CDS&E: Computation-Informed Learning of Melt Pool Dynamics for Real-Time Prognosis

金属增材制造（AM）为制造复杂零件提供了巨大的机会。然而，在打印过程中，复杂的零件几何形状、不均匀的散热和不同的激光扫描的共同影响经常导致熔池过热。过热问题导致各种质量问题。因此，对熔池行为的了解和快速预测是打印高质量零件的必要条件。数据科学模型（如深度学习或DL）可以使用不同类型的熔池数据来有效预测过热。但数据科学模型缺乏透明度，计算成本高，需要大量的训练数据。另一方面，计算模型可以理解复杂的熔池行为，但需要不断更新模型参数，不适合快速预测。该奖项提供了一种综合的方法，利用两种模型的强度来快速预测熔池过热。该项目的成果不仅有助于深度学习的基础知识，而且还使该项目的测试平台作为增材制造社区的公共工具得到广泛接受。研究结果将有助于许多行业，包括航空航天、医疗保健、工具和模具、汽车等。该项目的跨学科性质还有助于培养未来的数字制造劳动力，扩大女性和代表性不足的少数群体在数据科学驱动的研究和教育中的参与。本研究通过开发一种新的基于计算的深度学习（Co-DL）方法，弥合了对熔池动力学的基本理解和实时预测方面的知识差距。研究团队将：(1)建立选择性激光熔化（SLM）的计算流体动力学（CFD）模型，以生成无法测量的补充数据；(2)创建网络基础设施，以实现多模式数据管理、情境化、集成和互操作性，从数据分析中提取知识，并连接Co-DL测试平台；(3)开发Co-DL建模方法，将熔池动力学的物理规律和CFD模型的增强数据集成到DL训练和学习算法中；(4)建立一套基于小数据的深度学习加速和半监督学习方法；(5)为金属AM社区创建实时在线Co-DL测试平台。由此产生的方法将解决纯数据驱动的深度学习模型缺乏可解释性的主要限制，显着减少Co-DL模型训练和推理的时间延迟，并创建网络基础设施，以实现数据管理、上下文化、集成、互操作性和与Co-DL测试平台的接口。该奖项反映了美国国家科学基金会的法定使命，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。

Metal additive manufacturing (AM) offers a great opportunity for making complex parts. However, the collective impact of complex part geometry, nonuniform heat dissipation, and diverse laser scanning often cause overheating of the melt pool during the printing process. The overheating problem leads to various quality issues. Therefore, the understanding and fast prediction of melt pool behaviors are necessary for printing high-quality parts. Data science models (e.g., deep learning, or DL) may use diverse types of melt pool data for efficient prediction of overheating. But the data science models lack transparency, are computationally expensive, and need massive training data. On the other hand, computational models may understand the complex melt pool behaviors, but require continuous updates of model parameters and are not suitable for fast prediction. This award provides an integrated approach by using the strength of both models for fast prediction of melt pool overheating. The outcome of this project will not only contribute to the fundamental knowledge of deep learning but also enable the broad acceptance of the project's testbed as a public tool for the AM community. The results will help many industry sectors including aerospace, healthcare, tools, and mold, automotive, and others. The project’s interdisciplinary nature also helps train the future digital manufacturing workforce by broadening the participation of women and underrepresented minority groups in data science-driven research and education.This research bridges the knowledge gap in fundamental understanding and real-time prognosis of melt pool dynamics by developing a new computation-informed deep learning (Co-DL) approach. The research team will: (1) develop a computational fluid dynamics (CFD) model of selective laser melting (SLM) to generate complementary data which cannot be measured otherwise; (2) create cyberinfrastructure to enable multimodal data curation, contextualization, integration, and interoperability, extracting knowledge from data analytics, and interfacing Co-DL testbed; (3) develop a Co-DL modeling method to integrate physical laws of melt pool dynamics and augmented data from the CFD model into DL training and learning algorithm; (4) create a set of DL acceleration and semi-supervised learning approaches with small data; and (5) create a real-time online Co-DL testbed for the metal AM community. The resulting method will solve a major limitation of pure data-driven DL models for lacking explainability, significantly reduce the time-latency of the Co-DL model training and inference, and create cyberinfrastructure to enable data curation, contextualization, integration, interoperability, and interfacing with the Co-DL testbed.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.