SHF: Small: Perception-Based Analytics For Semiconductor Production and Test Data

点击翻译按钮获取中文摘要

Big data analytics is transforming multiple industries by impacting how companies operate their businesses. The semiconductor industry, a key driver for the high-growth technology sector, is in the vanguard of this development. A large part of data-analytics practice in the semiconductor industry relies on workflows that involve human-in-the-loop steps to make an expert decision. How to automate those human-in-the-loop steps remains a fundamental challenge, and requires capturing the decision-making perception of an expert into software. This project will develop a set of novel methods and software tools that provide the capability for a domain expert to effectively model their perception in an analytics workflow. This capability enables automation of the workflow. Workflow automation not only enhances productivity but also reduces the chance of human error and, consequently, improves the overall quality of an operation. The experimental framework developed in the project will serve as an education platform for undergraduate and graduate students with diverse backgrounds. The platform will support training across multiple fields, including semiconductor production and test, machine learning, and software engineering. Such training provides a broader choice of career paths for students and prepares them with skills to meet future demands in industry.Data analytics in the semiconductor industry largely involves prescriptive analytics, where analytics results are often used to support decision making that demands accountability. Today, there remains a substantial gap between what data-analytics and machine-learning technologies can provide and what is needed for automating a prescriptive analytics workflow. This project will fill the gap by developing a perception-based analytics software component which enables domain experts to model their perception into an automated workflow. The project addresses two fundamental challenges in such modeling. The first concerns the lack of sufficient data samples when training a model. The second concerns the requirements for robustness when using such a model. To address the first challenge, the project will pursue a novel approach that enables learning a model in one domain via learning a model in another domain that has sufficient data samples readily available for training. To address the second challenge, the project will provide an innovative approach enabling integration of multiple learning methods to achieve a required robustness level. The values and impacts of the research results will be demonstrated based on data collected from actual semiconductor product lines and through experiments that reflect actual industrial settings.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.