Physics-integrated neural differentiable (PiNDiff) model for composites manufacturing
Physics-integrated neural differentiable (PiNDiff) model for composites manufacturing
复制标题
DOI：
10.1016/j.cma.2023.115902
复制
发表时间：
2023-03
影响因子：
7.2
通讯作者：
Deepak Akhare;T. Luo;Jian-Xun Wang
中科院分区：
文献类型：
--
作者：
Deepak Akhare;T. Luo;Jian-Xun Wang
由于复合材料的重量轻、性能高，人们正在开发各种制造技术来改进复合材料的制造。复合材料的机械性能取决于制造过程的各种变量和参数，这些变量和参数即使不是不可能，也是具有挑战性的，难以通过实验确定和优化。由于涉及复杂的物理问题，传统的第一性原理建模方法是不可用的。一种混合模型，将不完整的物理知识与可用测量数据结合在一个可区分的编程框架内，开辟了应对挑战的新途径。在这项工作中，提出了一个物理集成神经可微(PiNDiff)模型，其中将部分已知的物理集成到递归网络结构中，以实现有效的学习和推广。通过对热固性复合材料厚板固化过程的模拟，验证了该方法的优越性和潜力，给出了厚板固化过程的部分主导物理模型。所提出的PiNDiff模型显示了从有限的、间接的数据中学习未知物理的能力，同时可以用于推断未观察到的变量和参数。比较了PiNDiff模型和两种最新的黑盒深度学习模型(SOTA)的性能，并详细讨论了其相对于纯数据驱动模型和基于第一性原理的物理模型的优势。已证明的PiNDiff策略可以提供对物理仅部分已知且稀疏的间接数据可用的现象进行建模的一般策略。
Various manufacturing technologies are being developed to improve the manufacturing of composites owing to their low weight and high performance. The mechanical properties of the composites depend on various variables and parameters of the manufacturing process, which are challenging, if not impossible, to determine and optimize experimentally. Traditional first-principle modeling approaches are not accessible due to the complex physics involved. A hybrid model that combines incomplete physics knowledge with available measurement data within a differentiable programming framework opens up new avenues to tackle the challenges. In this work, a physics-integrated neural differentiable (PiNDiff) model is developed, where the partially known physics is integrated into the recurrent network architecture to enable effective learning and generalization. The merit and potential of the proposed method have been demonstrated in modeling the curing process of thick thermoset composite laminates, whose governing physics is partially given. The proposed PiNDiff model shows the capability to learn unknown physics from the limited, indirect data and, meanwhile, can be used to infer unobserved variables and parameters. The performance of the PiNDiff model has been compared with two state-of-the-art (SOTA) black-box deep learning models, and its advantages over the purely data-driven models and first-principles physics-based models have been discussed in detail. The demonstrated PiNDiff strategy may provide a general strategy to model phenomena where physics is only partially known and sparse, indirect data are available.