期刊
JOURNAL OF COMPUTATIONAL PHYSICS
卷 496, 期 -, 页码 -出版社
ACADEMIC PRESS INC ELSEVIER SCIENCE
DOI: 10.1016/j.jcp.2023.112578
关键词
Composite neural network; Data-driven modeling; Surrogate modeling; Impulsive dynamics; Nonlinear dynamics
This study proposes a novel and robust composite neural network called cNN-DP, which is capable of handling impulsive time-transient dynamics. The network divides the prediction tasks into multiple sub-networks to improve the accuracy and performance of the model.
In mechanical engineering, abundant high-quality data from simulations and experimental observations can help develop practical and accurate data-driven models. However, when dynamics are complex and highly nonlinear, designing a suitable model and optimizing it accurately is challenging. In particular, when data comprise impulsive signals or high-frequency components, training a data-driven model becomes increasingly challenging. This study proposes a novel and robust composite neural network for impulsive time-transient dynamics by dividing the prediction of the dynamics into tasks for three sub-networks, one for approximating simplified dynamics and the other two for mapping lower-order derivatives to higher-order derivatives. The mapping serves as the temporal differential operator, hence, the name composite neural network with differential propagation (cNN-DP) for the suggested model. Furthermore, numerical investigations were conducted to compare cNN-DP with two baseline models, a conventional network and another employing the autogradient approach. Regarding the convergence rate of model optimizations and the generalization accuracy, the proposed network outperformed the baseline models by many orders of magnitude. In terms of computational efficiency, numerical tests showed that cNN-DP requires an acceptable and comparable computational load. Although the numerical studies and descriptions focus on accelerations, the proposed network can be easily extended to any other application involving impulsive data.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据