Partial-Physics-Informed Multi-fidelity Modeling of Manufacturing Processes
Partial-Physics-Informed Multi-fidelity Modeling of Manufacturing Processes
复制标题
DOI：
10.1016/j.jmatprotec.2023.118125
复制
发表时间：
2023-08
影响因子：
6.3
通讯作者：
J. Cleeman;K. Agrawala;Evan Nastarowicz;R. Malhotra
中科院分区：
文献类型：
--
作者：
J. Cleeman;K. Agrawala;Evan Nastarowicz;R. Malhotra
制造过程的设计和控制取决于其参数效应的预测建模。机器学习（ML）模型的可部署性使其在这方面越来越受到关注。最近的工作已经解决了创建必要的训练数据的实验和计算成本。但是这些方法由于需要直观地和迭代地导出用于生成训练数据的精确的基于物理的过程模型而招致物理开发成本。这个问题在文献中很少涉及。本文描述了一种科学信息多保真度辅助降低成本的机器学习（Smart-ML）方法来应对这一挑战。新颖之处在于放松现有的约束，即多保真度学习中基于物理的源必须定性地匹配实验基础事实，同时约束源使用守恒定律。具有不同物理理解水平的加法，减法和混合加法-变形过程被用作测试平台，以证明Smart-ML可以将物理开发成本降低多个人类年，实验成本降低60%，计算成本降低几个数量级。这些结果讨论的背景下，所提出的方法可以移动ML超越创建已知的基于物理的模型的副本，对加速和廉价的派生功能新的ML为基础的过程模型从部分已知的物理和小的实验数据集。
The design and control of manufacturing processes hinges on predictive modeling of its parametric effects. The deployability of Machine Learning (ML) models has made them of increasing interest for this purpose. Recent work has addressed the experimental and computational costs of creating the requisite training data. But these methods incur a physics development cost due to the need to intuitively and iteratively derive accurate physics-based process models for generating the training data. This issue is rarely addressed in the literature. This paper describes aScience-informedmultifidelity-aidedreduced-cost MachineLearning (Smart-ML) approach to tackle this challenge. The novelty lies in relaxing the existing constraint that the physics-based source in multi-fidelity learning must qualitatively match the experimental ground truth while constraining the source to use conservation laws. An additive, a subtractive, and a hybrid additive-deformative process with different levels of physical understanding are used as testbeds to demonstrate that Smart-ML can reduce the physics development cost by multiple human-years, experimental cost by as much as 60 %, and computational cost by orders of magnitude. These results are discussed in the context of how the proposed approach can move ML beyond the creation of copies of known physics-based models towards the accelerated and inexpensive derivation of functionally new ML-based process models from partially known physics and small experimental datasets.