A Physics-Guided Neural Network Dynamical Model for Droplet-Based Additive Manufacturing
A Physics-Guided Neural Network Dynamical Model for Droplet-Based Additive Manufacturing
复制标题
用于基于液滴的增材制造的物理引导神经网络动力学模型
DOI：
10.1109/tcst.2021.3128422
复制
发表时间：
2022-09
影响因子：
4.8
通讯作者：
Uduak Inyang-Udoh;Sandipan Mishra
中科院分区：
文献类型：
--
作者：
Uduak Inyang-Udoh;Sandipan Mishra
本文开发了一种物理引导的数据驱动模型，用于基于液滴的增材制造中打印的零件的高度演化。所提出的模型是一个卷积递归神经网络（ConvRNN），其结构是基于高度演化过程中质量守恒的物理理解而推导的。由于这种物理引导的模型结构，获得的模型参数对于打印部件的几何形状是不变的，因此可以从一种几何形状移植到另一种几何形状，进化的物理稳定性条件直接转化为神经网络的训练稳定性，并且与纯黑盒模型相比，训练该模型所需的数据要少得多。该模型的这些方面已在喷墨 3D 打印装置上进行了实验验证。通过使用大约两个数量级的数据进行训练，所提出的模型优于现成的黑盒多层感知器（神经网络），同时在测试数据上提供较小的 1.7 美元\times 均方根误差。所提出的模型还与最先进的降阶线性模型进行了比较，并显示测试数据的均方根误差小了 1.4 美元\ 倍。最后，实验结果还强调，学习的模型参数是几何不变的，也就是说，在一种几何体上训练的模型参数可以用于预测其他几何体的高度图演化，而无需重新学习。
This article develops a physics-guided data-driven model for the height evolution of parts printed in droplet-based additive manufacturing. The proposed model is a convolutional recurrent neural network (ConvRNN) whose structure is derived based on the physical understanding of mass conservation during the height evolution. Because of this physics-guided model structure, the model parameters obtained are invariant to the geometry of the printed part and thus portable from one geometry to another, the conditions on physical stability of the evolution translate directly to training stability of the neural network, and the data required to train this model are much less compared to a pure black-box model. These aspects of the model are validated experimentally on an inkjet 3-D printing setup. The proposed model outperforms a black-box off-the-shelf multilayer perceptron (neural network) by using about two orders of magnitude less data for training, at the same time delivering $1.7\times $ smaller rms error on test data. The proposed model is also compared with a state-of-the-art reduced order linear model and shows $1.4\times $ smaller rms error on test data. Finally, experimental results also underline that the model parameters learned are geometry invariant, that is, the model parameters trained on one geometry can be used to predict the height map evolution for other geometries without relearning.