Iman Nejati Ajbisheh, Javad Rezapour, Sina Gohari Rad,
Volume 16, Issue 1 (3-2026)
Abstract
Thin-walled tubes are effective crashworthiness structures that absorb energy under axial loading, while preventing the transmission of injurious acceleration and excessive forces to the protected section and thereby reducing damage severity. This study presents a high-accuracy artificial neural network (ANN) framework for predicting the dynamic response and impact resistance of thin-walled steel tubes under high-velocity axial impacts. The model was developed using a hybrid dataset comprising 300 experimental impact tests and 4000 finite element (FE) simulations, with systematic variations in tube geometry and impact conditions. After parameter optimization, the final model consisted of a 24-layer network with 180 neurons per layer and achieved high accuracy (R-value above 0.985). Error assessments across the four physical criteria (PFE, MFE, AEE, and SHE) show that the ANN predicts peak force, mean force, absorbed energy, and shortening with average errors generally below 10%, demonstrating strong predictive capability. Overall, the model not only offers a much faster alternative to FE simulations but also accurately reproduces oscillatory behavior and deformation progression, making it a reliable tool for impact response prediction.
Mr Mohamad Masoud Mohamadkhani, Dr Farzad Ghafoorian,
Volume 16, Issue 2 (6-2026)
Abstract
Adhesively bonded joints, particularly the Single-Lap Joint (SLJ), are widely used in structural applications. However, stress concentrations at the overlap edges significantly limit their load-bearing capacity. Modifying the adherend geometry by introducing notches is an effective technique for stress redistribution and strength improvement. In this study, a Machine Learning (ML) approach is proposed to predict the peak peel and shear stresses in notched SLJs based on geometric parameters, namely the notch depth, notch angle, notch width, and the notch distance from the end of the overlap. First, a comprehensive dataset comprising 1183 Finite Element Analysis (FEA) simulations was generated. Subsequently, four ML models—Artificial Neural Networks (ANN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Machines (SVM)—were developed to map the geometric features to the maximum stress values. To ensure robust performance and prevent overfitting, a Randomized Search Cross-Validation technique was employed for hyperparameter tuning across all models prior to final evaluation. The evaluation results demonstrated highly accurate predictions across all algorithms. Specifically, the ANN model achieved the most precise results, yielding the lowest Mean Absolute Percentage Error (MAPE) for both peel (0.28%) and shear (0.16%) stresses. The SVM, XGBoost, and RF models also exhibited excellent predictive capabilities, with all MAPE values remaining well below 1%. Feature importance analysis from XGBoost and SHapley Additive exPlanations (SHAP) applied to the optimal ANN revealed that notch distance from the edge and notch depth are the key design parameters influencing stress distributions.
