Piecewise Physics-Constrained Neural Networks for solving second-order impulsive differential equations

Due to the reliance of traditional numerical methods on grids, their performance significantly degrades when dealing with non-linear complex problems such as impulsive problems. In contrast, as mesh-free methods, physics-informed neural networks(PINNs) have more advantages in solving impulsive probl...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Neurocomputing (Amsterdam) Ročník 649; s. 130764
Hlavní autoři: Chen, Junxi, Mei, Liangcai, Yang, Chuan, Liu, Boyang, Han, Jiarong, Yao, Wulong
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 07.10.2025
Témata:
Hard enforcement
Impulsive differential equations
Numerical algorithm
Physics-Informed Neural Networks
Physics-Informed Neural Networks
Hard enforcement
Impulsive differential equations
Numerical algorithm
ISSN:0925-2312
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Due to the reliance of traditional numerical methods on grids, their performance significantly degrades when dealing with non-linear complex problems such as impulsive problems. In contrast, as mesh-free methods, physics-informed neural networks(PINNs) have more advantages in solving impulsive problems. This paper proposes a model named Piecewise-PCNNs. This model adopts multiple sub-networks architecture and hard enforcement method. It realizes the transformation from multi-objective optimization to single-objective optimization through prior information and improves the model performance by decoupling the solution space. To evaluate the performance of this model, this paper selects various types of second-order impulsive differential equations and conducts numerical experiments with other numerical methods. Experiments prove that under general circumstances, the Piecewise-PCNNs model has higher computational accuracy and efficiency, and this advantage can be extended to complex non-linear impulsive problems.
ISSN:0925-2312
DOI:10.1016/j.neucom.2025.130764