SBIR Phase I: Machine learning-powered simulation of additive manufacturing for real-time design and process optimization

这个小型企业创新研究(SBIR)第一阶段项目的更广泛影响是通过超高速工程模拟软件加速更大规模的添加剂制造(AM)的采用。2020年，AM行业的价值为126亿美元，在为美国航空航天、医疗和汽车行业提供先进设计和支持分布式供应链方面具有巨大潜力。然而，由于缺乏有效和可靠的工程工作流程而导致的试错过程，AM正面临着缓慢的采用。拟议的超快模拟技术可以在设计阶段实时预测AM零件可能出现的制造问题，从而减少制造失败和原型制作。该项目还寻求产生系统的知识，了解机器学习如何帮助克服科学计算中一些长期存在的根本挑战，并帮助推动用于数字制造的工程软件。该小型企业创新研究(SBIR)第一阶段项目将机器学习与有限元方法(FEM)相结合，为AM开发了一个概念验证软件，速度快3-5个数量级，用于预测由于高温、残余变形和残余应力造成的制造故障。传统的局部AM模拟计算方法需要数小时到数天的时间，并且依赖于迭代的、分层的方法。拟议的项目试图用深度学习取代传统模拟方法中最耗时的步骤，并实施一步法。提出的混合数据驱动加物理仿真框架包括开发特征驱动的深度学习模型和基于过程参数的迁移学习模型，并将这些模型与有限元方法相耦合。该项目还旨在应用AM物理建模、制造部件的三维(3D)扫描以及用于培训和模型可伸缩性的现场监测的混合数据集，并将其作为基准。该团队寻求通过在该项目中开发的简化用户界面和应用程序编程接口(API)的试点测试来展示技术优势。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is to accelerate larger-scale adoption of additive manufacturing (AM) through ultrafast engineering simulation software. The AM industry was worth $12.6 billion in 2020 and holds great potential in providing advanced designs and enabling distributed supply chains for the US aerospace, medical, and automotive industries. However, AM is facing slow adoption due to trial and error processes casued by the lack of an efficient and reliable engineering workflow. The proposed ultrafast simulation technology may provide real-time predictions of possible manufacturing issues for AM parts in the design phase, thereby reducing manufacturing failures and prototyping. The project also seeks to generate systematic knowledge of how machine learning can help overcome some long-lasting fundamental challenges in scientific computing and help advance engineering software used for digital manufacturing. This Small Business Innovation Research (SBIR) Phase I project integrates machine learning with finite element methods (FEM) to develop a proof-of-concept for 3-5 orders of magnitude faster process simulation software for AM used to predict manufacturing failures due to high temperature, residual distortion, and residual stresses. The traditional computation method for part-scale AM simulation takes hours to days and relies on an iterative, layer-wise approach. The proposed project seeks to replace the most time-consuming steps in the traditional simulation method with deep learning and implement a one-step approach. The proposed hybrid data-driven plus physical simulation framework includes the development a feature-driven, deep learning model and a process parameter-based transfer learning model, and coupling these models with the finite element method. The project also aims to apply and benchmark hybrid datasets from AM physical modelling, three-dimentional (3D) scanning of manufactured parts, and in-situ monitoring for training and model scalability. The team seeks to demonstrate technological advantages through pilot testing with streamlined user interfaces and application programming interfaces (APIs) developed in this project.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.