Machine learning based semiconductor manufacturing defect real-time prediction method and system

Abstract

The application discloses a kind of based on machine learning's semiconductor manufacturing defect real-time prediction method and system, belong to the cross field of semiconductor manufacturing and machine learning technique, the application is integrated the sensor time series data of semiconductor manufacturing multi-process by federal learning framework, using improved type LSTM-Transformer fusion time series model, realize the real-time prediction to specific wafer defect type and position, complete from "after the event detection" to "before prediction" technical change, process parameter adjustment is triggered in advance, significantly improve wafer manufacturing yield.The application does not need to share each process core sensitive data, give consideration to data privacy protection and multi-source data collaborative use, prediction response speed is fast, precision is high.

Description

Technical Field

[0001] This invention relates to the intersection of semiconductor manufacturing technology and machine learning technology, and in particular to a method and system for real-time prediction of semiconductor manufacturing defects based on machine learning. Background Technology

[0002] Semiconductor manufacturing is a complex, multi-stage collaborative process encompassing dozens of core steps, including photolithography, etching, deposition, ion implantation, and polishing. Even minute fluctuations in the process parameters (temperature, pressure, gas flow rate, etc.) of each step can lead to wafer defects, thus affecting product yield. Currently, the semiconductor manufacturing industry's defect management model primarily relies on "post-production inspection," which involves detecting, classifying, and processing defects after they occur through methods such as Adaptive Difference (ADI) and Advanced Etching Inspection (AEI). This model has significant technological limitations.

[0003] In existing technologies, some solutions attempt to use machine learning methods for defect prediction, but most suffer from the following problems: First, insufficient data utilization, analyzing only sensor data from a single process without achieving collaborative integration of data from multiple processes, failing to capture the impact of multi-process interactions on defects, resulting in low prediction accuracy; second, insufficient data privacy protection, as sensor data from each process in semiconductor manufacturing constitutes core sensitive information for enterprises, and direct centralized collection and analysis can easily lead to data leakage, failing to meet enterprise data security requirements; third, prediction lag, as existing models mostly rely on historical data for offline prediction, failing to achieve real-time response and failing to meet the need for real-time adjustment of process parameters during semiconductor mass production; fourth, poor adaptability, as semiconductor process nodes advance to 5nm and below, defect sizes shrink to the atomic level, traditional prediction models struggle to capture the defect risks caused by minute process fluctuations and cannot accurately locate defect positions, limiting their practicality.

[0004] Furthermore, in advanced processes such as the 3-nanometer node, traditional detection methods also face a conflict between detection sensitivity and sample integrity. Improving detection sensitivity requires increasing the electron dose, but excessive dose can lead to irreversible degradation of the photoresist, causing a sharp increase in false defect false alarms. At the same time, there are gaps in the process interaction diagnosis level, making it difficult to distinguish which process the defect originates from, causing process debugging to remain at the trial and error level, resulting in a large number of wafer scraps and cost waste. Summary of the Invention

[0005] The purpose of this invention is to propose a real-time prediction method for semiconductor manufacturing defects based on machine learning, which aims to solve the problems of delayed prediction of semiconductor manufacturing defects, fragmented data across multiple processes, insufficient data privacy protection, and low prediction accuracy in the prior art.

[0006] This invention is implemented as follows: a real-time prediction method for semiconductor manufacturing defects based on machine learning, the method comprising the following steps: Deploy multiple types of sensor groups in the core processes of semiconductor manufacturing to collect multi-dimensional time-series data of the corresponding processes in real time during wafer manufacturing; The collected time-series data is cleaned, normalized, and aligned to a unified timeline, generating a standardized time-series dataset. To construct and train a federated learning model; The multi-process time-series data, which are collected and preprocessed in real time, are input into the trained global prediction model. Based on the dynamic change characteristics of the time-series data, the model outputs the defect type, defect location and defect occurrence probability corresponding to a specific wafer in real time. Based on preset process adjustment trigger conditions, the system automatically generates process parameter adjustment suggestions, associates them with corresponding process nodes and sensor data anomalies, and pushes the adjustment suggestions to the equipment controller of the corresponding process in real time, thereby achieving real-time adaptive adjustment of process parameters.

[0007] Another objective of this invention is to propose a real-time prediction method system for semiconductor manufacturing defects based on machine learning; the system includes: The data acquisition module is used to deploy corresponding multi-type sensor groups in the core processes of semiconductor manufacturing to collect multi-dimensional time-series data of the corresponding processes in the wafer manufacturing process in real time. The data preprocessing module is used to clean, normalize, and align the collected time-series data, unifying the time-series data from different processes and sensors to the same time axis, and generating a standardized time-series dataset. The Federated Learning Model Building and Training module is used to build and train federated learning models. The real-time prediction module is used to input the multi-process time-series data that has been collected and preprocessed in real time into the trained global prediction model. Based on the dynamic change characteristics of the time-series data, the model outputs the defect type, defect location and defect occurrence probability corresponding to a specific wafer in real time. The process adjustment module is used to automatically generate process parameter adjustment suggestions based on preset process adjustment trigger conditions, associate the corresponding process nodes and sensor data anomalies, and push the adjustment suggestions to the equipment controller of the corresponding process in real time to achieve real-time adaptive adjustment of process parameters.

[0008] Beneficial effects of the present invention This invention discloses a real-time prediction method and system for semiconductor manufacturing defects based on machine learning, belonging to the interdisciplinary field of semiconductor manufacturing and machine learning technology. The method integrates sensor time-series data from multiple semiconductor manufacturing processes using a federated learning architecture and employs an improved LSTM-Transformer fusion time-series model to achieve real-time prediction of specific wafer defect types and locations. This represents a technological shift from "post-detection" to "pre-prediction," triggering process parameter adjustments in advance and significantly improving wafer manufacturing yield. This invention eliminates the need to share core sensitive data from each process, balancing data privacy protection with the collaborative utilization of multi-source data, and offers fast prediction response and high accuracy. Attached Figure Description

[0009] Figure 1 This is a flowchart of a preferred embodiment of the present invention: a real-time prediction method for semiconductor manufacturing defects based on machine learning. Figure 2 This is a structural diagram of a machine learning-based real-time prediction system for semiconductor manufacturing defects according to a preferred embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. For ease of explanation, only the parts related to the embodiments of this invention are shown. It should be understood that the specific embodiments described herein are merely for explaining this invention and are not intended to limit this invention.

[0011] This invention proposes a real-time prediction method and system for semiconductor manufacturing defects based on machine learning, belonging to the interdisciplinary field of semiconductor manufacturing and machine learning technology. The method integrates sensor time-series data from multiple semiconductor manufacturing processes using a federated learning architecture and employs an improved LSTM-Transformer fusion time-series model to achieve real-time prediction of specific wafer defect types and locations. This represents a technological shift from "post-detection" to "pre-detection," triggering process parameter adjustments in advance and significantly improving wafer manufacturing yield. The method eliminates the need to share core sensitive data from each process, balancing data privacy protection with the collaborative utilization of multi-source data, and offers fast prediction response and high accuracy.

[0012] Figure 1 This is a flowchart of a preferred embodiment of the present invention, a method for real-time prediction of semiconductor manufacturing defects based on machine learning; it includes the following steps: S1, deploys multiple types of sensor groups in the core processes of semiconductor manufacturing to collect multi-dimensional time-series data of the corresponding processes in the wafer manufacturing process in real time; The core processes include, but are not limited to, photolithography, etching, deposition, ion implantation, and polishing. The timing data includes, but is not limited to, temperature, pressure, gas flow rate, RF power, robot position, vacuum level, and wafer transfer speed; Each sensor data stream carries a unique process identifier, sensor identifier, and nanosecond-level timestamp. Real-time data transmission can be achieved using both SECS / GEM and OPC UA protocols. Time synchronization of data from multiple devices and sensors can be achieved through the IEEE 1588 PTP protocol. The acquisition frequency is dynamically adjusted according to the importance of the process to ensure accurate capture of data from critical processes.

[0013] For example, in one embodiment of the present invention, multiple types of sensor groups are deployed in the core processes of semiconductor manufacturing (photolithography, etching, deposition, ion implantation, and polishing). The photolithography process deploys temperature sensors, RF power sensors, and gas flow sensors; the etching process deploys pressure sensors, vacuum sensors, and gas flow sensors; and the deposition, ion implantation, and polishing processes deploy sensors adapted to the requirements of each process, such as temperature, position, and transmission speed. Multi-dimensional time-series data corresponding to each process are collected in real time, including the coating temperature and RF power in the photolithography process, the etching pressure and vacuum in the etching process, the robot position in the deposition process, the gas flow rate in the ion implantation process, and the wafer transfer speed in the polishing process. Each sensor data stream carries a unique process identifier (e.g., "photolithography-01", "etching-02"), a sensor identifier (e.g., "temperature-001", "vacuum-002"), and a nanosecond-level timestamp. The collected data is transmitted in real time to the edge gateway using both SECS / GEM and OPC UA protocols via IEEE 1588. The PTP protocol enables data synchronization across all sensors and devices. The acquisition frequency is dynamically adjusted based on the importance of the process. Specifically, the acquisition frequency for photolithography and etching processes is set to 10ms / time, for deposition and ion implantation processes it is set to 50ms / time, and for polishing processes it is set to 100ms / time, ensuring accurate capture of data from critical processes.

[0014] S2, cleans, normalizes and aligns the collected time-series data, removes outliers, missing values ​​and noisy data, uses the rotating door algorithm to compress high-frequency data, and aligns time-series data from different processes and sensors to the same time axis to generate a standardized time-series dataset. The standardized time-series dataset is associated with a unique wafer ID and process node information; In this embodiment of the invention, the 3σ criterion can be used to remove outliers from the collected time-series data. Data that exceeds the range of [μ-3σ,μ+3σ] (such as abrupt data caused by sensor failure) is deleted, where μ is the data mean and σ is the data standard deviation. For missing values ​​in the data, timing alignment can be based on the time when the wafer enters each process step, and linear interpolation can be used to supplement them to ensure the continuity of the data. The rotating door algorithm can be used to compress high-frequency acquired data, reducing data storage and processing pressure while retaining key features; The Min-Max normalization method can be used to map all time series data to the [0,1] interval, eliminate the difference in the dimensions of data of different dimensions, and generate a standardized time series dataset. This dataset is associated with a unique wafer ID and the corresponding process node information, providing standardized input for subsequent model training and prediction.

[0015] S3 is used for building and training federated learning models; Specifically, a federated learning architecture is constructed, including a central server and multiple process-end local nodes. Each process-end local node trains a locally improved LSTM-Transformer fusion time-series model based on its own preprocessed standardized time-series dataset and combined with historical defect annotation data (including defect type, defect location, and corresponding process time-series data). The fusion model includes an LSTM feature extraction layer, a Transformer attention mechanism layer, and a defect prediction output layer. The LSTM feature extraction layer captures the long-term and short-term dependencies of the time-series data, adapting to the dynamic changes in semiconductor manufacturing time-series data. The Transformer attention mechanism layer strengthens the feature weights of key process data, solving the problem of decreased prediction accuracy caused by data redundancy across multiple processes. The defect prediction output layer uses a softmax activation function to output the defect type, defect location, and probability of occurrence. Each local node only uploads model parameters to the central server and does not transmit raw sensitive data. The central server aggregates and optimizes the model parameters uploaded by each local node to generate a global prediction model (i.e., the trained and optimized federated learning model). The central server encrypts and distributes the aggregated global model parameters to each local node. Each local node updates its local model based on the global model parameters and enters the next round of iterative training. Through multiple rounds of iterative training, the prediction accuracy of the global prediction model meets the preset threshold (e.g., ≥98.5%).

[0016] In one embodiment of the present invention, the central server uses a federated average algorithm to weight and aggregate the model parameters uploaded by all local nodes according to the defect impact weight of each process (e.g., lithography process weight 0.35, etching process weight 0.25, deposition process weight 0.15, ion implantation process weight 0.15, polishing process weight 0.10) to generate a global prediction model. For example, a federated learning architecture is constructed, consisting of one central server and five local nodes at the process ends: lithography, etching, deposition, ion implantation, and polishing. The lithography local nodes train a locally improved LSTM-Transformer fusion temporal model based on their pre-processed standardized time-series datasets of coating temperature and RF power, combined with historical labeled data of particle defects and bubble defects and their corresponding time-series data. In this model, the LSTM feature extraction layer captures the fluctuation pattern of coating temperature over time, the Transformer attention mechanism layer strengthens the feature weights of the key parameter RF power, and the defect prediction output layer outputs the defect type and wafer information through a softmax activation function. The three-dimensional coordinates and probability of occurrence are determined. The local nodes of the other four processes, including etching and deposition, train their local models synchronously based on the standardized time-series data and corresponding historical defect annotation data of their respective processes. Each node only uploads the trained model parameters to the central server without transmitting any raw sensor data or defect annotation data. The central server uses a federated averaging algorithm to aggregate and optimize the model parameters of the five nodes to generate a global prediction model. The aggregated global model parameters are then encrypted and distributed to each local node. After each node updates its local model, it enters the next iteration. After 86 iterations, the prediction accuracy of the global prediction model reaches 98.7%, which meets the preset threshold of ≥98.5%, and training stops.

[0017] S4 inputs the real-time collected and preprocessed multi-process time series data into the trained global prediction model. Based on the dynamic change characteristics of the time series data, the model outputs the defect type, defect location and defect occurrence probability corresponding to a specific wafer in real time. Specifically, real-time collected and preprocessed multi-process time-series data is input into a trained global prediction model. Based on the dynamic changes of the time-series data, the model outputs in real time the defect type, defect location, and defect occurrence probability for a specific wafer. The defect location is accurately located and uniquely identified using a globally unified wafer three-dimensional coordinate system (XYZ). The X and Y axes represent the planar position of the defect on the wafer surface, and the Z axis represents the internal interlayer position of the defect along the wafer thickness direction. This is used to distinguish whether the defect originates on the wafer surface, in the dielectric layer, polysilicon layer, metal interconnect layer, or inside the silicon substrate, achieving full-domain defect spatial positioning from the wafer surface to each internal functional layer with a positioning accuracy of ≤10nm. At the same time, the defect occurrence probability output by the model can intuitively reflect the risk level of defect occurrence, providing a precise basis for triggering subsequent process adjustments.

[0018] Among them, the defect types are classified by spatial morphology into point defects, line defects, surface defects, and volume defects; including but not limited to particle defects, scratch defects, bubble defects, crack defects, lattice defects (vacancies, interstitial atoms, dislocations), interlayer voids, and doping concentration deviations. Each type of defect corresponds to the above spatial morphology classification. Furthermore, point defects include, but are not limited to, particle defects, bubble defects, vacancy lattice defects, and interstitial atom lattice defects; line defects include, but are not limited to, scratch defects and dislocation lattice defects; surface defects include, but are not limited to, interlayer voids; and bulk defects include, but are not limited to, crack defects and doping concentration deviations. For example, in an optional embodiment of the present invention, real-time collected and pre-processed time-series data of multiple processes such as photolithography, etching, and deposition (including the photolithography process coating temperature of 25.3°C, the etching process gas flow rate of 50 sccm, and the position data of the robot arm at X=10.2 mm in the deposition process, etc.) are input into the trained global prediction model. Based on the dynamic change characteristics of these time-series data (such as the abnormal fluctuation of the etching process gas flow rate from 50 sccm to 45 sccm within 100 ms), the model outputs in real time the defect type (scratch defect, which belongs to line defect), defect location, and defect occurrence probability corresponding to a specific wafer (wafer ID: W20260304001). Specifically, the defect location on the wafer is precisely located and uniquely identified using a globally unified wafer three-dimensional coordinate system (XYZ). The X-axis coordinate is 12.3 mm, the Y-axis coordinate is 15.7 mm, representing the planar position of the defect on the wafer surface, and the Z-axis coordinate is 2.1 μm, representing the internal interlayer position of the defect along the wafer thickness direction. This clarifies that the defect originates in the wafer surface etching area (excluding dielectric layers, metal interconnect layers, and other layers), achieving global defect spatial localization with a positioning accuracy of 8 nm (≤10 nm). At the same time, the model outputs an 88% probability of the defect occurrence, intuitively reflecting the high risk level of defect generation and providing a precise basis for triggering adjustments to process parameters in subsequent etching processes.

[0019] S5 automatically generates process parameter adjustment suggestions based on preset process adjustment trigger conditions, associates the corresponding process nodes and sensor data anomalies, and pushes the adjustment suggestions to the equipment controller of the corresponding process in real time to realize real-time adaptive adjustment of process parameters. At the same time, it records the adjustment process and prediction results for subsequent model iteration and optimization.

[0020] The process adjustment trigger condition is: the probability of defect occurrence ≥ the preset early warning threshold; In this embodiment of the invention, a preset defect warning threshold (≥85%) is established. When the predicted probability of a defect occurrence reaches or exceeds this threshold, the process adjustment trigger module automatically associates the corresponding defect type, process node, and sensor data anomalies to generate targeted process parameter adjustment suggestions. For example, when a scratch defect in the etching process is predicted, the etching gas flow rate (deviation ±5%) and etching time (deviation ±3%) are adjusted; when a bubble defect in the photolithography process is predicted, the resist coating temperature (deviation ±2℃) and vacuum degree (deviation ±0.01MPa) are adjusted. The adjustment suggestions are pushed to the equipment controller of the corresponding process in real time, realizing real-time adaptive adjustment of process parameters. After the adjustment is completed, the system collects the adjusted sensor data in real time and inputs it into the global prediction model for secondary verification to ensure that the probability of defect occurrence drops below the warning threshold, forming a closed-loop control of "prediction-adjustment-verification".

[0021] Corresponding to the real-time prediction method for semiconductor manufacturing defects based on machine learning described in the above embodiments, Figure 2 This diagram illustrates a structural block diagram of a real-time semiconductor manufacturing defect prediction system based on machine learning, according to an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. The system includes:

[0022] The data acquisition module is used to deploy corresponding multi-type sensor groups in the core processes of semiconductor manufacturing to collect multi-dimensional time-series data of the corresponding processes in the wafer manufacturing process in real time. The data preprocessing module is used to clean, normalize, and align the collected time-series data, unifying the time-series data from different processes and sensors to the same time axis, and generating a standardized time-series dataset. The Federated Learning Model Building and Training module is used to build and train federated learning models. The real-time prediction module is used to input the multi-process time-series data that has been collected and preprocessed in real time into the trained global prediction model. Based on the dynamic change characteristics of the time-series data, the model outputs the defect type, defect location and defect occurrence probability corresponding to a specific wafer in real time. The process adjustment module is used to automatically generate process parameter adjustment suggestions based on preset process adjustment trigger conditions, associate the corresponding process nodes and sensor data anomalies, and push the adjustment suggestions to the equipment controller of the corresponding process in real time to achieve real-time adaptive adjustment of process parameters.

Claims

1. A real-time prediction method for semiconductor manufacturing defects based on machine learning, characterized in that, Includes the following steps: Deploy multiple types of sensor groups in the core processes of semiconductor manufacturing to collect multi-dimensional time-series data of the corresponding processes in real time during wafer manufacturing; The collected time-series data is cleaned, normalized, and aligned to a unified time axis, generating a standardized time-series dataset. To construct and train a federated learning model; The multi-process time-series data, which are collected and preprocessed in real time, are input into the trained global prediction model. Based on the dynamic change characteristics of the time-series data, the model outputs the defect type, defect location and defect occurrence probability corresponding to a specific wafer in real time. Based on preset process adjustment trigger conditions, the system automatically generates process parameter adjustment suggestions, associates them with corresponding process nodes and sensor data anomalies, and pushes the adjustment suggestions to the equipment controller of the corresponding process in real time, thereby achieving real-time adaptive adjustment of process parameters.

2. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The standardized time-series dataset is associated with a unique wafer ID and process node information.

3. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The core processes include, but are not limited to, photolithography, etching, deposition, ion implantation, and polishing. The timing data includes, but is not limited to, temperature, pressure, gas flow rate, RF power, robot position, vacuum level, and wafer transfer speed; Each sensor data stream carries a unique process identifier, sensor identifier, and nanosecond-level timestamp.

4. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The standardized time-series dataset is associated with a unique wafer ID and process node information.

5. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The construction and training of the federated learning model are as follows: A federated learning architecture is constructed, comprising a central server and multiple process-end local nodes. Each process-end local node trains a locally improved LSTM-Transformer fusion time-series model based on its own preprocessed standardized time-series dataset and historical defect annotation data. The fusion model includes an LSTM feature extraction layer, a Transformer attention mechanism layer, and a defect prediction output layer. The LSTM feature extraction layer captures the long-term and short-term dependencies of the time-series data. The Transformer attention mechanism layer strengthens the feature weights of key process data. The defect prediction output layer uses a softmax activation function to output the defect type, defect location, and occurrence probability. Each local node only uploads model parameters to the central server and does not transmit raw sensitive data. The central server aggregates and optimizes the model parameters uploaded by each local node to generate a global prediction model. The central server encrypts and distributes the aggregated global model parameters to each local node. Each local node updates its local model based on the global model parameters and enters the next round of iterative training. Through multiple rounds of iterative training, the prediction accuracy of the global prediction model meets a preset threshold.

6. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 5, characterized in that, The central server uses a federated averaging algorithm to weight and aggregate the model parameters uploaded by all local nodes according to the defect impact weight of each process, thereby generating a global prediction model.

7. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The defect location is precisely located and uniquely identified using a globally unified wafer 3D coordinate system. The X and Y axes represent the planar position of the defect on the wafer surface, and the Z axis represents the internal interlayer position of the defect along the wafer thickness direction.

8. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The defect types are classified by spatial morphology into point defects, line defects, surface defects, and volume defects; including but not limited to particle defects, scratch defects, bubble defects, crack defects, lattice defects, interlayer voids, and doping concentration deviations, with each type of defect corresponding to the above spatial morphology classification; the lattice defects include but are not limited to vacancies, interstitial atoms, and dislocations.

9. The real-time prediction method for semiconductor manufacturing defects based on machine learning as described in claim 1, characterized in that, The process adjustment trigger condition is: the probability of defect occurrence is greater than or equal to the preset early warning threshold.

10. A real-time prediction system for semiconductor manufacturing defects based on machine learning, characterized in that, The system includes: The data acquisition module is used to deploy corresponding multi-type sensor groups in the core processes of semiconductor manufacturing to collect multi-dimensional time-series data of the corresponding processes in the wafer manufacturing process in real time. The data preprocessing module is used to clean, normalize, and align the collected time-series data, unifying the time-series data from different processes and sensors to the same time axis, and generating a standardized time-series dataset. The Federated Learning Model Building and Training module is used to build and train federated learning models. The real-time prediction module is used to input the multi-process time-series data that has been collected and preprocessed in real time into the trained global prediction model. Based on the dynamic change characteristics of the time-series data, the model outputs the defect type, defect location and defect occurrence probability corresponding to a specific wafer in real time. The process adjustment module is used to automatically generate process parameter adjustment suggestions based on preset process adjustment trigger conditions, associate the corresponding process nodes and sensor data anomalies, and push the adjustment suggestions to the equipment controller of the corresponding process in real time to achieve real-time adaptive adjustment of process parameters.

Images