CAREER: Privacy-preserving Transfer Learning for Process-defect Modeling toward Accelerated Cross-system Certification for Metal Additive Manufacturing

工艺缺陷关系是增材制造（AM）零件认证的关键要素之一，这一直是加速AM技术在行业中部署的主要挑战。先进的机器学习方法，利用大量的数据来表征过程缺陷的关系已经研究了AM认证。然而，一些增材制造和认证课程，特别是高价值的金属零件，是漫长和昂贵的;因此，如果认证可以在不同的增材制造系统之间转移，它可能会大大扩大增材制造技术的工业应用。虽然理论上可行，但出于认证目的将来自多个AM系统的数据组合在共享平台上是不切实际的，因为希望保护知识产权和敏感数据。因此，缺乏的是一个整体的策略来共享从不同的AM系统中学习到的知识，而不损害私人信息。这个教师早期职业发展（CAREER）奖支持隐私保护AM过程缺陷建模和认证手段在不同系统的基础研究。该项目旨在建立迁移学习基础，在保护工艺和零件机密性的同时，理解并建立不同系统之间的金属AM中的工艺-缺陷关系。此外，与研究紧密结合的教育活动将为来自不同群体（包括少数民族和女性）的下一代制造工程师提供隐私保护制造系统建模的基本培训。当前数据驱动的AM认证方案主要集中于表征各个系统的过程缺陷关系（即，一个模型用于每个单个系统，并且不能推广到其他系统，甚至类似的系统）。虽然最先进的迁移学习方法可以利用从多台机器收集的数据进行跨系统研究，但研究需要保持一定的机密性-对于零件和过程-以实现这种协作。因此，本项目的目标是通过建立一个数据共享平台，在不泄露关键部件和处理数据的情况下，在多个AM系统之间实现工艺缺陷建模，从而推动金属AM技术的规模化。如果成功的话，该研究项目的主要贡献将是一个隐私保护的迁移学习框架，该框架来自以下研究活动，使用定向能量沉积AM作为示例：1）使用物理学通知的张量分解方法通过解耦可分配给产品设计和过程质量的可变性分量来构建掩蔽的过程特征，2）通过多任务迁移学习建立跨系统的工艺缺陷关系，以表征系统内和系统间的可变性，以及3）通过基于基本物理原理集成部件级密度和工艺级热数据来增强AM认证能力。该项目由土木、机械和制造创新部门（CMMI）和激励竞争力研究的既定计划（EPSCoR）共同资助。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The process-defect relationship is one of the key elements to the certification of additive manufacturing (AM) parts, which has been a major challenge in accelerating AM technology deployments in the industry. Advanced machine learning methods that leverage massive data to characterize the process-defect relationship have been studied for AM certifications. However, some AM fabrications and certification courses, especially for high-valued metallic parts, are lengthy and costly; thus, if the certification could be transferrable between different AM systems, it may greatly broaden the industrial use of AM technologies. Though feasible in theory, combining data from multiple AM systems on a shared platform for the certification purpose is not practical because of the desire to protect intellectual properties and sensitive data. What is lacking, therefore, is a holistic strategy to share knowledge learned from different AM systems without compromising the private information. This Faculty Early Career Development (CAREER) award supports fundamental research on privacy-preserving AM process-defect modeling and certification means across different systems. The project aims to establish a transfer learning groundwork, while protecting the process and part confidentiality, to understand and establish the process-defect relationship in metal AM between different systems. In addition, educational activities closely integrated with the research will provide basic training in privacy-preserving manufacturing systems modeling to next-generation manufacturing engineers from diverse groups, including minorities and women. Current data-driven AM certification schemes largely focus on characterizing the process-defect relationship of individual systems (i.e., one model for each single system and not generalizable to other systems, even similar ones). Although the state-of-the-art transfer learning methods can leverage data collected from multiple machines for cross-system studies, the research need is to maintain certain confidentiality—for both the part and process—to realize such collaboration. The goal of this project, hence, is to advance the scale-up of metal AM technologies by establishing a data-sharing platform, which enables process-defect modeling among multiple AM systems without divulging critical part and processing data. If successful, the major contribution of the research project will be a privacy-preserving transfer learning framework derived from the following research activities using directed energy deposition AM as an example: 1) constructing masked process features through de-coupling variability components assignable to product designs and process quality using a physics-informed tensor decomposition method, 2) establishing cross-system process-defect relationship through multi-task transfer learning to characterize intra- and inter-system variability and 3) enhancing AM certification capability by integrating part-level density and process-level thermal data based on fundamental physics principles. This project is jointly funded by the division of Civil, Mechanical and Manufacturing Innovation (CMMI) and the Established Program to Stimulate Competitive Research (EPSCoR).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.