A Novel Feature-spanning Machine Learning Technology for Defect Inspection

Abstract

Defect inspection (to detect, classify, measure and analyze) has long been a challenging task in semiconductor manufacturing (MFG) domain. This paper discusses a new Machine Learning (ML) approach which can be used to assist defect inspection in different MFG scenarios. Associated solutions have been developed and applied to the surface defect inspection for substrate components, used in our IC packaging MFG.During the past decade, the help from image-based Automated Optical Inspection (AOI) equipment has significantly reduced manual efforts in substrate/PCB defect examination, but is still insufficient in defect classification automation. Recently, the adoption of ML and Convolution Neural Network (CNN) based Deep Learning technologies raised much hoping to advance the defect classification automation to a new level that acceptable for rigid MFG practices. Most of the published CNN models, however, tend to use large number of learning parameters (floating-point variables) during computing in order to gain high image recognition accuracy. The parameters’ massive blow-up often causes heavy power-gobbling computation. While applying CNN to our substrate defect datasets, each of which contains hundreds of thousands of high-resolution images, collected across different MFG processes and produced products, a training run could take days or even weeks to finish. Also, more the learning variables for training are used, longer the inference time will be. The situations make these complex CNN models hard to meet our overnight retraining and in-line inference time constraint. Therefore, to develop a more efficient ML method is strongly preferred in our MFG environment.In this paper, we develop a feature-spanning ML approach which takes the feature of an image as a base. Through mathematical transformations the base is spanned to fit into an accumulated and gradually divided feature-of-classes space. During the feature-spanning process, the initialization of parameters is tightly controlled, which means the redundancy blow-up is carefully calculated to optimize the usage of computation resource.To demonstrate the advantage of our ML method, the public dataset CIFAR-10 is used for benchmarking. CIFAR-10 dataset contains sufficient diversity and is small enough to have quick observations on computing performance. Our results indicate that the parameters used by our method are only 5.6% of ResNet-101’s. We also apply this new ML technology to the inspection of several substrate defect types. In particular, defects on solder mask are hard to be recognized because their pattern’s colors are quite similar to the substrate background. The benchmarking shows very competitive results as compared with other CNNs’. Further benchmarking shows our ML method holds high degree of shift invariance property, implying that our method can help to resist MFG condition changes during the defect inspection operation.Our ML approach requires much less learning variables, and thus can achieve very fast training and inference speed. While adopted in MFG production lines, it helps to reduce computing cost/energy and comply with Green Factory policy. Besides, our ML technology has high scalability, capable of performing heterogeneous learning on data combined from different aspects, such as image plus CAM design reference, Z-height or time-series signal waveform. For continuous development efforts, we are making the ML behavior in our AI-for-MFG applications more explainable, controllable, and be self-adaptive to changes in MFG environment.