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 116
您当前的位置：首页 > 期刊文章 > 当期目录 > Volume 116
A continual learning surrogate network for field prediction of fluidized beds with evolving internals
Tingting Liu a, Yong Jiang b, Ming Pan a, Zheng-Hong Luo a *, Xizhong Chen a *
a State Key Laboratory of Synergistic Chem-Bio Synthesis, Department of Chemical Engineering, School of Chemistry and Chemical Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China
b China Huanqiu Contracting & Engineering (Beijing) CO., LTD, Beijing, 100012, China
10.1016/j.partic.2026.07.001
Volume 116, September 2026, Pages 355-368
Received 28 April 2026, Revised 29 June 2026, Accepted 7 July 2026, Available online 7 July 2026, Version of Record 14 July 2026.
E-mail:
Highlights

• CLBedNet uses a dual-branch surrogate (point cloud + SDF) to predict voidage and axial gas velocity fields.

• A wall-proximity-aware gating module adaptively fuses near-wall geometry cues and bulk SDF features.

• The integrated continuous learning suite alleviates catastrophic forgetting during the update of incremental models.

• Gate–SDF trends identify a learned “effective thickness”, improving interpretability of geometry influence.

• CFD-DEM-level fidelity with ∼104× faster inference enables high-throughput design and digital-twin use.


Abstract

The design of complex internals in fluidized beds is hindered by the computational cost of numerical simulations and the lack of efficient surrogate models capable of capturing near-wall flow fields across evolving geometries. This study proposes a novel dual-branch architecture integrated with a continual learning suite for the rapid prediction of fluidization hydrodynamics around superquadric internals, named CLBedNet. The model fuses point cloud features with signed-distance representations, employing a wall-proximity-aware gating mechanism to adaptively weight each branch based on wall proximity. A continual training category comprising knowledge distillation, replay buffer, and online elastic weight consolidation are introduced to mitigate catastrophic forgetting during the introduction of evolving geometries. It is shown that CLBedNet can achieve high-fidelity predictions for voidage and axial gas velocity distributions of fluidized bed with varying internals, which is promising for establish digital twin and the high-throughput optimization of industrial fluidized-bed reactors.

Graphical abstract
Keywords
文章导读