.png){kind=icon}
.png){kind=icon}
Return
Efficient Prediction Method for 3D Mesoscopic Stress Fields in Composite Solid Propellants Based on Multi-slice Input
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compscitech.2026.111651.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Aiming at the exorbitant computational cost of the finite element method, inadequate feature characterization of existing deep learning models and scarce 3D samples in predicting the 3D mesoscopic stress field of composite solid propellants, this paper proposes an efficient prediction method integrating a two-tier dataset, dual-attention U-Net (DA-U-Net), multi-slice input and transfer learning. We constructed 2D pre-training and 3D slice fine-tuning datasets via an improved Random Sequential Adsorption algorithm, and designed the DA-U-Net with edge and channel attention to accurately capture the features of particle-matrix interfaces and high-stress regions. An encoder-freezing and decoder-only fine-tuning strategy was adopted, with 1–13 slice input schemes and six transfer learning strategies compared experimentally. Results show the DA-U-Net outperformed contrast models in pre-training, achieving R2=0.9361, MAE=0.0218 MPa, SSIM=0.9324 and Force Equilibrium Residual=0.0206 N. Considering prediction accuracy, computational efficiency and metric saturation characteristics comprehensively, 9-slice input was determined as the optimal scheme, reducing the prediction error by about 27% compared with single-slice input. Transfer learning with 2 3D samples and a fully unlocked decoder was the best strategy, with performance close to freezing the first decoder layer. The model also exhibited good generalization for 30%–45% particle volume fractions, providing an efficient solution for propellant stress field prediction and a reference for micromechanics research of particle-reinforced composites.
Keywords:
Deep learning
Stress field prediction
Composite solid propellants
Multi-slice input
Transfer learning
Journal
IF:
9.8
Papers:
703
Citations:
5.0W