The advancement of AI, IoT, and the increasing demand from hi-tech industries have paved the way for numerous opportunities and challenges in the semiconductor industry. Semiconductor manufacturing is a sophisticated and complex multistage process. Depending on its purpose, a unit product is developed by applying the most suitable processes with exact precision, such as epitaxy, etching, impurity injection, and light exposure, to the mother substrate.

 Once the manufacturing is complete, the product should undergo packaging to protect the semiconductor from the external environment before being released to the market. These semiconductor components are widely used in various fields, such as 5G mobile communication, military radar, electric vehicles, and the Internet of Things (IoT).

Defect Inspection Techniques in Semiconductor Manufacturing

As the raw materials used in the semiconductor industry are expensive, and defects caused by the manufacturing process can result in significant losses. Careful inspection of defects would help engineers locate and fix the problem at the source(s) and improve the yield. For example, tiny circuit patterns are often inspected manually using microscopes or semi-automated methods to identify possible defects and ensure non-defective production. Such techniques are prone to human errors, resulting in a high cost to quality, product wastage on the production line, and delays in time to market. Performing visually strenuous activities also affects the health of human inspectors.

Here're some of the challenges faced with manual and conventional defect inspection techniques:

Manual Inspection

In today's semiconductor industry, visible surface defects are still being inspected manually, resulting in erroneous classification when the inspectors become tired or lose objectivity. Manual inspection of defects results in high cost, lowered real-time performance, eye fatigue, and other health issues due to continuous product inspection allowing more defective parts to pass. The quality inspection involves a series of processes, including visual confirmation, to ensure components are of the right color, shape, and texture. This poses a challenge due to a wide product range variations. Quality inspectors need to be continuously updated with different quality requirements for different products.

Conventional Image Processing and Machine Learning Techniques

Classical image processing techniques fail due to arbitrariness in the size and shape of defects. Often, traditional machine learning approaches work with handcrafted features, making them application dependent and unable to scale. Additionally, it requires expensive and time-consuming manual feature engineering.

Using Deep Learning in Defect Inspection

Nowadays, semiconductor-manufacturing companies are applying deep learning approaches to images to improve the detection of defects that are difficult for humans to distinguish. Deep learning is an artificial intelligence technology that is powered by artificial neural networks. These deep neural networks mimic the human brain and learn patterns from the enormous data used to train them. This helps to classify future pieces of information. With visual inspection technology, integration of deep learning algorithms allows differentiating defects and other abnormalities with better accuracy than classical machine learning or image processing techniques.

With advancements in technology, it is now possible for high-precision manufacturing companies to enable real-time defect inspection, diagnosis and ensure optimal equipment usage, product quality inspections, and prognosis. Implementing deep learning training platforms together with continuous learning and deployment of models will help in smarter decision-making. These systems improve yield/throughput and ensure the maximum ROI from expensive equipment by targeting an ideal uptime of 24x7x365 operations. It also drastically reduces the long-term effects of stress and fatigue on human inspectors.

In deep learning, there is no handcrafting of features as the neural networks can automatically learn them. For the developed model to be accurate, reliable, and robust, it must be trained on huge amounts of data. To develop a model that provides accuracy in production, data scientists and AI engineers have to use quality data that are correctly annotated, choose the right network architecture, create multiple model versions, evaluate them, and perform cross-validation. They have to verify the different performance metrics before finalizing the model that can be deployed in production.

Because of the dynamic nature of the technology and open research initiatives for solving a specific problem, there are numerous network architectures available for different types of problems and use cases. There is also a need to explain why the model has a prediction result during inference. Automation of various stages of deep learning workflow will accelerate the model development process.

Let's consider the typical use case of defect inspection with deep learning. We can bring in automation in different phases like image annotation, image augmentation, network topology selection, training, evaluation, cross-validation, inference, and deployment. This leads to rapid prototyping and ensures that errors are at a minimum. Automation of development lifecycle of deep learning models leads to streamlined workflow and ecosystem to facilitate AI/DL based development, plug and play architecture stitching the different lifecycle phases like data preparation, neural network selection, training and inferencing, deployment, re-training, minimal manual intervention, life cycle management, and model reuse.

Way Forward

Democratizing deep learning through automation helps to accelerate the development time for defect inspection models in semiconductor manufacturing. It benefits the data engineers, data scientists, and data analysts involved in creating deep learning models.

Accelerating deep learning workflows through automation can help achieve improved productivity, reduce human error, and improve the return on investment. Data scientists' and data analysts' valuable time can be devoted to research and identifying the best network topologies, customization options, optimization of parameters, etc.