基于机器学习的矩形腔内Stokes流场预测  

Prediction of Stokes Flow Field in a Rectangular Cavity Based on Machine Learning

作　　者：吴桂群 钟家铭 邓辅秦 曾志强 洪智勇 徐百平 Wu Guiqun;Zhong Jiaming;Deng Fuqin;Zeng Zhiqiang;Hong Zhiyong;Xu Baiping(School of Intelligent Manufacturing,Wuyi University,Jiangmen,Guangdong 529020,China;School of Chemical Engineering and Technology,Guangdong Industry Polytechnic University,Guangzhou 510300,China)

机构地区：[1]五邑大学智能制造学部,广东江门529020 [2]广东轻工职业技术大学轻化工技术学院,广州510300

出　　处：《机电工程技术》2024年第9期117-120,139,共5页Mechanical & Electrical Engineering Technology

基　　金：国家自然科学基金资助项目(11972023,12102306);广东省教育厅重点项目(2020ZDZX2051)。

摘　　要：熔体输送过程发生在单螺杆挤出机均化段,涉及压力建立和混合机理。流体流动可简化为矩形腔内高黏度流体上盖拖动的流动问题,因此,速度场的精确求解非常重要。目前,流场求解多采用传统的数值模拟方法和可视化实验测试方法完成,所需的时间成本高。基于涡量-速度方法,通过数值模拟方法获得矩形腔内高黏度牛顿流体流场,抽取局部速度场作为数据集,尝试一种全新的从局部推测全局的流场预测方法。训练集、验证集和测试集的比例为1∶1∶8,使用机器学习方法,基于Tensor Flow建立多输入、多输出的6层人工神经网络模型,对腔体内全局速度场进行拓展预测研究。结果表明,在仅用10%训练数据的情况下,人工神经网络模型也能得到较精准的全局速度场预测结果。对比数值模拟全局数据,水平方向速度及垂直方向速度的MSE、L_(2)相对误差均达到了较小的范围,并且R_(2)均达到了0.99,证实了该方法的可行性与可靠性。所做工作为今后采集局部流场数据开展全局流场预测提供依据,也为采用神经网络模型实现挤出机内全局速度场插值进而开展流体运动数值模拟追踪提供了新思路。The melt conveying takes place in the metering zone of a single screw extruder,associated with the pressure establishment and mixing mechanism.Such fluid flow can be simplified as a lid-driven fluid flow problem in a rectangular cavity,hence the accurate solution of velocity field is very important.Recently,the solution of flow field is mainly obtained through the traditional numerical simulation method or the experimental visualization techniques,which is lengthy and time-consuming.So,the velocity field of highly viscous fluid in a rectangular cavity is obtained by the numerical simulation method based on vorticity-velocity method,a small part of velocity field is sampled as the data set,a kind of new prediction method of the global flow field in a viewpoint of extending from a part of the local data is attempted.The ratio of the training set to the verification set,and further to the test set is 1∶1∶8.Machine learning method is used to establish a multi-input and multi-output artificial neural network model(ANN)with a total of six layers based on TensorFlow to predict the global velocity field in the cavity.The results show that the ANN can give the relatively accurate prediction of global flow field even with only a 10%of the total data as the training data.By comparing to the global data,the relative errors of MSE and L_(2) of the horizontal and vertical velocity both reached a small range with R_(2) reaching 0.99.The feasibility and reliability of the method are verified.The work provides a basis for collecting a part of the local flow field data to predict the global flow field in the future,and also provides a new idea for performing the numerical front tracking to examine mixing behavior through the global velocity interpolation using ANN model in the screw extruders.

关 键 词：STOKES流 矩形腔 人工神经网络 机器学习 数值模拟 

分 类 号：O351.3[理学—流体力学] TP391[理学—力学]