.png){kind=icon}
.png){kind=icon}
Return
Prediction of interfacial delamination in SPIF sandwich panels based on Differential Evolution-tuned Extreme Learning Machine
.png){kind=icon}
PRE
AI
H
M
Y
J
DOI：10.1007/s00170-026-18666-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Interfacial delamination is a critical failure mode in single-point incremental forming (SPIF) of metal–composite sandwich panels, while repeated high-fidelity finite-element (FE) simulations for toolpath evaluation remain computationally expensive. This study develops a lightweight Differential-Evolution-tuned Extreme Learning Machine (DE-ELM) framework for path-wise prediction of interface damage in an Al/PA6-GFRP/Al sandwich panel during SPIF. Explicit FE simulations were conducted under 30 path conditions by combining five wall-angle levels, three tool-head diameters, and two step-down settings, resulting in 9,000 path-wise samples. A representative SPIF experiment was further conducted, in which digital image correlation (DIC) and three-dimensional scanning were used to compare the measured deformation response with the FE results. The experimental forming morphology and full-field strain distribution showed reasonable consistency with the numerical model, supporting the use of the FE model for generating path-wise training data within the investigated forming configuration. The cohesive stiffness-degradation variable SDEG was used as the prediction target. After feature relevance screening and multicollinearity diagnosis, a reduced input set consisting of forming depth, path curvature, wall angle, tool-head diameter, and step-down was adopted to improve model parsimony and interpretability. To avoid path-level data leakage, model performance was evaluated using a nested validation strategy, with an outer Leave-One-Angle-Out test scheme and an inner five-fold cross-validation procedure for parameter optimization. Under this stricter validation framework, DE-ELM achieved an average R2 of 0.9196, RMSE of 0.0522, and MAE of 0.0421 across the five held-out angle conditions. Comparisons with ELM, GA-ELM, PSO-ELM, XGBoost, SVM, and MLP show that DE-ELM provides the most favorable balance between prediction accuracy and stability within the investigated SPIF configuration. SHAP-based interpretation further indicates that forming depth and path curvature are the most influential retained descriptors for the predicted damage response. The proposed framework provides an efficient surrogate tool for preliminary damage-risk evaluation in SPIF of sandwich panels, with DIC- and scanning-based experimental comparisons providing initial support for the FE-predicted deformation response. Broader process-window extension and direct experimental characterization of interfacial damage remain important directions for future work.
Keywords:
Single-point incremental forming
Sandwich panel
Interfacial delamination
Extreme learning machine
Differential evolution
Journal
T
IF:
0
Papers:
2.0K
Citations:
0
Organization
Cited Papers