A Residual-Attention Physics-Informed Neural Network for Irregular Interfaces and Multi-Peak Transport Fields

TL;DR

This paper introduces RA-PINN, a neural network that combines residual learning and attention mechanisms to accurately predict complex multi-physics fields with irregular interfaces and multi-peak structures, outperforming standard PINNs.

Contribution

The paper proposes a novel Residual-Attention PINN that enhances local feature capturing and global consistency for complex engineering problems, addressing limitations of existing PINNs.

Findings

RA-PINN outperforms standard PINN and LSTM-PINN in benchmark tests.

It effectively captures irregular interfaces and multi-peak structures.

Demonstrates high accuracy and convergence in complex multi-physics scenarios.

Abstract

In complex engineering systems such as electro-thermal-fluid coupling, rapid and accurate prediction of multi-physics fields is essential for advanced applications like digital twins and real-time condition monitoring. Traditional numerical methods often suffer from high computational latency, whereas standard Physics-Informed Neural Networks (PINNs) frequently fail to capture critical local features, such as irregular interfaces, localized high-gradient regions, and multi-peak transport structures. To address these limitations and provide high-fidelity intelligent predictions for engineering decision-making, this paper proposes a Residual-Attention Physics-Informed Neural Network (RA-PINN) as a powerful surrogate modeling engine. The proposed method incorporates residual learning and attention enhancement into the network backbone to improve the representation of oblique transition structures, narrow charge layers, and distributed hotspots while strictly preserving global field consistency. To evaluate its effectiveness as an intelligent prediction framework, three representative benchmark cases are constructed, including an oblique asymmetric interface, a bipolar high-gradient charge layer, and a multi-peak Gaussian charge migration field. Under unified training settings, the proposed RA-PINN is systematically compared with a standard pure PINN and an LSTM-PINN in terms of average error, local maximum error, structural similarity, and convergence behavior. The results show that RA-PINN consistently achieves the best overall performance across all benchmark cases, demonstrating its tremendous potential as a highly reliable core inference engine for the condition monitoring and digital twin modeling of complex multi-physics engineering systems.