A method for predicting and evaluating the health of an integrated circuit chemical mechanical polishing machine

Abstract

The application discloses a kind of integrated circuit chemical mechanical polishing machine table health condition prediction evaluation method, sensor data under the health running condition of machine table are used to construct virtual measurement model, in actual production process, actual sensor data is input into model to obtain predicted polishing rate, compared with actual measured polishing rate, to make evaluation to the health state of machine table, make guidance for the maintenance of machine table, solve the stability problem caused by the periodic maintenance of existing experience, waste of consumables.The application uses the method of virtual measurement to construct CMP process model, avoids the construction of complex chemical and physical process model;The model constructed by the application using virtual measurement is used to predict the health condition of machine table, avoids the quality and cost problem caused by the rule of traditional machine maintenance and replacement of consumables, realizes cost reduction and efficiency increase.

Description

Technical Field

[0001] This invention relates to chemical mechanical polishing technology, and in particular to a method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine. Background Technology

[0002] Chemical mechanical polishing (CMP) is a key processing technology in semiconductor chip manufacturing. This process relies on the synergistic effect of chemicals and machinery to achieve ultra-precision material removal from the workpiece surface. A typical CMP system consists of four main parts: a polishing head that holds the workpiece, a polishing disc that carries the polishing pad, a dresser that trims the surface of the polishing pad, and a polishing slurry supply system. During polishing, the workpiece passes through the polishing head and is subjected to a given pressure on a rotating polishing disc with a polishing pad attached, while simultaneously rotating relative to its own axis and oscillating back and forth relative to the polishing disc. The polishing slurry is delivered to the contact interface between the workpiece and the polishing pad, continuously reacting chemically with the workpiece surface. Through the mechanical action of the abrasive particles and the chemical action of the polishing slurry, surface material is removed. A stable material removal rate is crucial for quality control in this semiconductor process.

[0003] The wear and tear of consumables in a chemical mechanical polishing (CMP) machine significantly affects the material removal rate during the CMP process. Therefore, in actual production, consumables need to be replaced to maintain the stability of the process flow when processing different batches of wafers. Consumables in the machine include: polishing pads, retaining rings, wafer adsorption films, and dressing wheels.

[0004] In existing technologies, foundries typically replace consumables at fixed intervals or after processing a fixed number of wafers, but this method has the following problems:

[0005] (1) The replacement cycle is determined by the engineer based on different processing techniques, and requires the assistance of a large number of experienced engineers;

[0006] (2) It is difficult to adjust the replacement time of consumables based on real-time tracking of processing status, which in turn affects the processing consistency of different batches of wafers;

[0007] (3) Considering the stability of the process, consumables will not be used until the end of their lifespan in actual production, resulting in waste of consumables. Summary of the Invention

[0008] Purpose of the Invention: The purpose of this invention is to provide a method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing (CMP) machine. A virtual measurement model is constructed using sensor data from the machine under healthy operating conditions. During actual production, the actual sensor data is input into the model to obtain a predicted polishing rate, which is then compared with the actual measured polishing rate. This allows for an assessment of the machine's health status, providing guidance for machine maintenance and addressing the issues of insufficient stability and wasted consumables associated with existing experience-based periodic maintenance.

[0009] Technical solution: The present invention provides a method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine, comprising the following steps:

[0010] (1) Construction of virtual measurement model: Sensor data and corresponding average MRR values ​​are collected from each wafer processing on the chemical mechanical polishing machine; the collected data are analyzed, outliers are removed, and statistical processing is performed; machine learning methods are used to build a model from the data and train a network model.

[0011] (2) Predictive assessment of the health status of the chemical mechanical polishing machine: Collect sensor data on the chemical mechanical polishing machine during wafer processing in the actual production process; perform statistical processing on the collected data; use the pre-trained network model to obtain the predicted average MRR value; measure the actual average MRR value; compare the predicted average MRR value with the actual average MRR value to assess the health status of the chemical mechanical polishing machine.

[0012] The specific steps (1) are as follows:

[0013] (1.1) Obtaining the material removal sample dataset: This dataset contains n samples, each containing a sensor data points and a corresponding average MRR value. The sensors include a pressure sensor, a fluid flow sensor, a consumable consumption sensor, and a speed sensor. Data collected by the pressure sensor includes CMP polishing head chamber pressure, main external pressure, center pressure, retaining ring pressure, corrugated pressure, and edge pressure. Data collected by the consumable consumption sensor includes polishing pad consumption, retaining ring consumption, dressing wheel consumption, and backing film consumption. Data collected by the speed sensor includes wafer rotation speed, polishing head rotation speed, and step movement speed.

[0014] (1.2) Analyze the data of each material sample and remove outliers.

[0015] (1.3) Extract the mean, standard deviation, skewness and kurtosis of the sensor data of each material sample to obtain 4*a feature vectors.

[0016] (1.4) The obtained 4*a feature vectors are normalized to obtain dataset A.

[0017] (1.5) The XGBoost regression model is used, with the training feature dataset A of n samples as input and the corresponding average MRR value as output, to train the XGBoost network model.

[0018] Step (2) specifically involves:

[0019] (2.1) In the actual production process, when a batch of wafers is ground using a chemical mechanical polishing machine, data from a sensor are collected during the process.

[0020] (2.2) Extract the mean, standard deviation, skewness and kurtosis of the sensor data during this batch of wafer processing to obtain 4*a feature vectors.

[0021] (2.3) The obtained 4*a feature vectors are normalized to obtain the test dataset B.

[0022] (2.4) Input the test dataset B into the trained XGBoost network model to obtain the predicted MRR value.

[0023] (2.5) Measure the actual MRR of the batch of wafers.

[0024] (2.6) Compare the predicted MRR value with the actual MRR value, and calculate the absolute percentage error (APE) to indicate the health status of the chemical mechanical grinding machine at this time.

[0025] (2.7) When APE is lower than L, it indicates that the process is unbalanced and chemical mechanical grinding machine maintenance or consumable replacement is required.

[0026] A computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine.

[0027] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine.

[0028] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0029] 1. This invention uses a virtual measurement method to construct a CMP process model, avoiding the construction of complex chemical and physical process models.

[0030] 2. This invention uses a model constructed by virtual measurement to predict the health status of machines, avoiding the quality and cost problems caused by the traditional rules for machine maintenance and replacement of consumables, and achieving cost reduction and efficiency improvement.

[0031] 3. This invention relies on a data-driven model building method, which can be adjusted according to the equipment status of the contract manufacturer, and is more in line with the differentiated needs of different types of contract manufacturers for processing consistency. Attached Figure Description

[0032] Figure 1 This is a flowchart of the steps of the method described in this invention;

[0033] Figure 2 This is a schematic diagram illustrating the changes in sensor data from a CMP machine.

[0034] Figure 3 This diagram illustrates the comparison between the predictions and actual results from the XGboost model. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] like Figure 1 As shown, a method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine includes the following steps:

[0037] (1) Construction of virtual measurement model: Sensor data and corresponding average MRR values ​​are collected from each wafer processing on the chemical mechanical polishing machine; the collected data are analyzed, outliers are removed, and statistical processing is performed; machine learning methods are used to build a model from the data and train a network model.

[0038] (2) Predictive assessment of the health status of the chemical mechanical polishing machine: Collect sensor data on the chemical mechanical polishing machine during wafer processing in the actual production process; perform statistical processing on the collected data; use the pre-trained network model to obtain the predicted average MRR value; measure the actual average MRR value; compare the predicted average MRR value with the actual average MRR value to assess the health status of the chemical mechanical polishing machine.

[0039] The specific steps (1) are as follows:

[0040] (1.1) As Figure 2As shown, a material removal sample dataset is obtained, containing n samples, each with a sensor data points and a corresponding average MRR value. The sensors include a pressure sensor, a fluid flow sensor, a consumable consumption sensor, and a speed sensor. Data collected by the pressure sensor includes CMP polishing head chamber pressure, main external pressure, center pressure, retaining ring pressure, corrugated pressure, and edge pressure. Data collected by the consumable consumption sensor includes polishing pad consumption, retaining ring consumption, dressing wheel consumption, and backing film consumption. Data collected by the speed sensor includes wafer rotation speed, polishing head rotation speed, and step movement speed.

[0041] (1.2) Analyze the data of each material sample and remove outliers.

[0042] The selected average grinding rate distribution is 50–100 nm / min.

[0043] (1.3) Extract the mean, standard deviation, skewness and kurtosis of the sensor data of each material sample to obtain 4*a feature vectors.

[0044] The statistical measures are: mean E[x], standard deviation σ(x) = E[x - μ] 1 / 2 Skewness s(x) = E[x-μ] 3 / σ 3 Kurtosis k(x) = E[x - μ] 4 / σ 4 .

[0045] (1.4) The obtained 4*a feature vectors are normalized to obtain dataset A.

[0046] The normalization formula is:

[0047] (1.5) The XGBoost regression model is used, with the training feature dataset A of n samples as input and the corresponding average MRR value as output, to train the XGBoost network model.

[0048] XGBoost stands for eXtremeGradientBoosting. Compared to traditional gradient boosting algorithms in machine learning, XGBoost has made many improvements. It is faster than other ensemble algorithms that use gradient boosting and has ultra-high performance in both classification and regression.

[0049] The training parameters for the XGBoost model are: maximum depth of decision trees (max_depth = 5), learning rate (learning_rate = 0.1), number of decision trees (n_estimators = 200), and learning task (objective = 'reg:gamma').

[0050] Using the performance of the XGBoost model with n_estimators = 200 (the number of trees) as the evaluation metric, we attempted to remove features one by one. After removing each feature, we evaluated the thin-film prediction performance of the XGBoost model after training. If R... 2 If an increase in MSE (Mean Sequence Size) and a decrease in MSE indicate that deleting that feature would improve model performance, then it should be deleted; otherwise, it should not be deleted. All features in the final feature set are those that positively impact the performance of the XGBoost model and contain a significant amount of useful information.

[0051] See the comparison between the predictions and actual results of the XGBoost model. Figure 3 .

[0052] Step (2) specifically involves:

[0053] (2.1) In the actual production process, when a batch of wafers is ground using a chemical mechanical polishing machine, data from a sensor are collected during the process.

[0054] (2.2) Extract the mean, standard deviation, skewness and kurtosis of the sensor data during this batch of wafer processing to obtain 4*a feature vectors.

[0055] (2.3) The obtained 4*a feature vectors are normalized to obtain the test dataset B.

[0056] (2.4) Input the test dataset B into the trained XGBoost network model to obtain the predicted MRR value.

[0057] (2.5) Measure the actual MRR of the batch of wafers.

[0058] (2.6) Compare the predicted MRR value with the actual MRR value, and calculate the absolute percentage error (APE) to indicate the health status of the chemical mechanical grinding machine at this time.

[0059] The formula for calculating the absolute percentage error (APE) is:

[0060] (2.7) When APE is lower than L, it indicates that the process is unbalanced and chemical mechanical grinding machine maintenance or consumable replacement is required.

[0061] The L value is set to 80%–90% depending on the process requirements.

[0062] A computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine.

[0063] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine.

Claims

1. A method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine, characterized in that, Includes the following steps: (1) Construction of virtual measurement model: Sensor data and corresponding average MRR values ​​were collected from each wafer processing on the chemical mechanical polishing machine; the collected data were analyzed, outliers were removed, and statistical processing was performed; machine learning methods were used to build a model from the data and train a network model. Step (1) specifically involves: (1.1) Obtain the material removal sample dataset: there are n samples, each containing a sensor data and the corresponding average MRR value; (1.2) Analyze the data of each material removal sample and remove outliers; (1.3) Extract the mean, standard deviation, skewness and kurtosis of the sensor data of each material removal sample to obtain 4*a feature vectors; (1.4) Standardize the obtained 4*a feature vectors to obtain dataset A; (1.5) Use the XGBoost regression model, with the training feature dataset A of n samples as input and the corresponding average MRR value as output, to train the model and obtain the XGBoost network model; (2) Predictive assessment of the health status of chemical mechanical polishing (CMP) machines: Collect sensor data on the CMP machines during wafer processing in actual production process; The collected data underwent statistical processing; The predicted average MRR value is obtained using a pre-trained network model; the actual average MRR value is measured; the predicted average MRR value is compared with the actual average MRR value to assess the health of the chemimechanical grinding machine. Step (2) is as follows: (2.1) In the actual production process, when a batch of wafers is ground using a chemical mechanical polishing machine, a sensor data points are collected during the process; (2.2) The mean, standard deviation, skewness, and kurtosis of the sensor data during the processing of this batch of wafers are extracted to obtain 4*a feature vectors; (2.3) The 4*a feature vectors are normalized to obtain test dataset B; (2.4) Test dataset B is input into the trained XGBoost network model to obtain the predicted MRR value; (2.5) Measure the actual MRR of the batch of wafers; (2.6) Compare the predicted MRR value with the actual MRR value, and calculate the absolute percentage error (APE) to indicate the health status of the chemical mechanical polishing machine at this time; (2.7) When the APE is lower than L, it indicates that the process is unbalanced and the chemical mechanical polishing machine needs maintenance or consumable replacement.

2. The method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine according to claim 1, characterized in that, The sensors include pressure sensors, fluid flow sensors, consumable consumption sensors, and speed sensors.

3. The method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine according to claim 2, characterized in that, The data collected by the pressure sensor includes the CMP grinding head chamber pressure, main external pressure, center pressure, retaining ring pressure, corrugated pressure, and edge pressure.

4. The method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine according to claim 2, characterized in that, The data collected by the fluid flow sensor includes the actual flow rates of three different grinding fluids: A, B, and C.

5. The method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine according to claim 2, characterized in that, The data collected by the consumable consumption sensor includes the consumption of polishing pads, retaining rings, dressing wheels, and backing film.

6. The method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine according to claim 2, characterized in that, The data collected by the speed sensor includes wafer rotation speed, grinding head rotation speed, and step movement speed.

7. A computer storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a method for predicting and assessing the health status of an integrated circuit chemical mechanical polishing machine as described in any one of claims 1-6.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for predicting and evaluating the health status of an integrated circuit chemical mechanical polishing machine as described in any one of claims 1-6.

Images