electrodes are easily corroded and scaled in a strong acid environment, and need to be frequently calibrated manually (2-3 times a week), which affects the production efficiency (such as the data of a certain wafer factory: the yield loss caused by the out-of-control concentration of the cleaning solution reaches 0.8%).

[0006] In summary, the defects in the existing technology include: insufficient measurement accuracy, unable to meet the cleaning requirements for processes below 3nm; lack of multi-parameter fusion and online self-calibration mechanism, poor stability; weak anti-interference ability, and high false alarm rate caused by working conditions such as bubbles and temperature mutations. Summary of the Invention

[0007] The purpose of the present invention is to provide a dynamic monitoring system for the concentration of a wafer cleaning solution to solve the problems raised in the above background art.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A dynamic monitoring system for the concentration of a wafer cleaning solution, comprising:

[0009] Multi-frequency conductivity detection module: An anti-corrosion ring electrode array is adopted, configured to measure the conductivity spectrum of the cleaning solution at different frequencies and extract the high-frequency impedance attenuation characteristics;

[0010] Temperature-viscosity joint compensation module: An integrated wall-mounted temperature sensor and an ultrasonic viscometer are used to obtain the temperature gradient distribution and dynamic viscosity value of the cleaning solution in real time;

[0011] Optical refractive index assisted calibration module: Equipped with a laser interferometer for obtaining data on the change in the refractive index of the cleaning liquid;

[0012] Data processing module: Configured to:

[0013] a) Construct a characteristic matrix of the ion concentration of the cleaning liquid based on the conductivity spectrum;

[0014] b) Generate a dynamic compensation coefficient by fusing temperature and viscosity data;

[0015] c) Calculate the concentration confidence by cross - validating the refractive index data with the conductivity characteristic matrix;

[0016] d) Adopt a convolutional neural network optimized by transfer learning (CNN - LSTM fusion model), input the characteristic matrix, compensation coefficient and confidence, and output the predicted value of the acid - base liquid concentration;

[0017] Self - calibration module, dynamically updates the neural network parameters according to the detection data of the offline mass spectrometer, and generates an electrode pollution warning signal.

[0018] Preferably, the corrosion - resistant annular electrode array includes:

[0019] An iridium layer plated on the surface of a titanium alloy substrate;

[0020] An electrode spacing dynamic adjustment mechanism that adaptively adjusts the electrode spacing according to the conductivity range of the cleaning liquid;

[0021] A pulsed reverse current cleaning unit that periodically eliminates the oxide layer on the electrode surface.

[0022] Preferably, the temperature - viscosity joint compensation module adopts the following algorithm:

[0023] Reconstruct the three - dimensional temperature field in the cleaning tank based on the heat conduction equation;

[0024] Calculate the ion mobility correction coefficient through the viscosity - temperature correlation model;

[0025] Combine the frequency - domain response characteristics of the conductivity spectrum and output the compensated equivalent concentration value.

[0026] Preferably, the CNN - LSTM fusion model optimized by transfer learning includes:

[0027] Feature extraction layer, extracting the local frequency - domain features of the conductivity spectrum through a convolution kernel;

[0028] Temporal correlation layer, using bidirectional LSTM to capture the dynamic law of concentration change;

[0029] Domain adaptation layer, eliminating the influence of equipment differences between different wafer production lines through adversarial training.

[0030] Preferably, a dynamic monitoring system for the concentration of a wafer cleaning solution. It further includes:

[0031] A cleaning solution activity monitoring unit that detects the attenuation rate of ozone concentration in the cleaning solution in real time through an oxidation-reduction potential (ORP) sensor;

[0032] A concentration balance control module that triggers the replenishment valve to adjust the cleaning solution ratio when the detected concentration rate ratio of hydrofluoric acid (HF) to ozone (O3) exceeds a preset threshold.

[0033] Preferably, a method for measuring the concentration of a wafer cleaning solution in a dynamic monitoring system for the concentration of a wafer cleaning solution,

[0034] including:

[0035] Simultaneously collect the conductivity spectrum, temperature-viscosity data, and refractive index change value;

[0036] Map the laboratory calibration data to the real-time production line environment based on a transfer learning model;

[0037] When the concentration confidence level is lower than 90%, start the optical refractive index assisted calibration process;

[0038] Output the concentration value and the corresponding uncertainty evaluation report, and record it in the wafer process traceability system.

[0039] Preferably, the transfer learning model is optimized in the following manner:

[0040] Construct a virtual sample library to simulate different wafer sizes, cleaning tank geometries, and fluid perturbation scenarios;

[0041] Adopt a domain generalization algorithm to improve the model's adaptability to unseen production line equipment.

[0042] Preferably, the method for determining the concentration rate ratio threshold of hydrofluoric acid (HF) to ozone (O3) includes:

[0043] Based on the correlation experiment between the surface roughness of the wafer and the cleaning efficiency, set the molar concentration change rate ratio of HF:O3 to be 1:2 to 1:3;

[0044] When it is detected that the rate ratio exceeds the threshold range, the adjustment amount ΔQ of the replenishment valve is calculated according to the following formula:

[0045]

[0046] where:

[0047] kk is the proportionality coefficient (dynamically adjusted through the transfer learning model),

[0048] αα is the ideal rate ratio (take 1:2.5),

[0049] Vtank: The volume of the cleaning tank;

[0050] The experimental data support for the threshold range includes: at a rate ratio of 1:2.5, the particle residue on the wafer surface is ≤ 5 particles per square inch, and when the range exceeds 1:1.8 - 1:3.2, the surface defect rate increases to more than 0.5%.

[0051] Preferably, the ultrasonic defoaming unit includes:

[0052] A piezoelectric ceramic transducer array disposed on the side wall of the cleaning tank, with an operating frequency of 1 - 3 MHz;

[0053] A bubble detection module that determines the bubble density in real time through laser scattering intensity analysis;

[0054] A defoaming control logic that, when the bubble density exceeds 200 bubbles / cm 3 initiates frequency modulation linked to the fluid viscosity:

[0055]

[0056] Wherein:

[0057] f0 is the reference frequency (1.5 MHz),

[0058] η is the viscosity of the cleaning liquid detected in real time,

[0059] η0 is the calibrated viscosity (taking 1.2 cP).

[0060] Preferably, the trigger conditions for the self - calibration module include:

[0061] Monitoring of the electrode impedance change rate: when the change rate of the AC impedance modulus value of the annular electrode array

[0062] Δ|Z| / |Z0|≥5% (calculation formula:

[0063]

[0064] where the baseline impedance |Z0| takes the system initial calibration value);

[0065] The cumulative continuous working time exceeds 72 hours;

[0066] The temperature fluctuation exceeds ±2°C per minute;

[0067] When any condition is met, perform the following operations:

[0068] a) Generate an electrode contamination warning signal and mark the position of the abnormal electrode;

[0069] b) Initiate the pulsed reverse current cleaning process (current density 5 - 10 mA / cm 2 , duration 30 - 60 seconds); c) Switch to the backup electrode group and recalibrate the measurement reference.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] 1. Improved accuracy:

[0072] Multi-parameter cross-validation (conductivity - refractive index) increases the confidence level to 99%;

[0073] The HF / O3 concentration rate ratio control (1:2.5 ± 0.3) makes the particle residue on the wafer surface ≤ 5 per square inch (SEMI standard).

[0074] 2. Enhanced stability:

[0075] The self-calibration module reduces the frequency of manual intervention (from 3 times a week to 1 time a month);

[0076] The anti-corrosion electrode design reduces the replacement cost by more than 60%.

[0077] 3. Adaptability to the production line:

[0078] The transfer learning model supports cross-device deployment, and the adaptation time is shortened from 24 hours to 2 hours;

[0079] The dynamic viscosity compensation enables the system to remain stable in an ambient temperature range of 5 - 50 °C. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 is a block diagram of the system architecture of the present invention;

[0081] Figure 2 : A cross-sectional view of the anti-corrosion electrode structure;

[0082] Figure 3 : A flowchart of the transfer learning model training;

[0083] Figure 4 : A logic diagram of the frequency adaptive control of the defoaming unit.

[0084] In the figure: 1 titanium alloy substrate, 2 iridium coating, 3, bidirectional pulsed cleaning current. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0085] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0086] Please refer to Figure 1 , the present invention provides a technical solution: a dynamic monitoring system for the concentration of a wafer cleaning liquid, including:

[0087] Multi-frequency conductivity detection module: Adopting an anti-corrosion ring electrode array, configured to measure the conductivity spectrum of the cleaning liquid at different frequencies and extract the high-frequency impedance attenuation characteristics;

[0088] Temperature-viscosity joint compensation module: Integrating a wall-mounted temperature sensor and an ultrasonic viscometer to obtain the temperature gradient distribution and dynamic viscosity value of the cleaning liquid in real time;

[0089] Optical refractive index auxiliary calibration module: Equipped with a laser interferometer for obtaining the refractive index change data of the cleaning liquid;

[0090] Data processing module: Configured as:

[0091] a) Construct an ion concentration characteristic matrix of the cleaning liquid based on the conductivity spectrum;

[0092] b) Generate a dynamic compensation coefficient by fusing temperature and viscosity data;

[0093] c) Calculate the concentration confidence through the cross-validation of the refractive index data and the conductivity characteristic matrix;

[0094] d) Adopt a convolutional neural network optimized by transfer learning (CNN-LSTM fusion model), input the

[0095] characteristic matrix, compensation coefficient and confidence, and output the predicted value of the acid-base liquid concentration;

[0096] Self-calibration module, dynamically update the neural network parameters according to the detection data of the offline mass spectrometer, and generate an electrode pollution warning signal; The system also includes:

[0097] Cleaning liquid activity monitoring unit, which detects the ozone concentration attenuation rate in the cleaning liquid in real time through an oxidation-reduction potential (ORP) sensor;

[0098] Concentration balance control module, when the concentration rate ratio of hydrofluoric acid (HF) and ozone (O3) is detected to exceed the preset threshold, trigger the replenishing valve to adjust the cleaning liquid ratio.

[0099] Figure 2As shown, the corrosion-resistant annular electrode array includes: an iridium layer plated on the surface of a titanium alloy substrate;

[0100] An electrode spacing dynamic adjustment mechanism that adaptively adjusts the electrode spacing according to the conductivity range of the cleaning liquid;

[0101] A pulsed reverse current cleaning unit that periodically eliminates the oxide layer on the electrode surface.

[0102] The temperature-viscosity combined compensation module adopts the following algorithm:

[0103] Reconstruct the three-dimensional temperature field in the cleaning tank based on the heat conduction equation;

[0104] Calculate the ion mobility correction coefficient through the viscosity-temperature correlation model;

[0105] Combined with the frequency-domain response characteristics of the conductivity spectrum, output the compensated equivalent concentration value.

[0106] The CNN-LSTM fusion model optimized by transfer learning includes:

[0107] The CNN-LSTM fusion model is composed of the following core components ( Figure 3 as shown):

[0108] Feature extraction layer (CNN module):

[0109] Input: Conductivity frequency-domain feature matrix (6 frequency points × 3 dimensions)

[0110] Structure: 3-layer convolution (kernel size 3×3, stride 1) + max pooling

[0111] Function: Extract local frequency-domain response characteristics (such as high-frequency impedance attenuation characteristics)

[0112] Spatio-temporal feature extraction layer (bidirectional LSTM module):

[0113] Input: Feature sequence output by CNN (time step = number of sampling periods)

[0114] Structure: 2-layer bidirectional LSTM (number of hidden units 128)

[0115] Function: Capture the dynamic time-series law of concentration change

[0116] Domain adaptation layer (adversarial training module):

[0117] Gradient reversal layer (GRL): Reverse the domain classification gradient during backpropagation

[0118] Domain classifier: 2-layer fully connected network, discriminates the sample source (laboratory / production line A / production line B)

[0119] Feature extractor: The optimization objective is to confuse the domain classifier and force the generation of domain-invariant features.

[0120] Adversarial training process

[0121] Step 1: Multi-source data preparation

[0122] Source domain: Laboratory-calibrated data (500 groups, including full parameters of conductivity, temperature, and refractive index)

[0123] Target domain: Unlabeled real-time production line data (only conductivity and temperature)

[0124] Step 2: Joint optimization objective

[0125] The total loss function of the model is:

[0126] L = L task -λL domain

[0127] L task : Mean squared error (MSE) of concentration prediction;

[0128] L domain : Cross-entropy loss of domain classification

[0129] λλ: Dynamically adjust the weight (initial value 0.1, linearly increased to 1.0)

[0130] Step 3: Adversarial training process

[0131] 1. Forward propagation:

[0132] The feature extractor generates domain-invariant features;

[0133] The domain classifier predicts the sample source;

[0134] The task predictor outputs the concentration value.

[0135] 2. Backward propagation:

[0136] Reverse the domain classification gradient through GRL to make the feature extractor learn to deceive the domain classifier;

[0137] The task prediction gradient is updated normally to maintain the concentration prediction ability.

[0138] Step 4: Dynamic weight strategy

[0139] When the domain classification accuracy > 85%, increase λλ to strengthen feature alignment;

[0140] When the task prediction error rises, decrease λλ to prevent over-alignment.

[0141] The transfer learning model is optimized in the following way:

[0142] Build a virtual sample library to simulate different wafer sizes, cleaning tank geometries, and fluid disturbance scenarios;

[0143] Adopt the Domain Generalization algorithm to improve the model's adaptability to unseen production line equipment.

[0144] The method for determining the concentration rate ratio threshold of hydrofluoric acid (HF) and ozone (O3) includes:

[0145] Based on the correlation experiment between wafer surface roughness and cleaning efficiency, set the molar concentration change rate ratio of HF:O3 to be 1:2 to 1:3;

[0146] When it is detected that the rate ratio exceeds the threshold range, the adjustment amount ΔQ of the liquid supply valve is calculated according to the following formula:

[0147]

[0148] Where:

[0149] kk is the proportionality coefficient (dynamically adjusted by the transfer learning model),

[0150] αα is the ideal rate ratio (take 1:2.5),

[0151] VtankVtank is the volume of the cleaning tank;

[0152] The experimental data support for the threshold range includes: at a rate ratio of 1:2.5, the particle residue on the wafer surface ≤ 5 per square inch, and when exceeding the range of 1:1.8 - 1:3.2, the surface defect rate increases to more than 0.5%.

[0153] A method for measuring the concentration of the wafer cleaning liquid in a wafer cleaning liquid concentration dynamic monitoring system:

[0154] Synchronously collect the conductivity spectrum, temperature-viscosity data, and refractive index change value;

[0155] Map the laboratory calibration data to the real-time production line environment based on the transfer learning model;

[0156] When the concentration confidence level is lower than 90%, start the optical refractive index auxiliary calibration process;

[0157] Figure 4 As shown, output the concentration value and the corresponding uncertainty evaluation report, and record it in the wafer process traceability system.

[0158] The ultrasonic defoaming unit includes:

[0159] A piezoelectric ceramic transducer array disposed on the side wall of the cleaning tank, with a working frequency of 1 - 3 MHz;

[0160] Bubble detection module, which judges the bubble density in real time through laser scattering intensity analysis;

[0161] Defoaming control logic. When the bubble density exceeds 200 bubbles / cm 3 , start the frequency modulation linked to the fluid viscosity:

[0162]

[0163] Where:

[0164] f0f0 is the reference frequency (the reference frequency (1.5 MHz),

[0165] ηη is the viscosity of the cleaning liquid detected in real time,

[0166] η0η0 is the calibrated viscosity (take 1.2 cP).

[0167] The trigger conditions of the self-calibration module include:

[0168] Monitoring the change rate of electrode impedance: When the change rate of the AC impedance modulus value of the annular electrode array

[0169] Δ|Z| / |Z0|≥5% (Calculation formula:

[0170]

[0171] where the baseline impedance |Z0| takes the system initial calibration value);

[0172] The cumulative continuous working time exceeds 72 hours;

[0173] The temperature fluctuation exceeds ±2°C / minute;

[0174] When any condition is met, perform the following operations:

[0175] a) Generate an electrode pollution warning signal and mark the abnormal electrode position;

[0176] b) Start the pulsed reverse current cleaning program (current density 5-10 mA / cm 2 , duration 30-60 seconds); c) Switch to the standby electrode group and recalibrate the measurement reference.

[0177] Example 1: Concentration control of hydrofluoric acid-ozone mixture

[0178] Step 1: Data acquisition

[0179] In a 300 mm wafer cleaning tank, start the multi-frequency conductivity detection module (0.1 kHz, 1 kHz, 10 kHz,

[0180] 100 kHz, 1 MHz, 10 MHz);

[0181] Synchronously collect temperature gradient data (adhesive sensors, spacing 10 cm) and ultrasonic viscosity values (sampling rate 10 Hz);

[0182] The laser interferometer records the refractive index change at a speed of 5 frames per second.

[0183] Step 2: Data processing

[0184] Construct a conductivity frequency-domain feature matrix with a size of 6 (frequency points) × 3 (real part / imaginary part / phase angle);

[0185] Reconstruct the three-dimensional temperature field through the heat conduction equation, and calculate the dynamic compensation coefficient in combination with the viscosity data

[0186] α = η / (η0·T 2 )

[0187] Cross-validate the refractive index data with the conductivity matrix. If the deviation > 0.5%, start the optical auxiliary calibration process.

[0188] Step 3: Concentration prediction and control

[0189] Input the feature matrix into the CNN-LSTM model (pre-trained on laboratory data and adapted to production line equipment through transfer learning);

[0190] When the detected HF:O3 rate ratio > 1:1.8, according to the formula

[0191] ΔQ = 0.2 × |(d[HF] / dt) / (d[O3] / dt) - 2.5| × V_tank to adjust the liquid supplement volume;

[0192] Output the concentration value (accuracy ±0.05 mol / L) and 95% confidence interval in real time.

[0193] Example 2: Bubble interference suppression

[0194] Defoaming control process:

[0195] 1. The laser scattering detection module judges the bubble density: when > 200 bubbles / cm 3 , trigger the piezoelectric ceramic transducer array;

[0196] 2. According to the real-time viscosity η (detected value 1.8 cP), adjust the frequency according to the formula f = 1.5 MHz × √(1.8 / 1.2) = 1.84 MHz;

[0197] 3. After continuous defoaming for 30 seconds, re-verify the refractive index data until the deviation < 0.1%.

[0198] Example 3: Self-calibration module operation

[0199] Trigger conditions and responses:

[0200] When the electrode impedance change rate Δ|Z| = 7% (baseline impedance |Z0| = 100 Ω), the system performs:

[0201] a) Mark the abnormal electrode position (such as electrode pair #3);

[0202] b) Start the pulsed reverse current cleaning (current density 8 mA / cm 2 , duration 45 seconds);

[0203] c) Switch to the backup electrode group #4 and recalibrate the baseline value.

[0204] Data comparison after calibration: The deviation of the concentration measurement value is restored from 1.2% to 0.3%.

[0205] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic monitoring system for the concentration of a wafer cleaning solution, characterized in that, Including: Multi-frequency conductivity detection module: Adopting an anti-corrosion annular electrode array, configured to measure the conductivity spectrum of the cleaning liquid at different frequencies and extract the high-frequency impedance attenuation characteristics; Temperature-viscosity combined compensation module: Integrating a wall-mounted temperature sensor and an ultrasonic viscometer to obtain the temperature gradient distribution and dynamic viscosity value of the cleaning liquid in real time; Optical refractive index auxiliary calibration module: Equipped with a laser interferometer for obtaining data on the change in the refractive index of the cleaning liquid; Data processing module: Configured to: a) Construct an ion concentration characteristic matrix of the cleaning liquid based on the conductivity spectrum; b) Generate a dynamic compensation coefficient by fusing temperature and viscosity data; c) Calculate the concentration confidence level through cross-validation of the refractive index data and the conductivity characteristic matrix; d) Adopt a convolutional neural network optimized by transfer learning (CNN-LSTM fusion model), input the characteristic matrix, compensation coefficient, and confidence level, and output the predicted value of the acid-base liquid concentration; Self-calibration module, dynamically updating the neural network parameters according to the detection data of the offline mass spectrometer and generating an electrode contamination warning signal.

2. The dynamic monitoring system for the concentration of a wafer cleaning solution according to claim 1, characterized in that The anti-corrosion annular electrode array includes: An iridium coating on the surface of a titanium alloy substrate; An electrode spacing dynamic adjustment mechanism that adaptively adjusts the electrode spacing according to the conductivity range of the cleaning liquid; A pulsed reverse current cleaning unit that periodically eliminates the oxide layer on the electrode surface.

3. The dynamic monitoring system for the concentration of a wafer cleaning liquid as claimed in claim 1, wherein, The temperature-viscosity combined compensation module adopts the following algorithm: Reconstruct the three-dimensional temperature field in the cleaning tank based on the heat conduction equation; Calculate the ion mobility correction coefficient through the viscosity-temperature correlation model; Combined with the frequency-domain response characteristics of the conductivity spectrum, output the compensated equivalent concentration value.

4. A dynamic monitoring system for the concentration of a wafer cleaning solution according to claim 1, characterized in that, The CNN-LSTM fusion model optimized by transfer learning includes: Feature extraction layer, extracting the local frequency-domain features of the conductivity spectrum through convolutional kernels; Temporal correlation layer, using bidirectional LSTM to capture the dynamic law of concentration change; Domain adaptation layer, eliminating the influence of equipment differences between different wafer production lines through adversarial training.

5. The dynamic monitoring system for the concentration of a wafer cleaning solution according to claim 1, characterized in that Also including: Cleaning liquid activity monitoring unit, real-time detecting the ozone concentration decay rate in the cleaning liquid through an oxidation-reduction potential (ORP) sensor; Concentration balance control module, triggering the replenishing valve to adjust the cleaning liquid ratio when the detected concentration rate ratio of hydrofluoric acid (HF) to ozone (O3) exceeds the preset threshold.

6. A method for measuring the concentration of a wafer cleaning liquid based on the wafer cleaning liquid concentration dynamic monitoring system according to any one of claims 1-5, characterized in that, Including: Synchronously collecting the conductivity spectrum, temperature-viscosity data, and refractive index change value; Mapping the laboratory calibration data to the real-time production line environment based on the transfer learning model; When the concentration confidence level is lower than 90%, starting the optical refractive index auxiliary calibration process; Outputting the concentration value and the corresponding uncertainty evaluation report, and recording them in the wafer process traceability system.

7. The wafer cleaning liquid concentration measurement method according to claim 6, characterized in that, The transfer learning model is optimized in the following ways: Constructing a virtual sample library to simulate different wafer sizes, cleaning tank geometries, and fluid disturbance scenarios; adopting a domain generalization algorithm to improve the model's adaptability to unseen production line equipment.

8. A dynamic monitoring system for the concentration of a wafer cleaning solution according to claim 5, characterized in that, The method for determining the concentration rate ratio threshold of hydrofluoric acid (HF) to ozone (O3) includes: Based on the correlation experiment between the surface roughness of the wafer and the cleaning efficiency, setting the molar concentration change rate ratio of HF:O3 to be 1:2 to 1:3; When the detected rate ratio exceeds the threshold range, the adjustment amount ΔQ of the liquid replenishment valve is calculated according to the following formula: Where: kk is the proportionality coefficient (dynamically adjusted by the transfer learning model), αα is the ideal rate ratio (take 1:2.5), VtankVtank is the volume of the cleaning tank; The experimental data support for the threshold range includes: at a rate ratio of 1:2.5, the particle residue on the wafer surface is ≤ 5 per square inch, and when it exceeds the range of 1:1.8 - 1:3.2, the surface defect rate increases to more than 0.5%.

9. The dynamic monitoring system for the concentration of a wafer cleaning liquid according to claim 1, characterized in that It also includes an ultrasonic defoaming unit, and the ultrasonic defoaming unit includes: A piezoelectric ceramic transducer array arranged on the side wall of the cleaning tank, with an operating frequency of 1 - 3 MHz; A bubble detection module that judges the bubble density in real time by analyzing the laser scattering intensity; Defoaming control logic. When the bubble density exceeds 200 bubbles / cm 3 , start the frequency modulation linked to the fluid viscosity: Where: f0f0 is the reference frequency (1.5 MHz), ηη is the viscosity of the cleaning liquid detected in real time, η0η0 is the calibrated viscosity (take 1.2 cP).

10. A dynamic monitoring system for the concentration of a wafer cleaning solution according to claim 1, characterized in that, The trigger conditions of the self-calibration module include: Monitoring the change rate of electrode impedance: when the change rate of the AC impedance modulus value of the annular electrode array Δ|Z| / |Z0|≥5% (calculation formula: where the baseline impedance |Z0| takes the initial calibration value of the system); The cumulative continuous working time exceeds 72 hours; The temperature fluctuation exceeds ±2°C per minute; When any condition is met, perform the following operations: a) Generate an electrode pollution warning signal and mark the position of the abnormal electrode; b) Start the pulsed reverse current cleaning program (current density 5 - 10 mA / cm 2 , duration 30 - 60 seconds); c) Switch to the standby electrode group and re-calibrate the measurement reference.

Images