En
期刊文章
文章精选
当期目录
过刊浏览
高被引
高下载
特邀文章
专刊合集
待刊文章
Volume 114
您当前的位置：首页 > 期刊文章 > 过刊浏览 > Volumes 108-119 (2025) > Volume 114
MAP-Net: A multi-layer attention physics-informed network for accurate prediction of cyclone separation efficiency
Lixin Yang a b, Peixiao Ji a b, Jianyi Chen a b *
a State Key Laboratory of Heavy Oil Processing, China University of Petroleum, Beijing, 102249, China
b Beijing Key Laboratory of Process Fluid Filtration and Separation, 102249, China
10.1016/j.partic.2026.04.003
Volume 114, July 2026, Pages 130-140
Received 2 March 2026, Revised 31 March 2026, Accepted 2 April 2026, Available online 15 April 2026, Version of Record 30 April 2026.
E-mail: jychen@cup.edu.cn
Highlights

• A novel MAP-Net integrating physical constraints and attention mechanisms is proposed for separation efficiency prediction.

• The hybrid model significantly improves prediction accuracy and generalization capability compared to pure data-driven approaches.

• MAP-Net exhibits strong interpretability by successfully capturing both macroscopic physical trends and complex microscopic nonlinearities.


Abstract

Accurate prediction of the separation efficiency of cyclone separators is of great significance for industrial process optimization and energy saving and emission reduction. Addressing the limitations of traditional empirical models (narrow application range), numerical simulations (high computational costs), and pure data-driven models (poor generalization ability under small sample sizes and tendency to violate physical mechanisms), this study proposes a deep learning model integrating physical priors and attention mechanisms—the Multi-layer Attention Physics Network (MAP-Net). First, a high-fidelity comprehensive database covering multi-dimensional geometric dimensions and operating conditions was constructed. On this basis, MAP-Net utilizes a Multi-Layer Perceptron (MLP) as its backbone architecture. On one hand, a Multi-layer Feature Attention (MFA) mechanism is introduced to enhance the adaptive extraction capability for key flow field features; on the other hand, a corresponding physics-constrained loss function is constructed based on the gas-solid two-phase separation mechanism within the cyclone separator and embedded into the model training process. The results indicate that compared with the common MLP, the coefficient of determination (R2) of MAP-Net increased from 0.7230 to 0.9688, and the Mean Absolute Error (MAE) decreased significantly from 1.4773 to 0.4965, demonstrating a significant improvement in prediction accuracy. The proposed model possesses high accuracy, strong robustness, and physical interpretability, offering corresponding guidance and support for the rapid design and separation prediction of cyclone separators.

Graphical abstract
Keywords
Cyclone separator; Separation efficiency; Neural network; Physics constraint; Attention mechanism
文章导读