Process Optimization and Product Design for Metal Additive Manufacturing via Knowledge-Assisted Machine Learning

增材制造（AM）是一项颠覆性技术，它可以制造复杂的形状、定制的零件，并且无需昂贵的工具就可以直接制造物体。在各种增材制造技术中，金属增材制造因其潜在的广泛工业应用前景而备受关注。金属增材制造的关键技术挑战包括增材制造的工艺建模、工艺参数优化和产品设计。工艺参数（如激光功率、激光速度和层厚）显著影响气孔率、抗拉强度和硬度等最终机械性能。为了理解工艺参数和产品机械性能之间的关系，需要建立基于物理的模型或数据驱动的模型。基于物理的模型使用微分方程来控制潜在的热力学过程。由于金属增材制造技术的高工艺复杂性，目前还缺乏可靠的仿真工具。数据驱动模型来源于不同工艺参数设置的实验。然而，这种方法需要指数增长的实验数量，因为参数的数量变得相对较大，这是金属增材制造的情况。实验费用很快就变得太高了。此外，这种方法需要应用于每一种不同的技术，甚至每一台机器，这阻碍了它的实际应用。增材制造设计的另一个挑战是拓扑优化结果不是参数化的，需要在增材制造之前进行繁琐的人工处理。

Additive Manufacturing (AM) is a disruptive technology that can fabricate complex shapes, customized parts, and directly make objects without expensive tooling. Among various AM technologies, metal AM has attracted the most attention due to the potentially wide application in industry. Key technical challenges to metal AM include process modeling, process parameter optimization, and product design for AM. Process parameters (e.g., laser power, laser velocity, and layer thickness) influence the final mechanical properties such as porosity, tensile strength, and hardness significantly. To understand the relationship between the process parameters and the product mechanical properties, one builds physics-based models or data-driven models. Physics-based models use differential equations that govern the underlying thermomechanical process. Due to high process complexity, there are still lack of credible and reliable simulation tools for metal AM technologies. Data-driven models are derived from experiments on different process parameter settings. This approach, however, needs an exponentially growing number of experiments as the number of parameters becomes relatively large, which is the case for metal AM. The experiment costs soon become too high. Moreover, this approach needs to be applied for every different technology and even every machine, which prevents its practice use. Another challenge in design for AM is that the topology optimization results are not parametric and need tedious manual processing before AM. This proposal aims to research and develop methods for process parameter optimization and product design for metal AM to address the above challenges. I propose to use knowledge-assisted artificial neural networks to build a data-driven model with fewer experiments, which in return help to fine-tune a physics-based model and define key parameters and their effective value ranges. An integrated and simplified hybrid model will be defined that is easily transferable between different machines and technologies. Dedicated optimization algorithms are to be developed to optimize process parameters for both Laser Beam Melting and Laser Beam Deposition metal AM machines. I also propose to use primitive geometries such as holes, cylinders and triangles to be the basic elements for topology optimization so that the output of topology optimization is parametric geometry that can be directly used for optimization and AM. The outcome of the proposal will be methods and a software prototype that can generate parametric topology and geometry, model metal AM processes, and optimize process parameters for maximum efficiency, lowest cost, or highest product quality from metal AM. This research will fill a gap in the literature and meet the needs of an emerging US$6 billion industry. In total, 4 PhD, 2 MSc and 5 undergraduate students will be directly trained via this research program.