摘要
格子玻尔兹曼算法(LBM,Lattice Boltzmann method)相较于传统计算流体力学方法具有程序结构简单,对复杂边界和非线性问题适应性强以及便于并行计算等诸多优点。然而,其作为一种显式算法,在计算过程中的迭代次数较多,进而消耗大量计算资源。利用深度学习在预测与回归方面的强大能力,基于LBM设计了一个由卷积层与卷积长短期记忆层组成的人工神经网络预测模型并将其命名为C-LBM(compressed LBM)。该模型能等效替代多个普通LBM迭代。对于方腔环流问题,模型完成训练后,对测试集均方差在5×10^(-6)以下,对泛化算例在10-5以下,精度得到了保障。相较于串行LBM程序,C-LBM模型计算效率提升约15倍。
Compared to the traditional computational fluid dynamics techniques,the Lattice Boltzmann method has the advantages of simple structure of program,strong adaptability to complex boundaries as well as nonlinear problems,and high parallelism.However,since LBM is an explicit algorithm,its calculation usually involves many iteration steps,and thereby consumes a huge amount of computing resources.This study takes advantage of deep learning in prediction and regression to accelerate LBM calculations innovatively.We establish a prediction model(compressed LBM or C-LBM),which involves an artificial neural network composed of convolution layers and convolution long-term and short-term memory layers.The prediction model is an equivalent substitution of multiple ordinary LBM iterations.For the two dimensional driven cavity circulation problem,the mean square error of C-LBM is less than 5×10^(-6) for the test set,and is less than 10-5 for the generalized calculation example.The calculation efficiency of C-LBM is about 15 times higher than that of the serial LBM.
作者
陈辛阳
聂滋森
蒋子超
杨耿超
姚清河
CHEN Xinyang;NIE Zisen;JIANG Zichao;YANG Gengchao;YAO Qinghe(School of Aeronautics and Astronautics,Sun Yat-sen University,Guangzhou 510006,China)
出处
《中山大学学报（自然科学版）（中英文）》
CAS
CSCD
北大核心
2021年第5期39-49,共11页
Acta Scientiarum Naturalium Universitatis Sunyatseni
基金
国家重点研发计划国际合作项目(2018YFE9103900)
高性能计算专项(2016YFB0200603)
国家自然科学基金(11972384)
广东省促进经济高质量发展专项资金(GDOE[2019]A01)
广州市科技计划项目产学研协同创新重大专项(201704030089)。
