SEMICONDUCTOR WAFER DEFECT ANALYSIS USING AI/ML

Problem Statement:

Semiconductor manufacturing is a highly complex and competitive branch of industry, comprising hundreds of process steps, which do not allow any deviations from the specification. Depending on the application area of the products, the production chain is subject to strict quality requirements. Therefore, the entire manufacturing process  can involve lengthy repetitive steps, and takes place in a sterile clean fabrication room designed to prevent even the tiniest speck of dust from falling on the pristine wafers. These wafers are processed through various equipment during the whole manufacturing process. It is extremely critical that the gas flows, temperatures, film thicknesses etc are controlled precisely. However, it is quite common that the equipment malfunction for various reasons. So, monitoring these equipment’s with on board sensors gives an early indication of any malfunction. Often, these are equipment’s results in production yield losses. If these equipment errors/process deviations are not addressed at the appropriate stage, then there might be decrease in test yield that results in the wafer manufacturing costs. After a detailed analysis and research, Ai Labs developed and implemented a Proof of Concept (POC) using its proprietary engine Minsky to highlight wafer test data pattern recognition that contains sensor recordings and high-tech production equipment. The objective of this case study is to find correlations between the wafer test data and the sensor data in order to identify the root cause and classify them (Good, Bad) in the process using machine learning classification model based on the threshold value.

Solution Overview:

After comprehensively evaluating the challenges/data, we used Minsky to accurately model the historical data which comprised of some wafer test data and sensor data parameters like Wafer (Wafer number), Timestamp (the timestamp of the respective sensor (Called as Params) recordings (176 timestamps per wafer - represented as approximately every second one recording for the sensors),Sensor data (sensor1,sensor 2 etc).The first 24 sensors belong to equipment 1 and remaining 32 belong to equipment 2 to predict the wafer defects based on 2 categories(good, bad). This result can be used for the identification of root cause of equipment failure.

Minsky modelled and analyzed the measurement attributes using machine learning algorithms for historical labelled wafer data that was mapped to corresponding sensor data. Trained Minsky models were used with Real-time equipment data to predict the corresponding equipment failure and also the potential yield loss in wafer fabrication process.

• Provides you a list of % dependency features that can be used to optimize the outcomes.
• Ability to fine tune or optimize the models by trying different algorithms / prediction attributes
• Easy integration with other third-party solutions to further analyze yield information from Wafer probe and Test data.

Minsky Key Benefits:

• User-Friendly cloud-based AI platform
• Scalable across various domains/data.
• Provides you a list of % dependency features that can be used to optimize the outcomes.
• Ability to fine tune or optimize the models by trying different algorithms / prediction attributes
• Easy integration with other third-party solutions to further analyze yield information from Wafer probe and Test data.

Results/Benefits

• Improved production yield
• Decrease in production and operation costs.
• Identification of product defects /anomalies in early stages with better accuracy.
• Reduce scrap and improvement in production process.

Fig.1: Minsky ML Based Results for Semiconductor Fab

Based on the data from Fig 1 above, our analysis shows that Decision Tree Classifier (93%) and Random Forest Classifier (92.7%) gave the highest model accuracy. In terms of the equipment measurements attributes data, Param_42 showed the highest dependency cause for the equipment failure followed by Param_19